Merge pull request #620 from Routstr/fix/fee-payout-crash-guard

fix: guard fee payouts against crash double-payments

.env.example

@@ -2,6 +2,9 @@
 UPSTREAM_BASE_URL=https://api.openai.com/v1
 UPSTREAM_API_KEY=your-upstream-api-key
 
+# Tinfoil (confidential inference enclaves, EHBP)
+# TINFOIL_API_KEY=your-tinfoil-api-key
+
 # ADMIN_PASSWORD=secure-admin-password
 
 # Database

.gitignore

@@ -16,6 +16,9 @@ dist/
 *.db-shm
 *.db-wal
 .*wallet.sqlite3
+.wallet/
+AGENTS.md
+TEST_SUITE_OVERVIEW.md
 *models.json
 .cashu
 .relay
@@ -38,3 +41,4 @@ proof_backups
 
 *.todo
 ui_out
+.worktrees

Dockerfile.full

@@ -7,7 +7,7 @@ WORKDIR /app/ui
 RUN corepack enable pnpm && corepack prepare pnpm@10.15.0 --activate
 
 # Copy UI source
-COPY ui/package.json ui/pnpm-lock.yaml* ./
+COPY ui/package.json ui/pnpm-lock.yaml* ui/pnpm-workspace.yaml* ./
 RUN pnpm install --frozen-lockfile
 
 COPY ui/ ./

Makefile

@@ -98,7 +98,7 @@ docker-down:
 lint:
 	@echo "🔍 Running linting checks..."
 	$(RUFF) check .
-	$(MYPY) routstr/ --ignore-missing-imports
+	$(MYPY) .
 
 format:
 	@echo "✨ Formatting code..."
@@ -107,7 +107,7 @@ format:
 
 type-check:
 	@echo "🔎 Running type checks..."
-	$(MYPY) routstr/ --ignore-missing-imports
+	$(MYPY) .
 
 # Development setup
 dev-setup:
@@ -234,7 +234,7 @@ ci-test:
 ci-lint:
 	@echo "🤖 Running CI linting..."
 	$(RUFF) check . --exit-non-zero-on-fix
-	$(MYPY) routstr/ --ignore-missing-imports --no-error-summary
+	$(MYPY) . --no-error-summary
 
 # Debug helpers
 test-debug:

compose.yml

@@ -13,6 +13,7 @@ services:
       - ./ui_out:/output:z
     command:
       ["sh", "-c", "mkdir -p /output && cp -r /app/built/. /output/ && echo 'UI build copied to mounted volume' && ls -la /output/ && echo 'UI built and ready' && tail -f /dev/null"]
+    restart: unless-stopped
 
   routstr:
     build: .
@@ -31,6 +32,7 @@ services:
       - 8000:8000
     extra_hosts: # Needed to access locally running models
       - "host.docker.internal:host-gateway"
+    restart: unless-stopped
 
   tor:
     image: ghcr.io/hundehausen/tor-hidden-service:latest
@@ -41,6 +43,7 @@ services:
       - HS_ROUTER=routstr:8000:80
     depends_on:
       - routstr
+    restart: unless-stopped
 
 volumes:
   tor-data:

docs/api/errors.md

@@ -113,42 +113,109 @@ All errors follow a consistent JSON structure:
 **Status:** 402  
 **Resolution:** Top up API key balance
 
-#### Invalid Token
+### Cashu Token Redemption Errors
+
+These errors are returned when a Cashu token you pay with cannot be redeemed.
+They apply to every endpoint that accepts a token:
+
+- **Per-request payment** via the `X-Cashu` header (chat completions + Responses API).
+- **API key top-up** via `POST /v1/wallet/topup`.
+- **Minting an API key** from a token sent in `Authorization: Bearer <cashu-token>`.
+
+All three share one classifier, so the same failure yields the same HTTP status
+and sanitized message everywhere. Structured error envelopes (`X-Cashu` and
+`Authorization: Bearer <cashu-token>`) also expose the same `type` and `code` —
+branch on `type` (or `code` for finer granularity). `POST /v1/wallet/topup`
+keeps its existing plain-string `detail` envelope, so branch on status there.
+
+| `type` | Status | `code` | Retryable | Meaning |
+|--------|--------|--------|-----------|---------|
+| `token_already_spent` | 400 | `cashu_token_already_spent` | No | The token was already redeemed. |
+| `invalid_token` | 400 | `invalid_cashu_token` | No | The token is malformed or cannot be decoded. |
+| `mint_error` | 422 | `cashu_token_swap_fees_exceed_amount` | No | Token value is too small to cover the mint's swap/melt fees. |
+| `mint_error` | 422 | `cashu_foreign_mint_swap_failed` | No | Swapping the token from a foreign mint to the primary mint failed. |
+| `mint_unreachable` | 503 | `cashu_mint_unreachable` | **Yes** | The mint could not be reached (DNS failure, refused/reset connection, timeout). The token is fine — retry once the mint recovers. |
+| `cashu_error` | 400 | `cashu_token_redemption_failed` | No | The token could not be redeemed for another expected reason. |
+| `cashu_error` | 400 | `cashu_token_zero_value` | No | The token redeemed to zero (empty/dust token, or value fully consumed by fees). |
+| `token_consumed` | 500 | `cashu_token_consumed` | No | The token was **spent** (melted/redeemed) but crediting it then failed. Do not retry — the token is gone; contact support to reconcile. |
+| `api_error` | 500 | `internal_error` | Maybe | Unexpected server-side fault during redemption. |
+
+!!! important "Retry only `mint_unreachable`"
+    Only `mint_unreachable` (503) means the same token will work again later —
+    everything else is a permanent property of the token and must not be
+    blindly retried. Use exponential backoff for the 503. In particular, a
+    `token_consumed` 500 means the mint already spent the token, so a retry
+    would fail as `token_already_spent`.
+
+#### Mint Unreachable (retryable)
 
 ```json
 {
   "error": {
-    "type": "payment_error",
-    "message": "Invalid Cashu token",
-    "code": "invalid_token",
-    "details": {
-      "reason": "Token already spent"
-    }
+    "type": "mint_unreachable",
+    "message": "Cashu mint is unreachable",
+    "code": "cashu_mint_unreachable"
   }
 }
 ```
 
-**Status:** 400  
-**Resolution:** Use a valid, unspent token
+**Status:** 503
 
-#### Mint Unavailable
+**Resolution:** The token is valid — the mint is temporarily down. Retry with
+backoff, or pay with a token from a different mint.
+
+#### Token Already Spent
 
 ```json
 {
   "error": {
-    "type": "payment_error",
-    "message": "Cannot connect to Cashu mint",
-    "code": "mint_unavailable",
-    "details": {
-      "mint_url": "https://mint.example.com",
-      "retry_after": 60
-    }
+    "type": "token_already_spent",
+    "message": "Cashu token already spent",
+    "code": "cashu_token_already_spent"
   }
 }
 ```
 
-**Status:** 503  
-**Resolution:** Try again later or use different mint
+**Status:** 400
+
+**Resolution:** Use a fresh, unspent token. Do not retry with the same token.
+
+#### Response envelope differs by endpoint
+
+The `error` object above is identical everywhere, but the surrounding envelope
+depends on how you paid:
+
+- **`X-Cashu` header payments** (chat + Responses API) return the object at the
+  top level, alongside a `request_id`:
+
+  ```json
+  {
+    "error": { "type": "mint_unreachable", "message": "Cashu mint is unreachable", "code": "cashu_mint_unreachable" },
+    "request_id": "req-abc123"
+  }
+  ```
+
+  The original token is echoed back in the `X-Cashu` **response header only when
+  it is still spendable** (e.g. `mint_unreachable`, `invalid_cashu_token`, fee
+  errors) so you can recover/retry it. It is **not** echoed for spent/consumed
+  tokens (`cashu_token_already_spent`, `cashu_token_consumed`,
+  `cashu_token_zero_value`, `internal_error`) — retrying those can never succeed.
+
+- **`Authorization: Bearer <cashu-token>`** (API key minting) wraps it in
+  FastAPI's `detail` field:
+
+  ```json
+  { "detail": { "error": { "type": "mint_unreachable", "message": "Cashu mint is unreachable", "code": "cashu_mint_unreachable" } } }
+  ```
+
+- **`POST /v1/wallet/topup`** returns a plain string message under `detail` —
+  it carries the shared HTTP **status** and **message** (e.g. `503` for an
+  unreachable mint) but not the structured `type`/`code`, so branch on the
+  status code here:
+
+  ```json
+  { "detail": "Cashu mint is unreachable" }
+  ```
 
 ### Validation Errors
 
@@ -347,7 +414,7 @@ class ErrorHandler:
         'rate_limit',
         'upstream_timeout',
         'model_overloaded',
-        'mint_unavailable'
+        'cashu_mint_unreachable'
     }
     
     # Errors requiring user action

docs/ehbp-proxy-support.md

@@ -0,0 +1,150 @@
+# EHBP Proxy Support for Tinfoil Models
+
+## Problem
+
+The SDK's `SecureClient.fetch` encrypts request bodies with HPKE (EHBP protocol)
+and sends them to the Routstr provider. The Routstr proxy had no EHBP handling:
+
+1. It tried to `json.loads()` the binary HPKE-sealed body → failed with a 400
+2. The upstream PPQ.AI public endpoint (`/v1/chat/completions`) doesn't speak
+   EHBP, so the response had no `Ehbp-Response-Nonce` header
+3. `SecureClient` threw `Missing Ehbp-Response-Nonce header` because it expects
+   every response from an EHBP-configured `baseURL` to carry that header
+
+## Root cause
+
+PPQ.AI exposes EHBP-aware inference at `/private/`, separate from the public
+`/v1/` endpoint. The Routstr proxy was forwarding to `/v1/` (the public
+endpoint) instead of `/private/` (the enclave endpoint). The public endpoint
+can't decrypt the body, returns a normal HTTP response, and the SDK can't
+decrypt it because there's no nonce header.
+
+The PPQ private-mode proxy (`ppq-private-mode-proxy/lib/proxy.ts`) shows the
+correct pattern: `SecureClient` talks to `api.ppq.ai/private/v1/chat/completions`,
+which decrypts inside the attested enclave and returns an EHBP-encrypted
+response with the `Ehbp-Response-Nonce` header.
+
+## What was changed
+
+### `routstr/proxy.py`
+
+Detects EHBP requests by checking for the `Ehbp-Encapsulated-Key` header (set
+by the EHBP transport on every encrypted request). For EHBP requests:
+
+- Skips JSON body parsing (the body is binary ciphertext, not JSON)
+- Reads the model ID from the `X-Routstr-Model` header (set by the SDK) instead
+  of from `body.model`
+- Routes through new `forward_ehbp_request` (bearer auth) and
+  `forward_ehbp_x_cashu_request` (x-cashu auth) methods
+- Skips reactive 400 param correction (can't parse encrypted response body)
+- Still charges the user via `pay_for_request` (uses `max_cost_for_model` from
+  the model registry, not the body)
+
+### `routstr/upstream/base.py`
+
+Keeps EHBP as an explicit opt-in provider capability instead of making every
+upstream provider appear EHBP-capable:
+
+- `supports_ehbp = False` by default
+- `get_ehbp_forwarding_target(path, model_obj)` raises `NotImplementedError`
+  unless a provider opts in and returns a provider-specific EHBP target
+
+The actual EHBP forwarding logic does **not** live in `base.py`.
+
+### `routstr/upstream/ehbp.py`
+
+Contains the shared opaque EHBP transport and billing helpers:
+
+- `EHBPForwardingTarget` — provider-specific target URL plus extra headers
+- `forward_ehbp_request()` — forwards the encrypted body, captures Tinfoil
+  usage from a response header or streaming HTTP trailer, and finalizes bearer
+  billing at actual cost (falling back to max cost when usage is unavailable)
+- `forward_ehbp_x_cashu_request()` — redeems the Cashu token, refunds the full
+  token on upstream failure, and refunds the difference between the redeemed
+  amount and actual cost (or max cost when usage is unavailable)
+
+### Provider support
+
+EHBP is currently enabled only for `TinfoilUpstreamProvider`. It forwards to
+Tinfoil's attested enclave and requests `X-Tinfoil-Usage-Metrics` for billing.
+PPQ.AI retains its private-target implementation, but `supports_ehbp = False`
+until it has a provider-specific trusted usage/model-binding strategy.
+
+## Why it's done this way
+
+The proxy is a **blind relay** for EHBP requests. It cannot decrypt the body
+(only the attested enclave can), so it must:
+
+1. Get the model ID from a header, not the body
+2. Forward the raw bytes without parsing or transformation
+3. Stream the response back without SSE/cost parsing
+4. Pass through EHBP protocol headers (`Ehbp-Encapsulated-Key` on request,
+   `Ehbp-Response-Nonce` on response)
+
+Cost tracking happens at the proxy level. Routstr reserves or redeems up to
+`max_cost_for_model`, then Tinfoil's out-of-band usage header/trailer allows it
+to finalize at actual token cost. If trusted usage is missing or invalid, the
+proxy safely falls back to max-cost billing.
+
+## End-to-end flow
+
+```
+SDK                                  Routstr Proxy                    PPQ.AI /private/
+  │                                       │                                │
+  │── X-Routstr-Model: tinfoil-kimi-k2-6 ─│                                │
+  │── Ehbp-Encapsulated-Key: <hex> ───────│                                │
+  │── Authorization: Bearer <cashu> ──────│                                │
+  │── body = HPKE-encrypted(kimi-k2-6) ───│                                │
+  │                                       │                                │
+  │                                  detects Ehbp-Encapsulated-Key         │
+  │                                  reads model from X-Routstr-Model      │
+  │                                  does billing/routing                  │
+  │                                       │                                │
+  │                                  adds X-Private-Model: private/kimi-k2-6
+  │                                  forwards raw body to /private/v1/...  │
+  │                                       │──────────────────────────────▶│
+  │                                       │                          enclave decrypts
+  │                                       │                          runs inference
+  │                                       │◀── Ehbp-Response-Nonce ──────│
+  │                                       │◀── encrypted response ────────│
+  │                                       │                                │
+  │                                  streams response back untouched       │
+  │◀── encrypted response ────────────────│                                │
+  │                                        │                                │
+  SecureClient reads nonce, decrypts       │                                │
+  SDK SSE processing sees plaintext        │                                │
+```
+
+## Model ID mapping
+
+Three parties see three different model IDs:
+
+| Party | Header/Body | Value | Source |
+|---|---|---|---|
+| Routstr proxy | `X-Routstr-Model` header | `tinfoil-kimi-k2-6` | SDK sends full caller-facing id |
+| Tinfoil usage metrics | `model` field | `kimi-k2-6` | Enclave reports the model actually served |
+| Tinfoil enclave | `body.model` (encrypted) | `kimi-k2-6` | SDK strips `tinfoil-` prefix before encryption |
+
+## Implementation status
+
+A dedicated `TinfoilUpstreamProvider` (`routstr/upstream/tinfoil.py`) now
+implements the direct blind-upstream pattern described above. The shared EHBP
+helpers in `routstr/upstream/ehbp.py` were extended to:
+
+- Request usage metrics via `X-Tinfoil-Request-Usage-Metrics: true`.
+- Parse `X-Tinfoil-Usage-Metrics` from the response header (non-streaming) or
+  HTTP trailer (streaming).
+- Override the forwarding URL with a validated `X-Tinfoil-Enclave-Url` when the
+  SDK sends it.
+- Finalize bearer billing with the dedicated EHBP actual-cost finalizer.
+- Compute X-Cashu refunds from actual cost instead of max cost.
+
+See `docs/tinfoil-direct-integration.md` for the full implementation notes.
+
+## Verification status
+
+Unit coverage includes usage parsing, target validation, HTTP trailer capture,
+response-size limits, and bearer payment finalization. End-to-end requests have
+verified both non-streaming usage headers and streaming usage trailers against
+Tinfoil. SDK behavior on proxy-generated non-2xx responses without an
+`Ehbp-Response-Nonce` still merits explicit end-to-end coverage.

docs/tinfoil-direct-integration.md

@@ -0,0 +1,493 @@
+# Tinfoil / PPQ Private-Mode Integration Notes
+
+This document summarizes the current options for integrating Tinfoil/PPQ private models with Routstr, based on the EHBP work in this branch and local testing against `ppq-private-mode-proxy`.
+
+## Background
+
+PPQ private models run behind a Tinfoil/EHBP flow:
+
+- Request bodies are HPKE-encrypted by a Tinfoil client.
+- The PPQ `/private/` endpoint routes ciphertext to the attested enclave.
+- The enclave decrypts, runs inference, and returns an encrypted response.
+- The caller's Tinfoil client decrypts the response locally.
+
+The PPQ private-mode proxy (`~/projects/ppq-private-mode-proxy`) uses this pattern with the JavaScript `tinfoil` SDK:
+
+```ts
+import { SecureClient } from "tinfoil";
+
+const apiBase = "https://api.ppq.ai";
+
+const client = new SecureClient({
+  baseURL: `${apiBase}/private/`,
+  attestationBundleURL: `${apiBase}/private`,
+  transport: "ehbp",
+});
+
+await client.ready();
+
+const response = await client.fetch(`${apiBase}/private/v1/chat/completions`, {
+  method: "POST",
+  headers: {
+    "Content-Type": "application/json",
+    Authorization: `Bearer ${process.env.PPQ_API_KEY}`,
+    "X-Private-Model": "private/gpt-oss-120b",
+    "x-query-source": "api",
+  },
+  body: JSON.stringify({
+    model: "gpt-oss-120b", // enclave-internal model id
+    messages: [{ role: "user", content: "Hello" }],
+  }),
+});
+
+const json = await response.json();
+console.log(json.usage);
+```
+
+`X-Private-Model` carries the PPQ-facing private model id, while the encrypted JSON body uses the enclave-internal model id without the `private/` prefix.
+
+## PPQ private model pricing
+
+PPQ exposes private model pricing through:
+
+```text
+GET https://api.ppq.ai/v1/models?type=all
+```
+
+Filter models whose IDs start with `private/`.
+
+Example private pricing observed:
+
+| Model | Input USD / 1M tokens | Output USD / 1M tokens |
+|---|---:|---:|
+| `private/gpt-oss-120b` | `0.79125` | `1.31875` |
+| `private/llama3-3-70b` | `1.84625` | `2.90125` |
+| `private/qwen3-vl-30b` | `1.31875` | `4.22` |
+| `private/glm-5-2` | `1.5825` | `5.53875` |
+| `private/gemma4-31b` | `0.47475` | `1.055` |
+| `private/kimi-k2-6` | `1.5825` | `5.53875` |
+
+The `ppq-private-mode-proxy` commit `ba984214793d3bca0f7d046b6955d42abc1c6843` changed OpenClaw display metadata to align with a 5% API margin. The live PPQ model endpoint returns the more precise rates above, which already include that margin.
+
+Actual PPQ private billing is:
+
+```text
+price_usd =
+  input_tokens  * input_per_1M_tokens  / 1_000_000
++ output_tokens * output_per_1M_tokens / 1_000_000
+```
+
+A test request to `private/gpt-oss-120b` produced:
+
+```json
+{
+  "input_count": 74,
+  "output_count": 3,
+  "price_in_usd": 0.00006250875
+}
+```
+
+which exactly matches:
+
+```text
+74 * 0.79125 / 1_000_000 + 3 * 1.31875 / 1_000_000
+= 0.00006250875 USD
+```
+
+## Integration architectures
+
+There are three materially different ways Routstr could integrate Tinfoil/PPQ private inference.
+
+## Option A: User integrates Tinfoil directly
+
+```text
+User app / SDK
+  -> Tinfoil SecureClient / TinfoilAI
+    -> EHBP-encrypted request
+      -> Tinfoil/PPQ private enclave
+```
+
+Properties:
+
+- Best privacy for the user.
+- Routstr is not in the request path.
+- User's client encrypts requests and decrypts responses.
+- Usage is visible to the user's app after decryption.
+- Billing is handled directly by Tinfoil/PPQ.
+
+Example with Tinfoil's OpenAI-compatible client:
+
+```ts
+import { TinfoilAI } from "tinfoil";
+
+const client = new TinfoilAI({
+  apiKey: process.env.TINFOIL_API_KEY,
+  transport: "ehbp",
+});
+
+const res = await client.chat.completions.create({
+  model: "llama3-3-70b",
+  messages: [{ role: "user", content: "Hello" }],
+});
+
+console.log(res.choices[0].message.content);
+console.log(res.usage);
+```
+
+This is not a Routstr marketplace flow unless Routstr only acts as discovery/UI around direct Tinfoil/PPQ usage.
+
+## Option B: Routstr integrates Tinfoil as an upstream client
+
+```text
+User -> Routstr plaintext request
+  -> Routstr Tinfoil SecureClient encrypts to PPQ/Tinfoil
+    -> PPQ private enclave
+  -> Routstr receives decrypted response
+  -> Routstr bills from decrypted usage
+  -> Routstr returns plaintext response to user
+```
+
+Properties:
+
+- Easier exact billing.
+- Routstr can read the decrypted OpenAI response and `usage` object.
+- Routstr can charge exact PPQ token pricing.
+- Privacy is different: the user sends plaintext to Routstr, and Routstr sees prompts/responses.
+- End-to-end encryption is only Routstr-to-enclave, not user-to-enclave.
+
+This should be considered a separate product/provider mode, not the same as an end-to-end private relay.
+
+Practical implementation options:
+
+1. Run a Node sidecar that uses the `tinfoil` npm package and expose it as a local HTTP upstream to Routstr.
+2. Port EHBP client behavior to Python.
+3. Reuse or adapt `ppq-private-mode-proxy` as a local upstream.
+
+A Node sidecar is probably the quickest implementation path because `ppq-private-mode-proxy` already demonstrates the full flow.
+
+## Option C: Routstr uses Tinfoil as a direct blind upstream
+
+This is the current branch's design intent and is likely the best fit if Tinfoil/PPQ exposes usage metadata headers:
+
+```text
+User Tinfoil SecureClient
+  -> encrypted request body
+    -> Routstr proxy
+      -> Tinfoil/PPQ private API
+         with X-Tinfoil-Request-Usage-Metrics: true
+        -> PPQ private enclave
+      <- encrypted response
+         plus X-Tinfoil-Usage-Metrics / cost headers
+    <- Routstr proxy
+  -> user decrypts response
+```
+
+In this mode Routstr is still a normal upstream proxy from the user's point of view, but the upstream is Tinfoil/PPQ private inference and the body remains opaque to Routstr.
+
+Properties:
+
+- Strongest privacy with Routstr in the path.
+- Routstr never sees plaintext prompt or plaintext response.
+- Routstr can authenticate and route based on plaintext headers.
+- Routstr should request Tinfoil usage metadata by adding `X-Tinfoil-Request-Usage-Metrics: true` to the upstream request.
+- Tinfoil documents `X-Tinfoil-Usage-Metrics` as an upstream response header for non-streaming requests.
+- For streaming requests, Tinfoil documents `X-Tinfoil-Usage-Metrics` as an HTTP trailer available only after the response body completes.
+- If Tinfoil/PPQ returns `X-Tinfoil-Usage-Metrics` or a cost header on the encrypted response, Routstr can bill exactly without decrypting the body.
+- If usage is only present inside the encrypted response body, Routstr still cannot read it and exact billing is not possible without a separate metadata path.
+
+This is the only architecture that preserves end-to-end encryption from the user to the PPQ/Tinfoil enclave while still letting Routstr mediate payment. The key requirement is that usage/cost metadata must be returned outside the encrypted body, ideally as a response header available before body streaming begins.
+
+## Current Routstr problem
+
+The current EHBP implementation charges successful EHBP requests at `max_cost_for_model` because Routstr cannot decrypt the response body:
+
+```text
+successful EHBP request -> charge full reserved max cost
+```
+
+That is incorrect for PPQ private models. Max cost should be only a reservation/solvency ceiling. Final charge should use actual PPQ private token pricing.
+
+Desired behavior:
+
+```text
+reserve max cost
+forward encrypted request
+obtain actual usage/cost metadata
+finalize actual cost
+refund/release the difference
+```
+
+## Usage/cost metadata requirement
+
+For blind-relay exact billing, PPQ should return one of the following outside the encrypted response body:
+
+```http
+X-PPQ-Cost-USD: 0.00006250875
+```
+
+or:
+
+```http
+X-Private-Usage-Metrics: input=74,output=3
+```
+
+or:
+
+```http
+X-Tinfoil-Usage-Metrics: prompt=74,completion=3,total=77
+```
+
+Tinfoil proxy documentation references a usage-metrics flow where a proxy can request usage via:
+
+```http
+X-Tinfoil-Request-Usage-Metrics: true
+```
+
+and read usage from:
+
+```http
+X-Tinfoil-Usage-Metrics
+```
+
+Docs/example references:
+
+- https://docs.tinfoil.sh/guides/proxy-server
+- https://github.com/tinfoilsh/encrypted-request-proxy-example
+
+During local PPQ testing, PPQ responses included this CORS exposure header:
+
+```http
+Access-Control-Expose-Headers: Ehbp-Response-Nonce, X-Private-Usage-Metrics, X-Encrypted-Usage-Metrics, X-Tinfoil-Usage-Metrics
+```
+
+However, the actual tested non-streaming response did not include any of these usage headers, even when `X-Tinfoil-Request-Usage-Metrics: true` was sent.
+
+The decrypted body did include normal OpenAI usage, but only the decrypting Tinfoil client can see that body.
+
+## Query-history fallback
+
+PPQ's query history endpoint exposes actual usage and cost:
+
+```text
+GET https://api.ppq.ai/queries/history?page=1&page_count=...
+```
+
+A record includes:
+
+```json
+{
+  "timestamp": "...",
+  "model": "private/gpt-oss-120b",
+  "input_count": 74,
+  "output_count": 3,
+  "price_in_usd": 0.00006250875,
+  "query_type": "chat_completion",
+  "query_source": "api"
+}
+```
+
+This could be used as a fallback, but it is less robust than response headers/trailers because matching a request to a history row can be race-prone under concurrency. It would need a reliable request identifier or metadata field that PPQ stores in history.
+
+## Recommended Routstr direction
+
+Use Tinfoil/PPQ as a direct blind upstream and have Routstr explicitly request usage metadata:
+
+```text
+User encrypts body
+Routstr reserves max cost
+Routstr forwards encrypted body to Tinfoil/PPQ /private/
+Routstr includes X-Tinfoil-Request-Usage-Metrics: true
+Tinfoil/PPQ returns X-Tinfoil-Usage-Metrics as a response header for non-streaming,
+or as an HTTP trailer after the body completes for streaming
+Routstr finalizes exact charge
+User decrypts encrypted response
+```
+
+This preserves both:
+
+- privacy: Routstr cannot read prompts/responses;
+- exact billing: Routstr can charge actual PPQ private model cost, assuming Tinfoil/PPQ returns usage/cost metadata outside the encrypted body.
+
+Implementation steps:
+
+1. Update PPQ model fetching to include private models:
+
+   ```text
+   GET https://api.ppq.ai/v1/models?type=all
+   ```
+
+2. Register `private/*` models and any Routstr-facing aliases with correct `forwarded_model_id`.
+
+3. Keep max-cost reservation for bearer keys and X-Cashu solvency checks.
+
+4. Add parsing support for possible usage/cost headers:
+
+   ```http
+   X-PPQ-Cost-USD
+   X-Private-Usage-Metrics
+   X-Tinfoil-Usage-Metrics
+   X-Encrypted-Usage-Metrics
+   ```
+
+5. Finalize by actual cost instead of max cost:
+
+   ```text
+   actual_msats = ceil((actual_usd / sats_usd_price()) * 1000)
+   actual_msats = max(actual_msats, settings.min_request_msat)
+   actual_msats = min(actual_msats, reserved_msats)
+   ```
+
+   or, if only token counts are available:
+
+   ```text
+   actual_usd =
+     input_tokens  * input_per_1M_tokens  / 1_000_000
+   + output_tokens * output_per_1M_tokens / 1_000_000
+   ```
+
+6. If usage/cost metadata is missing, choose an explicit policy:
+
+   - fail closed and refund/revert;
+   - query PPQ history as a fallback;
+   - fallback to max-cost billing only if explicitly configured and clearly disclosed.
+
+Silent max-cost billing should not be the default for PPQ private requests.
+
+## X-Cashu consideration
+
+For bearer-auth requests, finalization can happen after the response stream completes if usage is delivered as a trailer.
+
+For `X-Cashu`, Routstr needs to return the refund token in the response headers. If usage/cost is only available after consuming the encrypted response stream, Routstr may need to buffer EHBP responses before sending them to the client so it can compute the refund amount first.
+
+Possible approaches:
+
+1. Prefer a non-trailer response header with actual cost, available before streaming body starts.
+2. Buffer EHBP X-Cashu responses and then return `X-Cashu` refund.
+3. Introduce a later/refund-claim mechanism, which would be a larger protocol change.
+
+## Summary
+
+- PPQ private models are billed per actual input/output tokens.
+- Private model rates are available from `GET /v1/models?type=all`.
+- Current Routstr EHBP billing at max cost is wrong for PPQ private models.
+- Direct Tinfoil integration inside Routstr would enable exact usage billing but would make Routstr see plaintext.
+- A blind EHBP relay preserves privacy but requires PPQ/Tinfoil to expose usage/cost in plaintext headers/trailers.
+- The preferred solution is to keep Routstr blind and have PPQ return billing metadata outside the encrypted body.
+
+## Implementation status
+
+Direct Tinfoil upstream integration is implemented in `routstr/upstream/tinfoil.py`
+and `routstr/upstream/ehbp.py`.
+
+### What was built
+
+- `TinfoilUpstreamProvider` (`provider_type = "tinfoil"`):
+  - Base URL: `https://inference.tinfoil.sh`
+  - Fetches models from the public `GET /v1/models` endpoint (no auth needed).
+  - Parses Tinfoil's pricing (`inputTokenPricePer1M`, `outputTokenPricePer1M`,
+    `requestPrice`) into the standard `Model`/`Pricing` schema.
+  - `supports_ehbp = True` — acts as a blind EHBP relay.
+  - `get_ehbp_forwarding_target()` returns a target that includes
+    `X-Tinfoil-Request-Usage-Metrics: true`.
+  - `forward_get_request()` proxies `/attestation` to `https://atc.tinfoil.sh/attestation`
+    so the SDK can fetch attestation bundles through Routstr.
+  - Registered in `routstr/upstream/__init__.py` and seeded from
+    `TINFOIL_API_KEY` env var.
+
+- `routstr/upstream/ehbp.py`:
+  - `parse_tinfoil_usage_metrics()` parses
+    `prompt=N,completion=N[,total=N][,model=<name>]` into an OpenAI-style
+    usage dict. The `model` field (added in tinfoilsh/confidential-model-router
+    PR #385) is extracted as a string.
+  - `_resolve_ehbp_target_url()` overrides the forwarding URL with
+    `X-Tinfoil-Enclave-Url` when the SDK sends it.
+  - `_strip_proxy_headers()` removes `X-Routstr-Model`,
+    `X-Tinfoil-Enclave-Url`, and `X-Tinfoil-Request-Usage-Metrics` before
+    forwarding to the enclave.
+  - `_compute_ehbp_actual_cost()` converts the usage header into msats via
+    `calculate_cost()`, clamped to `[min_request_msat, max_cost_for_model]`.
+    When the header's `model=<name>` differs from the requested model, the
+    actual served model's pricing is used for cost calculation.
+  - `forward_ehbp_request()` (bearer auth): if `X-Tinfoil-Usage-Metrics` is
+    present in the response header, finalizes with `adjust_payment_for_tokens()`
+    for exact billing; otherwise falls back to max-cost. Billing uses the
+    actual served model when it differs from the requested one.
+  - `forward_ehbp_x_cashu_request()`: if usage is available, computes the
+    refund from actual cost instead of max cost, using the actual served
+    model's pricing when applicable.
+
+- `routstr/proxy.py`: `/attestation` and `/tee/attestation` paths are forwarded
+  to Tinfoil upstreams without model/cost/auth lookups.
+
+### Billing behavior
+
+| Request shape | Usage source | Billing |
+|---|---|---|
+| Bearer, non-streaming | `X-Tinfoil-Usage-Metrics` response header | Exact token cost via `adjust_payment_for_tokens` |
+| Bearer, streaming | `X-Tinfoil-Usage-Metrics` HTTP trailer | Exact token cost (h11 captures trailers) |
+| Bearer, no usage header/trailer | N/A | Max-cost fallback |
+| X-Cashu, non-streaming | `X-Tinfoil-Usage-Metrics` response header | Refund = `redeemed - actual_cost` |
+| X-Cashu, streaming | `X-Tinfoil-Usage-Metrics` HTTP trailer | Refund = `redeemed - actual_cost` (h11 captures trailers) |
+| X-Cashu, no usage header/trailer | N/A | Refund = `redeemed - max_cost` |
+
+### Cost response headers
+
+Since EHBP response bodies are opaque encrypted blobs, per-request cost cannot
+be injected into the JSON body (as done in the normal proxy flow). Instead,
+Routstr returns cost info as response headers:
+
+| Header | Auth | Description |
+|---|---|---|
+| `X-Routstr-Cost-Msats` | Bearer, X-Cashu | Total msats charged for this request |
+| `X-Routstr-Cost-Usd` | Bearer | USD equivalent of the charge |
+| `X-Routstr-Input-Cost-Msats` | Bearer, X-Cashu | msats attributed to input tokens |
+| `X-Routstr-Output-Cost-Msats` | Bearer, X-Cashu | msats attributed to output tokens |
+
+The client/Tinfoil SDK can read these headers from the HTTP response without
+needing to decrypt the body.
+
+### Setup
+
+```bash
+TINFOIL_API_KEY=your-tinfoil-api-key
+```
+
+The provider is auto-seeded on first startup.
+
+### Usage metrics header format
+
+Tinfoil returns usage metrics in the `X-Tinfoil-Usage-Metrics` response header
+(non-streaming) or HTTP trailer (streaming) when `X-Tinfoil-Request-Usage-Metrics:
+true` is sent. As of tinfoilsh/confidential-model-router PR #385, the format is:
+
+```
+prompt=<prompt_tokens>,completion=<completion_tokens>,total=<total_tokens>,model=<served_model>
+```
+
+The `model` field carries the actual model name served by the enclave.
+Routstr uses this to:
+
+- Verify the served model matches the expected upstream model. The comparison
+  uses ``model_obj.forwarded_model_id`` (the actual upstream ID, e.g.
+  ``glm-5-2``) rather than ``model_obj.id`` (the client-facing alias, e.g.
+  ``tinfoil-glm-5-2``), so aliased models don't trigger a spurious mismatch.
+- When they genuinely differ (Tinfoil served a different upstream model than
+  expected), look up the actual served model's pricing and use it for billing.
+  The reverse lookup uses ``get_model_instance``, which resolves
+  ``forwarded_model_id`` values registered as routable aliases.
+- Log the discrepancy for observability.
+
+If the actual model is not found in Routstr's model registry, billing falls
+back to the requested model's pricing.
+
+### What still needs verification
+
+- ~~End-to-end test with a real Tinfoil SDK client against a Routstr node with
+  `TINFOIL_API_KEY` set.~~ Verified: both non-streaming (header) and streaming
+  (trailer) responses include `model=<name>`.
+- Streaming trailer capture is implemented by buffering the encrypted response
+  in `forward_with_trailer()` and then using the dedicated EHBP payment
+  finalizers for bearer and X-Cashu requests. This provides actual-cost billing
+  today, at the cost of full time-to-last-byte latency for streaming responses.
+- Whether Tinfoil's `/v1/responses` endpoint also returns usage metrics
+  headers or trailers.

migrations/versions/c6d7e8f9a0b1_add_slug_to_upstream_providers.py

@@ -0,0 +1,88 @@
+"""add slug to upstream_providers
+
+Revision ID: c6d7e8f9a0b1
+Revises: b5e7c9d1f3a2
+Create Date: 2026-06-29 00:00:00.000000
+"""
+
+from __future__ import annotations
+
+import sqlalchemy as sa
+from alembic import op
+
+from routstr.core.provider_slugs import provider_slug_base, provider_slug_candidate
+
+revision = "c6d7e8f9a0b1"
+down_revision = "b5e7c9d1f3a2"
+branch_labels = None
+depends_on = None
+
+
+def _allocate_backfill_slug(provider_type: str, reserved_slugs: set[str]) -> str:
+    base = provider_slug_base(provider_type)
+    suffix_number = 1
+    while True:
+        candidate = provider_slug_candidate(base, suffix_number)
+        if candidate not in reserved_slugs:
+            reserved_slugs.add(candidate)
+            return candidate
+        suffix_number += 1
+
+
+def _backfill_provider_slugs(conn: sa.Connection) -> None:
+    existing_rows = conn.execute(
+        sa.text(
+            "SELECT slug FROM upstream_providers "
+            "WHERE slug IS NOT NULL AND slug != ''"
+        )
+    )
+    reserved_slugs = {str(row.slug).lower() for row in existing_rows}
+
+    rows_to_backfill = conn.execute(
+        sa.text(
+            "SELECT id, provider_type FROM upstream_providers "
+            "WHERE slug IS NULL OR slug = '' "
+            "ORDER BY id"
+        )
+    )
+    for row in rows_to_backfill:
+        slug = _allocate_backfill_slug(str(row.provider_type), reserved_slugs)
+        conn.execute(
+            sa.text("UPDATE upstream_providers SET slug = :slug WHERE id = :id"),
+            {"slug": slug, "id": row.id},
+        )
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+    inspector = sa.inspect(conn)
+    columns = {c["name"] for c in inspector.get_columns("upstream_providers")}
+
+    if "slug" not in columns:
+        op.add_column(
+            "upstream_providers",
+            sa.Column("slug", sa.String(), nullable=True),
+        )
+
+    _backfill_provider_slugs(conn)
+
+    existing_indexes = {idx["name"] for idx in inspector.get_indexes("upstream_providers")}
+    if "ix_upstream_providers_slug" not in existing_indexes:
+        op.create_index(
+            "ix_upstream_providers_slug",
+            "upstream_providers",
+            ["slug"],
+            unique=True,
+        )
+
+
+def downgrade() -> None:
+    conn = op.get_bind()
+    inspector = sa.inspect(conn)
+    existing_indexes = {idx["name"] for idx in inspector.get_indexes("upstream_providers")}
+    if "ix_upstream_providers_slug" in existing_indexes:
+        op.drop_index("ix_upstream_providers_slug", table_name="upstream_providers")
+
+    columns = {c["name"] for c in inspector.get_columns("upstream_providers")}
+    if "slug" in columns:
+        op.drop_column("upstream_providers", "slug")

migrations/versions/d7e8f9a0b1c2_add_fee_payout_checkpoint.py

@@ -0,0 +1,37 @@
+"""add fee payout checkpoint
+
+Revision ID: d7e8f9a0b1c2
+Revises: c6d7e8f9a0b1
+Create Date: 2026-07-18 00:00:00.000000
+"""
+
+from __future__ import annotations
+
+import sqlalchemy as sa
+from alembic import op
+
+revision = "d7e8f9a0b1c2"
+down_revision = "c6d7e8f9a0b1"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "routstr_fees",
+        sa.Column(
+            "payout_in_progress_msats",
+            sa.Integer(),
+            nullable=False,
+            server_default="0",
+        ),
+    )
+    op.add_column(
+        "routstr_fees",
+        sa.Column("payout_started_at", sa.Integer(), nullable=True),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("routstr_fees", "payout_started_at")
+    op.drop_column("routstr_fees", "payout_in_progress_msats")

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "routstr"
-version = "0.4.3"
+version = "0.4.4"
 description = "Payment proxy for your LLM endpoint using cashu and nostr."
 readme = "README.md"
 requires-python = ">=3.11"
@@ -10,6 +10,7 @@ dependencies = [
     "aiosqlite>=0.20",
     "sqlmodel>=0.0.24",
     "httpx[socks]>=0.25.2",
+    "h11>=0.14",
     "greenlet>=3.2.1",
     "alembic>=1.13",
     "python-json-logger>=2.0.0",

routstr/algorithm.py

@@ -86,8 +86,8 @@ def get_provider_penalty(provider: "BaseUpstreamProvider") -> float:
 
 def create_model_mappings(
     upstreams: list["BaseUpstreamProvider"],
-    overrides_by_id: dict[str, tuple],
-    disabled_model_ids: set[str],
+    overrides_by_key: dict[tuple[str, int], tuple],
+    disabled_model_keys: set[tuple[str, int]],
 ) -> tuple[
     dict[str, "Model"], dict[str, list["BaseUpstreamProvider"]], dict[str, "Model"]
 ]:
@@ -107,8 +107,9 @@ def create_model_mappings(
 
     Args:
         upstreams: List of all upstream provider instances
-        overrides_by_id: Dict of model overrides from database {model_id: (ModelRow, fee)}
-        disabled_model_ids: Set of model IDs that should be excluded
+        overrides_by_key: Dict of model overrides from database
+            {(model_id_lower, upstream_provider_id): (ModelRow, fee)}
+        disabled_model_keys: Set of provider-scoped model keys that should be excluded
 
     Returns:
         Tuple of (model_instances, provider_map, unique_models)
@@ -179,14 +180,22 @@ def create_model_mappings(
         """Process all models from a given provider."""
         upstream_prefix = getattr(upstream, "upstream_name", None)
         provider_key = get_provider_identity(upstream)
+        upstream_db_id = getattr(upstream, "db_id", None)
 
         for model in upstream.get_cached_models():
-            if not model.enabled or model.id in disabled_model_ids:
+            model_key = (
+                (model.id.lower(), upstream_db_id)
+                if isinstance(upstream_db_id, int)
+                else None
+            )
+            if not model.enabled or (
+                model_key is not None and model_key in disabled_model_keys
+            ):
                 continue
 
-            # Apply overrides if present
-            if model.id in overrides_by_id:
-                override_row, provider_fee = overrides_by_id[model.id]
+            # Apply overrides only for this provider's model row.
+            if model_key is not None and model_key in overrides_by_key:
+                override_row, provider_fee = overrides_by_key[model_key]
                 model_to_use = _row_to_model(
                     override_row, apply_provider_fee=True, provider_fee=provider_fee
                 )
@@ -237,13 +246,10 @@ def create_model_mappings(
 
     # Include enabled DB overrides even when provider discovery misses models.
     # This is important for deployment-based providers like Azure.
-    for model_id, override_data in overrides_by_id.items():
-        if model_id in disabled_model_ids:
+    for (model_id, upstream_provider_id), override_data in overrides_by_key.items():
+        if (model_id, upstream_provider_id) in disabled_model_keys:
             continue
         override_row, provider_fee = override_data
-        upstream_provider_id = getattr(override_row, "upstream_provider_id", None)
-        if not isinstance(upstream_provider_id, int):
-            continue
 
         upstream_for_override = providers_by_db_id.get(upstream_provider_id)
         if upstream_for_override is None:

routstr/auth.py

@@ -20,7 +20,11 @@ from .payment.cost_calculation import (
     MaxCostData,
     calculate_cost,
 )
-from .wallet import credit_balance, deserialize_token_from_string
+from .wallet import (
+    classify_redemption_error,
+    credit_balance,
+    deserialize_token_from_string,
+)
 
 logger = get_logger(__name__)
 payments_logger = get_logger("routstr.payments")
@@ -78,6 +82,37 @@ async def check_and_reset_limit(key: ApiKey, session: AsyncSession) -> bool:
     return False
 
 
+def redemption_error_to_http_exception(error: Exception) -> HTTPException:
+    """Map a Cashu token redemption failure to a sanitized client-facing error.
+
+    Thin wrapper over the shared :func:`classify_redemption_error` so the bearer
+    path stays identical to the X-Cashu and top-up paths.
+    """
+    classified = classify_redemption_error(error)
+    if classified is None:
+        return HTTPException(
+            status_code=500,
+            detail={
+                "error": {
+                    "message": "Internal error during token redemption",
+                    "type": "api_error",
+                    "code": "internal_error",
+                }
+            },
+        )
+    error_type, status_code, message, error_code = classified
+    return HTTPException(
+        status_code=status_code,
+        detail={
+            "error": {
+                "message": message,
+                "type": error_type,
+                "code": error_code,
+            }
+        },
+    )
+
+
 async def validate_bearer_key(
     bearer_key: str,
     session: AsyncSession,
@@ -216,7 +251,17 @@ async def validate_bearer_key(
 
         try:
             hashed_key = hashlib.sha256(bearer_key.encode()).hexdigest()
-            token_obj = deserialize_token_from_string(bearer_key)
+            try:
+                token_obj = deserialize_token_from_string(bearer_key)
+            except Exception as decode_error:
+                # A malformed token is a bad token (400 invalid_cashu_token via
+                # the shared taxonomy), not an auth failure (401) — otherwise it
+                # would fall through to the generic "Invalid API key" handler.
+                raise redemption_error_to_http_exception(
+                    ValueError(
+                        f"Invalid Cashu token: could not decode token ({decode_error})"
+                    )
+                ) from decode_error
             logger.debug(
                 "Generated token hash", extra={"hash_preview": hashed_key[:16] + "..."}
             )
@@ -334,19 +379,32 @@ async def validate_bearer_key(
                         "error_type": type(credit_error).__name__,
                     },
                 )
-                raise credit_error
+                await session.rollback()
+                raise redemption_error_to_http_exception(credit_error) from credit_error
 
             if msats <= 0:
                 logger.error(
                     "Token redemption returned zero or negative amount",
                     extra={"msats": msats, "key_hash": hashed_key[:8] + "..."},
                 )
-                # Defense-in-depth: credit_balance now refuses to commit on a
-                # zero/negative redemption, but if a row was nonetheless
-                # persisted, drop it so we never leave an orphan zero-balance key.
+                # Defense-in-depth: credit_balance already raises
+                # ValueError("Redeemed token amount must be positive…") before
+                # returning (wallet.py), so this branch is only reachable if a
+                # zero/negative row was somehow persisted; drop it so we never
+                # leave an orphan zero-balance key. Reuse the shared taxonomy
+                # (cashu_error) so the envelope matches the mapper above.
                 await session.delete(new_key)
                 await session.commit()
-                raise Exception("Token redemption failed")
+                raise HTTPException(
+                    status_code=400,
+                    detail={
+                        "error": {
+                            "message": "Failed to redeem Cashu token: token yielded no value",
+                            "type": "cashu_error",
+                            "code": "cashu_token_zero_value",
+                        }
+                    },
+                )
 
             await session.refresh(new_key)
             await session.commit()
@@ -379,7 +437,7 @@ async def validate_bearer_key(
                 status_code=401,
                 detail={
                     "error": {
-                        "message": f"Invalid or expired Cashu key: {str(e)}",
+                        "message": "Invalid or expired Cashu key",
                         "type": "invalid_request_error",
                         "code": "invalid_api_key",
                     }
@@ -709,12 +767,18 @@ async def revert_pay_for_request(
 
 
 async def adjust_payment_for_tokens(
-    key: ApiKey, response_data: dict, session: AsyncSession, deducted_max_cost: int
+    key: ApiKey,
+    response_data: dict,
+    session: AsyncSession,
+    deducted_max_cost: int,
 ) -> dict:
     """
     Adjusts the payment based on token usage in the response.
     This is called after the initial payment and the upstream request is complete.
     Returns cost data to be included in the response.
+
+    The response's usage object is normalized with the default union parser in
+    ``calculate_cost``.
     """
     billing_key = await get_billing_key(key, session)
     model = response_data.get("model", "unknown")
@@ -797,7 +861,7 @@ async def adjust_payment_for_tokens(
                     extra={"error": str(e), "fee_msats": fee_msats},
                 )
 
-    match await calculate_cost(response_data, deducted_max_cost, session):
+    match await calculate_cost(response_data, deducted_max_cost):
         case MaxCostData() as cost:
             logger.debug(
                 "Using max cost data (no token adjustment)",
@@ -1015,32 +1079,37 @@ async def adjust_payment_for_tokens(
             # actual cost exceeded discounted reservation (due to tolerance_percentage)
             if cost_difference > 0:
                 # Always release the reservation and charge min(actual_cost, balance).
-                # Using a CASE expression makes this a single atomic UPDATE — no
-                # multi-level fallback needed and balance can never go negative.
+                # CASE expressions keep this atomic and safe even when the
+                # stale-reservation sweeper has already released the reservation.
                 chargeable = case(
                     (col(ApiKey.balance) >= total_cost_msats, total_cost_msats),
                     else_=col(ApiKey.balance),
                 )
+                overrun_safe_reserved = case(
+                    (
+                        col(ApiKey.reserved_balance) >= deducted_max_cost,
+                        col(ApiKey.reserved_balance) - deducted_max_cost,
+                    ),
+                    else_=0,
+                )
 
                 finalize_stmt = (
                     update(ApiKey)
                     .where(col(ApiKey.hashed_key) == billing_key.hashed_key)
-                    .where(col(ApiKey.reserved_balance) >= deducted_max_cost)
                     .values(
-                        reserved_balance=col(ApiKey.reserved_balance) - deducted_max_cost,
+                        reserved_balance=overrun_safe_reserved,
                         balance=col(ApiKey.balance) - chargeable,
                         total_spent=col(ApiKey.total_spent) + chargeable,
                     )
                 )
-                result = await session.exec(finalize_stmt)  # type: ignore[call-overload]
+                await session.exec(finalize_stmt)  # type: ignore[call-overload]
 
                 if billing_key.hashed_key != key.hashed_key:
                     child_stmt = (
                         update(ApiKey)
                         .where(col(ApiKey.hashed_key) == key.hashed_key)
-                        .where(col(ApiKey.reserved_balance) >= deducted_max_cost)
                         .values(
-                            reserved_balance=col(ApiKey.reserved_balance) - deducted_max_cost,
+                            reserved_balance=overrun_safe_reserved,
                             total_spent=col(ApiKey.total_spent) + min(billing_key.balance, total_cost_msats),
                         )
                     )
@@ -1048,51 +1117,38 @@ async def adjust_payment_for_tokens(
 
                 await session.commit()
 
-                if result.rowcount:
-                    await session.refresh(billing_key)
-                    if billing_key.hashed_key != key.hashed_key:
-                        await session.refresh(key)
-                    cost.total_msats = total_cost_msats
-                    logger.info(
-                        "Finalized payment with additional charge",
-                        extra={
-                            "key_hash": key.hashed_key[:8] + "...",
-                            "billing_key_hash": billing_key.hashed_key[:8] + "...",
-                            "charged_amount": total_cost_msats,
-                            "new_balance": billing_key.balance,
-                            "model": model,
-                        },
-                    )
-                    await _accumulate_fee(total_cost_msats)
-                    payments_logger.info(
-                        "FINALIZE",
-                        extra={
-                            "event": "finalize",
-                            "key_hash": key.hashed_key[:8] + "...",
-                            "billing_key_hash": billing_key.hashed_key[:8] + "...",
-                            "model": model,
-                            "cost_reserved": deducted_max_cost,
-                            "cost_charged": total_cost_msats,
-                            "input_tokens": cost.input_tokens,
-                            "output_tokens": cost.output_tokens,
-                            "balance": billing_key.balance,
-                            "reserved_balance": billing_key.reserved_balance,
-                            "total_spent": billing_key.total_spent,
-                            "finalize_type": "overrun",
-                        },
-                    )
-                else:
-                    # Guard fired: reservation was already released by a concurrent
-                    # finalization for this key. Nothing left to do.
-                    logger.warning(
-                        "Finalization skipped - reservation already released",
-                        extra={
-                            "key_hash": key.hashed_key[:8] + "...",
-                            "billing_key_hash": billing_key.hashed_key[:8] + "...",
-                            "attempted_charge": total_cost_msats,
-                            "model": model,
-                        },
-                    )
+                await session.refresh(billing_key)
+                if billing_key.hashed_key != key.hashed_key:
+                    await session.refresh(key)
+                cost.total_msats = total_cost_msats
+                logger.info(
+                    "Finalized payment with additional charge",
+                    extra={
+                        "key_hash": key.hashed_key[:8] + "...",
+                        "billing_key_hash": billing_key.hashed_key[:8] + "...",
+                        "charged_amount": total_cost_msats,
+                        "new_balance": billing_key.balance,
+                        "model": model,
+                    },
+                )
+                await _accumulate_fee(total_cost_msats)
+                payments_logger.info(
+                    "FINALIZE",
+                    extra={
+                        "event": "finalize",
+                        "key_hash": key.hashed_key[:8] + "...",
+                        "billing_key_hash": billing_key.hashed_key[:8] + "...",
+                        "model": model,
+                        "cost_reserved": deducted_max_cost,
+                        "cost_charged": total_cost_msats,
+                        "input_tokens": cost.input_tokens,
+                        "output_tokens": cost.output_tokens,
+                        "balance": billing_key.balance,
+                        "reserved_balance": billing_key.reserved_balance,
+                        "total_spent": billing_key.total_spent,
+                        "finalize_type": "overrun",
+                    },
+                )
             else:
                 # Refund some of the base cost
                 refund = abs(cost_difference)
@@ -1299,6 +1355,35 @@ async def periodic_key_reset() -> None:
             logger.error(f"Error in periodic_key_reset: {e}")
 
 
+async def periodic_dead_key_prune() -> None:
+    """Periodically prune dead API keys. Interval <= 0 disables it.
+
+    See ``prune_dead_api_keys`` for eligibility.
+    """
+    from .core.db import create_session, prune_dead_api_keys
+
+    interval = settings.dead_key_prune_interval_seconds
+    if interval <= 0:
+        logger.info("Dead-key pruning disabled (interval <= 0)")
+        return
+
+    while True:
+        try:
+            await asyncio.sleep(interval)
+        except asyncio.CancelledError:
+            break
+
+        try:
+            async with create_session() as session:
+                await prune_dead_api_keys(
+                    session, settings.dead_key_min_age_seconds
+                )
+        except asyncio.CancelledError:
+            break
+        except Exception as e:
+            logger.error(f"Error in periodic_dead_key_prune: {e}")
+
+
 STALE_RESERVATION_SWEEP_INTERVAL_SECONDS: int = 60
 
 

routstr/balance.py

@@ -20,7 +20,14 @@ from .core.db import (
 from .core.logging import get_logger
 from .core.settings import settings
 from .lightning import lightning_router
-from .wallet import credit_balance, recieve_token, send_to_lnurl, send_token
+from .wallet import (
+    classify_redemption_error,
+    credit_balance,
+    is_mint_connection_error,
+    recieve_token,
+    send_to_lnurl,
+    send_token,
+)
 
 router = APIRouter()
 balance_router = APIRouter(prefix="/v1/balance")
@@ -156,30 +163,18 @@ async def topup_wallet_endpoint(
         raise HTTPException(status_code=400, detail="Invalid token format")
     try:
         amount_msats = await credit_balance(cashu_token, billing_key, session)
-    except ValueError as e:
-        error_msg = str(e)
-        if "already spent" in error_msg.lower():
-            raise HTTPException(status_code=400, detail="Token already spent")
-        elif "invalid" in error_msg.lower() or "decode" in error_msg.lower():
-            raise HTTPException(status_code=400, detail="Invalid token format")
-        elif "insufficient" in error_msg.lower() or "melt fee" in error_msg.lower():
-            raise HTTPException(
-                status_code=400,
-                detail=f"Token value is too small to cover swap fees. {error_msg}",
-            )
-        elif "failed to melt" in error_msg.lower():
-            raise HTTPException(
-                status_code=400,
-                detail=f"Failed to swap foreign mint token. {error_msg}",
-            )
-        else:
-            raise HTTPException(status_code=400, detail=f"Failed to redeem token: {error_msg}")
     except Exception as e:
-        logger.error(
-            "topup_wallet_endpoint: unhandled error",
-            extra={"error": str(e), "error_type": type(e).__name__},
-        )
-        raise HTTPException(status_code=500, detail="Internal server error")
+        # Shared taxonomy so top-up matches the bearer/X-Cashu paths (503 for an
+        # unreachable mint, 422 for fee/swap failures, 400 for token faults).
+        classified = classify_redemption_error(e)
+        if classified is None:
+            logger.error(
+                "topup_wallet_endpoint: unhandled error",
+                extra={"error": str(e), "error_type": type(e).__name__},
+            )
+            raise HTTPException(status_code=500, detail="Internal server error")
+        _type, status_code, message, _code = classified
+        raise HTTPException(status_code=status_code, detail=message)
     return {"msats": amount_msats}
 
 
@@ -274,7 +269,20 @@ async def refund_wallet_endpoint(
         )
         out_tx = out_tx_result.first()
         if out_tx is None:
-            raise HTTPException(status_code=404, detail="Refund not found")
+            # The "in" row exists with a request_id, but the "out" (refund)
+            # row hasn't been written yet — the upstream request is still in
+            # flight and the refund will be minted once it completes. Tell the
+            # client to retry instead of 404ing permanently (race condition
+            # where /v1/wallet/refund is polled before the refund exists).
+            logger.debug(
+                "refund_wallet_endpoint: refund pending (in row exists, out row not yet created)",
+                extra={"request_id": in_tx.request_id},
+            )
+            raise HTTPException(
+                status_code=425,
+                detail="Refund is pending; retry shortly.",
+                headers={"Retry-After": "2"},
+            )
         if out_tx.swept:
             raise HTTPException(status_code=410, detail="Refund has been swept")
 
@@ -441,14 +449,10 @@ async def refund_wallet_endpoint(
                 "has_refund_address": bool(key.refund_address),
             },
         )
-        if (
-            "mint" in error_msg.lower()
-            or "connection" in error_msg.lower()
-            or "ConnectError" in str(type(e))
-        ):
-            raise HTTPException(status_code=503, detail=f"Mint service unavailable: {error_msg}")
+        if is_mint_connection_error(e):
+            raise HTTPException(status_code=503, detail="Mint service unavailable")
         else:
-            raise HTTPException(status_code=500, detail=f"Refund failed: {error_msg}")
+            raise HTTPException(status_code=500, detail="Refund failed")
 
     await _refund_cache_set(bearer_value, result)
 

routstr/core/admin.py

@@ -1,5 +1,6 @@
 import asyncio
 import json
+import re
 import secrets
 from datetime import datetime, timezone
 from pathlib import Path
@@ -8,6 +9,7 @@ from fastapi import APIRouter, Depends, HTTPException, Query, Request
 from pydantic import BaseModel, RootModel
 from pydantic.v1 import ValidationError as PydanticValidationError
 from sqlmodel import select
+from sqlmodel.ext.asyncio.session import AsyncSession
 
 from ..payment.models import _row_to_model, list_models
 from ..proxy import refresh_model_maps, reinitialize_upstreams
@@ -26,9 +28,11 @@ from .db import (
     ModelRow,
     UpstreamProviderRow,
     create_session,
+    store_cashu_transaction,
 )
 from .log_manager import log_manager
 from .logging import get_logger
+from .provider_slugs import allocate_unique_provider_slug
 from .settings import SettingsService, settings
 
 logger = get_logger(__name__)
@@ -408,12 +412,11 @@ async def withdraw(
     # Get wallet and check balance
     from .settings import settings as global_settings
 
-    wallet = await get_wallet(
-        withdraw_request.mint_url or global_settings.primary_mint, withdraw_request.unit
-    )
+    effective_mint = withdraw_request.mint_url or global_settings.primary_mint
+    wallet = await get_wallet(effective_mint, withdraw_request.unit)
     proofs = get_proofs_per_mint_and_unit(
         wallet,
-        withdraw_request.mint_url or global_settings.primary_mint,
+        effective_mint,
         withdraw_request.unit,
         not_reserved=True,
     )
@@ -429,7 +432,16 @@ async def withdraw(
         raise HTTPException(status_code=400, detail="Insufficient wallet balance")
 
     token = await send_token(
-        withdraw_request.amount, withdraw_request.unit, withdraw_request.mint_url
+        withdraw_request.amount, withdraw_request.unit, effective_mint
+    )
+    await store_cashu_transaction(
+        token=token,
+        amount=withdraw_request.amount,
+        unit=withdraw_request.unit,
+        mint_url=effective_mint,
+        typ="out",
+        collected=False,
+        source="admin",
     )
     return {"token": token}
 
@@ -456,19 +468,18 @@ class ModelCreate(BaseModel):
     dependencies=[Depends(require_admin_api)],
 )
 async def upsert_provider_model(
-    provider_id: int, payload: ModelCreate
+    provider_id: str, payload: ModelCreate
 ) -> dict[str, object]:
     print(payload)
     logger.info(
         f"UPSERT_PROVIDER_MODEL called: provider_id={provider_id}, model_id={payload.id}"
     )
     async with create_session() as session:
-        provider = await session.get(UpstreamProviderRow, provider_id)
-        if not provider:
-            raise HTTPException(status_code=404, detail="Provider not found")
+        provider = await _get_upstream_provider_by_ref(session, provider_id)
+        provider_pk = _provider_pk(provider)
 
         # Try to get existing model
-        existing_row = await session.get(ModelRow, (payload.id, provider_id))
+        existing_row = await session.get(ModelRow, (payload.id, provider_pk))
 
         if existing_row:
             # Update existing model
@@ -524,7 +535,7 @@ async def upsert_provider_model(
                 alias_ids=(
                     json.dumps(payload.alias_ids) if payload.alias_ids else None
                 ),
-                upstream_provider_id=provider_id,
+                upstream_provider_id=provider_pk,
                 enabled=payload.enabled,
                 forwarded_model_id=payload.forwarded_model_id or payload.id,
             )
@@ -543,7 +554,7 @@ async def upsert_provider_model(
     dependencies=[Depends(require_admin_api)],
 )
 async def update_provider_model_legacy(
-    provider_id: int, model_id: str, payload: ModelCreate
+    provider_id: str, model_id: str, payload: ModelCreate
 ) -> dict[str, object]:
     """Legacy PATCH endpoint - redirects to upsert POST endpoint for backward compatibility."""
     logger.info(
@@ -556,13 +567,12 @@ async def update_provider_model_legacy(
     "/api/upstream-providers/{provider_id}/models/{model_id:path}",
     dependencies=[Depends(require_admin_api)],
 )
-async def get_provider_model(provider_id: int, model_id: str) -> dict[str, object]:
+async def get_provider_model(provider_id: str, model_id: str) -> dict[str, object]:
     async with create_session() as session:
-        provider = await session.get(UpstreamProviderRow, provider_id)
-        if not provider:
-            raise HTTPException(status_code=404, detail="Provider not found")
+        provider = await _get_upstream_provider_by_ref(session, provider_id)
+        provider_pk = _provider_pk(provider)
 
-        row = await session.get(ModelRow, (model_id, provider_id))
+        row = await session.get(ModelRow, (model_id, provider_pk))
         if not row:
             raise HTTPException(
                 status_code=404, detail="Model not found for this provider"
@@ -576,9 +586,11 @@ async def get_provider_model(provider_id: int, model_id: str) -> dict[str, objec
     "/api/upstream-providers/{provider_id}/models/{model_id:path}",
     dependencies=[Depends(require_admin_api)],
 )
-async def delete_provider_model(provider_id: int, model_id: str) -> dict[str, object]:
+async def delete_provider_model(provider_id: str, model_id: str) -> dict[str, object]:
     async with create_session() as session:
-        row = await session.get(ModelRow, (model_id, provider_id))
+        provider = await _get_upstream_provider_by_ref(session, provider_id)
+        provider_pk = _provider_pk(provider)
+        row = await session.get(ModelRow, (model_id, provider_pk))
         if not row:
             raise HTTPException(
                 status_code=404, detail="Model not found for this provider"
@@ -593,10 +605,12 @@ async def delete_provider_model(provider_id: int, model_id: str) -> dict[str, ob
     "/api/upstream-providers/{provider_id}/models",
     dependencies=[Depends(require_admin_api)],
 )
-async def delete_all_provider_models(provider_id: int) -> dict[str, object]:
+async def delete_all_provider_models(provider_id: str) -> dict[str, object]:
     async with create_session() as session:
+        provider = await _get_upstream_provider_by_ref(session, provider_id)
+        provider_pk = _provider_pk(provider)
         result = await session.exec(
-            select(ModelRow).where(ModelRow.upstream_provider_id == provider_id)
+            select(ModelRow).where(ModelRow.upstream_provider_id == provider_pk)
         )  # type: ignore
         rows = result.all()
         for row in rows:
@@ -615,7 +629,7 @@ class BatchOverrideRequest(BaseModel):
     dependencies=[Depends(require_admin_api)],
 )
 async def batch_override_provider_models(
-    provider_id: int, payload: BatchOverrideRequest
+    provider_id: str, payload: BatchOverrideRequest
 ) -> dict[str, object]:
     """Batch override models for a specific provider."""
     logger.info(
@@ -623,15 +637,14 @@ async def batch_override_provider_models(
     )
 
     async with create_session() as session:
-        provider = await session.get(UpstreamProviderRow, provider_id)
-        if not provider:
-            raise HTTPException(status_code=404, detail="Provider not found")
+        provider = await _get_upstream_provider_by_ref(session, provider_id)
+        provider_pk = _provider_pk(provider)
 
         overridden_count = 0
 
         for model_data in payload.models:
             # Try to get existing model regardless of whether it's enabled or not
-            existing_row = await session.get(ModelRow, (model_data.id, provider_id))
+            existing_row = await session.get(ModelRow, (model_data.id, provider_pk))
 
             if existing_row:
                 # Update existing
@@ -685,7 +698,7 @@ async def batch_override_provider_models(
                         if model_data.alias_ids
                         else None
                     ),
-                    upstream_provider_id=provider_id,
+                    upstream_provider_id=provider_pk,
                     enabled=model_data.enabled,
                 )
                 session.add(row)
@@ -702,6 +715,85 @@ async def batch_override_provider_models(
     }
 
 
+_SLUG_PATTERN = re.compile(r"^[a-z0-9][a-z0-9-]{1,62}[a-z0-9]$")
+
+
+def _validate_slug(value: str) -> str:
+    candidate = value.strip().lower()
+    if not _SLUG_PATTERN.fullmatch(candidate):
+        raise HTTPException(
+            status_code=400,
+            detail=(
+                "slug must be 3-64 chars, lowercase letters/digits/hyphens, "
+                "and may not start or end with a hyphen"
+            ),
+        )
+    if candidate.isdigit():
+        raise HTTPException(
+            status_code=400,
+            detail="slug must not be all digits",
+        )
+    return candidate
+
+
+async def _ensure_unique_slug(
+    session: AsyncSession, slug: str, exclude_id: int | None = None
+) -> None:
+    stmt = select(UpstreamProviderRow).where(UpstreamProviderRow.slug == slug)
+    result = await session.exec(stmt)
+    existing = result.first()
+    if existing and existing.id != exclude_id:
+        raise HTTPException(
+            status_code=409,
+            detail="Provider with this slug already exists",
+        )
+
+
+async def _get_upstream_provider_by_ref(
+    session: AsyncSession, provider_ref: str
+) -> UpstreamProviderRow:
+    if provider_ref.isdigit():
+        provider = await session.get(UpstreamProviderRow, int(provider_ref))
+    else:
+        slug = _validate_slug(provider_ref)
+        result = await session.exec(
+            select(UpstreamProviderRow).where(UpstreamProviderRow.slug == slug)
+        )
+        provider = result.first()
+
+    if not provider:
+        raise HTTPException(status_code=404, detail="Provider not found")
+    return provider
+
+
+def _provider_pk(provider: UpstreamProviderRow) -> int:
+    if provider.id is None:
+        raise HTTPException(status_code=500, detail="Provider has no database id")
+    return provider.id
+
+
+def _serialize_provider(
+    provider: UpstreamProviderRow, redact_api_key: bool = True
+) -> dict[str, object]:
+    return {
+        "id": provider.id,
+        "slug": provider.slug,
+        "provider_type": provider.provider_type,
+        "base_url": provider.base_url,
+        "api_key": "[REDACTED]"
+        if (redact_api_key and provider.api_key)
+        else provider.api_key
+        if not redact_api_key
+        else "",
+        "api_version": provider.api_version,
+        "enabled": provider.enabled,
+        "provider_fee": provider.provider_fee,
+        "provider_settings": json.loads(provider.provider_settings)
+        if provider.provider_settings
+        else None,
+    }
+
+
 class UpstreamProviderCreate(BaseModel):
     provider_type: str
     base_url: str
@@ -710,6 +802,7 @@ class UpstreamProviderCreate(BaseModel):
     enabled: bool = True
     provider_fee: float = 1.01
     provider_settings: dict | None = None
+    slug: str | None = None
 
 
 class UpstreamProviderUpdate(BaseModel):
@@ -720,6 +813,50 @@ class UpstreamProviderUpdate(BaseModel):
     enabled: bool | None = None
     provider_fee: float | None = None
     provider_settings: dict | None = None
+    slug: str | None = None
+
+
+class UpstreamProviderUpdateBySlug(BaseModel):
+    slug: str
+    new_slug: str | None = None
+    provider_type: str | None = None
+    base_url: str | None = None
+    api_key: str | None = None
+    api_version: str | None = None
+    enabled: bool | None = None
+    provider_fee: float | None = None
+    provider_settings: dict | None = None
+
+
+async def _apply_provider_update(
+    session: AsyncSession,
+    provider: UpstreamProviderRow,
+    payload: UpstreamProviderUpdate,
+    new_slug: str | None = None,
+) -> None:
+    if new_slug is not None:
+        validated = _validate_slug(new_slug)
+        await _ensure_unique_slug(session, validated, exclude_id=provider.id)
+        provider.slug = validated
+
+    if payload.provider_type is not None:
+        provider.provider_type = payload.provider_type
+    if payload.base_url is not None:
+        provider.base_url = payload.base_url
+    if payload.api_key is not None:
+        provider.api_key = payload.api_key
+    if payload.api_version is not None:
+        provider.api_version = payload.api_version
+    if payload.enabled is not None:
+        provider.enabled = payload.enabled
+    if payload.provider_fee is not None:
+        provider.provider_fee = payload.provider_fee
+    if payload.provider_settings is not None:
+        provider.provider_settings = json.dumps(payload.provider_settings)
+
+    session.add(provider)
+    await session.commit()
+    await session.refresh(provider)
 
 
 @admin_router.get("/api/upstream-providers", dependencies=[Depends(require_admin_api)])
@@ -727,21 +864,7 @@ async def get_upstream_providers() -> list[dict[str, object]]:
     async with create_session() as session:
         result = await session.exec(select(UpstreamProviderRow))
         providers = result.all()
-        return [
-            {
-                "id": p.id,
-                "provider_type": p.provider_type,
-                "base_url": p.base_url,
-                "api_key": "[REDACTED]" if p.api_key else "",
-                "api_version": p.api_version,
-                "enabled": p.enabled,
-                "provider_fee": p.provider_fee,
-                "provider_settings": json.loads(p.provider_settings)
-                if p.provider_settings
-                else None,
-            }
-            for p in providers
-        ]
+        return [_serialize_provider(p) for p in providers]
 
 
 @admin_router.post("/api/upstream-providers", dependencies=[Depends(require_admin_api)])
@@ -761,7 +884,14 @@ async def create_upstream_provider(
                 detail="Provider with this base URL and API key already exists",
             )
 
+        if payload.slug:
+            slug = _validate_slug(payload.slug)
+            await _ensure_unique_slug(session, slug)
+        else:
+            slug = await allocate_unique_provider_slug(session, payload.provider_type)
+
         provider = UpstreamProviderRow(
+            slug=slug,
             provider_type=payload.provider_type,
             base_url=payload.base_url,
             api_key=payload.api_key,
@@ -778,99 +908,81 @@ async def create_upstream_provider(
 
     await reinitialize_upstreams()
     await refresh_model_maps()
-    return {
-        "id": provider.id,
-        "provider_type": provider.provider_type,
-        "base_url": provider.base_url,
-        "api_key": "[REDACTED]",
-        "api_version": provider.api_version,
-        "enabled": provider.enabled,
-        "provider_fee": provider.provider_fee,
-        "provider_settings": payload.provider_settings,
-    }
+    return _serialize_provider(provider)
 
 
 @admin_router.get(
     "/api/upstream-providers/{provider_id}", dependencies=[Depends(require_admin_api)]
 )
-async def get_upstream_provider(provider_id: int) -> dict[str, object]:
+async def get_upstream_provider(provider_id: str) -> dict[str, object]:
     async with create_session() as session:
-        provider = await session.get(UpstreamProviderRow, provider_id)
-        if not provider:
-            raise HTTPException(status_code=404, detail="Provider not found")
-        return {
-            "id": provider.id,
-            "provider_type": provider.provider_type,
-            "base_url": provider.base_url,
-            "api_key": "[REDACTED]" if provider.api_key else "",
-            "api_version": provider.api_version,
-            "enabled": provider.enabled,
-            "provider_fee": provider.provider_fee,
-            "provider_settings": json.loads(provider.provider_settings)
-            if provider.provider_settings
-            else None,
-        }
+        provider = await _get_upstream_provider_by_ref(session, provider_id)
+        return _serialize_provider(provider)
 
 
 @admin_router.patch(
     "/api/upstream-providers/{provider_id}", dependencies=[Depends(require_admin_api)]
 )
 async def update_upstream_provider(
-    provider_id: int, payload: UpstreamProviderUpdate
+    provider_id: str, payload: UpstreamProviderUpdate
 ) -> dict[str, object]:
     async with create_session() as session:
-        provider = await session.get(UpstreamProviderRow, provider_id)
-        if not provider:
-            raise HTTPException(status_code=404, detail="Provider not found")
+        provider = await _get_upstream_provider_by_ref(session, provider_id)
 
-        if payload.provider_type is not None:
-            provider.provider_type = payload.provider_type
-        if payload.base_url is not None:
-            provider.base_url = payload.base_url
-        if payload.api_key is not None:
-            provider.api_key = payload.api_key
-        if payload.api_version is not None:
-            provider.api_version = payload.api_version
-        if payload.enabled is not None:
-            provider.enabled = payload.enabled
-        if payload.provider_fee is not None:
-            provider.provider_fee = payload.provider_fee
-        if payload.provider_settings is not None:
-            provider.provider_settings = json.dumps(payload.provider_settings)
-
-        session.add(provider)
-        await session.commit()
-        await session.refresh(provider)
+        await _apply_provider_update(session, provider, payload, new_slug=payload.slug)
 
     await reinitialize_upstreams()
     await refresh_model_maps()
-    return {
-        "id": provider.id,
-        "provider_type": provider.provider_type,
-        "base_url": provider.base_url,
-        "api_key": "[REDACTED]",
-        "api_version": provider.api_version,
-        "enabled": provider.enabled,
-        "provider_fee": provider.provider_fee,
-        "provider_settings": json.loads(provider.provider_settings)
-        if provider.provider_settings
-        else None,
-    }
+    return _serialize_provider(provider)
+
+
+@admin_router.patch(
+    "/api/upstream-providers", dependencies=[Depends(require_admin_api)]
+)
+async def update_upstream_provider_by_slug(
+    payload: UpstreamProviderUpdateBySlug,
+) -> dict[str, object]:
+    lookup = _validate_slug(payload.slug)
+    async with create_session() as session:
+        result = await session.exec(
+            select(UpstreamProviderRow).where(
+                UpstreamProviderRow.slug == lookup
+            )
+        )
+        provider = result.first()
+        if not provider:
+            raise HTTPException(status_code=404, detail="Provider not found")
+
+        update_payload = UpstreamProviderUpdate(
+            provider_type=payload.provider_type,
+            base_url=payload.base_url,
+            api_key=payload.api_key,
+            api_version=payload.api_version,
+            enabled=payload.enabled,
+            provider_fee=payload.provider_fee,
+            provider_settings=payload.provider_settings,
+        )
+        await _apply_provider_update(
+            session, provider, update_payload, new_slug=payload.new_slug
+        )
+
+    await reinitialize_upstreams()
+    await refresh_model_maps()
+    return _serialize_provider(provider)
 
 
 @admin_router.delete(
     "/api/upstream-providers/{provider_id}", dependencies=[Depends(require_admin_api)]
 )
-async def delete_upstream_provider(provider_id: int) -> dict[str, object]:
+async def delete_upstream_provider(provider_id: str) -> dict[str, object]:
     async with create_session() as session:
-        provider = await session.get(UpstreamProviderRow, provider_id)
-        if not provider:
-            raise HTTPException(status_code=404, detail="Provider not found")
+        provider = await _get_upstream_provider_by_ref(session, provider_id)
+        deleted_id = _provider_pk(provider)
         await session.delete(provider)
         await session.commit()
     await reinitialize_upstreams()
     await refresh_model_maps()
-    return {"ok": True, "deleted_id": provider_id}
+    return {"ok": True, "deleted_id": deleted_id}
 
 
 @admin_router.get("/api/provider-types", dependencies=[Depends(require_admin_api)])
@@ -885,17 +997,16 @@ async def get_provider_types() -> list[dict[str, object]]:
     "/api/upstream-providers/{provider_id}/models",
     dependencies=[Depends(require_admin_api)],
 )
-async def get_provider_models(provider_id: int) -> dict[str, object]:
+async def get_provider_models(provider_id: str) -> dict[str, object]:
     from ..upstream.helpers import _instantiate_provider
 
     async with create_session() as session:
-        provider = await session.get(UpstreamProviderRow, provider_id)
-        if not provider:
-            raise HTTPException(status_code=404, detail="Provider not found")
+        provider = await _get_upstream_provider_by_ref(session, provider_id)
+        provider_pk = _provider_pk(provider)
 
         db_models = await list_models(
             session=session,
-            upstream_id=provider_id,
+            upstream_id=provider_pk,
             include_disabled=True,
             apply_fees=False,
         )
@@ -985,13 +1096,11 @@ class TopupTokenRequest(BaseModel):
     dependencies=[Depends(require_admin_api)],
 )
 async def topup_provider_with_token(
-    provider_id: int, payload: TopupTokenRequest
+    provider_id: str, payload: TopupTokenRequest
 ) -> dict:
     """Redeem a Cashu token for an upstream provider."""
     async with create_session() as session:
-        provider = await session.get(UpstreamProviderRow, provider_id)
-        if not provider:
-            raise HTTPException(status_code=404, detail="Provider not found")
+        provider = await _get_upstream_provider_by_ref(session, provider_id)
 
         import httpx
 
@@ -1022,15 +1131,13 @@ async def topup_provider_with_token(
     dependencies=[Depends(require_admin_api)],
 )
 async def initiate_provider_topup(
-    provider_id: int, payload: TopupRequest
+    provider_id: str, payload: TopupRequest
 ) -> dict[str, object]:
     """Initiate a Lightning Network top-up for the upstream provider account."""
     from ..upstream.helpers import _instantiate_provider
 
     async with create_session() as session:
-        provider = await session.get(UpstreamProviderRow, provider_id)
-        if not provider:
-            raise HTTPException(status_code=404, detail="Provider not found")
+        provider = await _get_upstream_provider_by_ref(session, provider_id)
 
         try:
             logger.info(
@@ -1150,15 +1257,13 @@ async def initiate_provider_topup(
     "/api/upstream-providers/{provider_id}/topup/{invoice_id}/status",
     dependencies=[Depends(require_admin_api)],
 )
-async def check_topup_status(provider_id: int, invoice_id: str) -> dict[str, object]:
+async def check_topup_status(provider_id: str, invoice_id: str) -> dict[str, object]:
     """Check the status of a Lightning Network top-up invoice."""
     from ..upstream.helpers import _instantiate_provider
     from ..upstream.ppqai import PPQAIUpstreamProvider
 
     async with create_session() as session:
-        provider = await session.get(UpstreamProviderRow, provider_id)
-        if not provider:
-            raise HTTPException(status_code=404, detail="Provider not found")
+        provider = await _get_upstream_provider_by_ref(session, provider_id)
 
         # For Routstr providers, proxy the status check
         if provider.provider_type == "routstr":
@@ -1205,14 +1310,12 @@ async def check_topup_status(provider_id: int, invoice_id: str) -> dict[str, obj
     "/api/upstream-providers/{provider_id}/balance",
     dependencies=[Depends(require_admin_api)],
 )
-async def get_provider_balance(provider_id: int) -> dict[str, object]:
+async def get_provider_balance(provider_id: str) -> dict[str, object]:
     """Get the current balance for an upstream provider account."""
     from ..upstream.helpers import _instantiate_provider
 
     async with create_session() as session:
-        provider = await session.get(UpstreamProviderRow, provider_id)
-        if not provider:
-            raise HTTPException(status_code=404, detail="Provider not found")
+        provider = await _get_upstream_provider_by_ref(session, provider_id)
 
         # For Routstr providers, proxy the balance check
         if provider.provider_type == "routstr":
@@ -1591,15 +1694,13 @@ async def get_lightning_invoices_api(
     "/api/upstream-providers/{provider_id}/routstr/refund",
     dependencies=[Depends(require_admin_api)],
 )
-async def refund_routstr_provider_balance(provider_id: int) -> dict[str, object]:
+async def refund_routstr_provider_balance(provider_id: str) -> dict[str, object]:
     """Refund balance from an upstream Routstr provider back to the local wallet."""
     from ..upstream.helpers import _instantiate_provider
     from ..upstream.routstr import RoutstrUpstreamProvider
 
     async with create_session() as session:
-        provider_row = await session.get(UpstreamProviderRow, provider_id)
-        if not provider_row:
-            raise HTTPException(status_code=404, detail="Provider not found")
+        provider_row = await _get_upstream_provider_by_ref(session, provider_id)
 
         if provider_row.provider_type != "routstr":
             raise HTTPException(

routstr/core/db.py

@@ -9,9 +9,10 @@ from typing import AsyncGenerator
 from alembic import command
 from alembic.config import Config
 from alembic.util.exc import CommandError
-from sqlalchemy import UniqueConstraint
+from sqlalchemy import UniqueConstraint, delete
 from sqlalchemy.exc import OperationalError
 from sqlalchemy.ext.asyncio.engine import create_async_engine
+from sqlalchemy.orm import aliased
 from sqlmodel import Field, Relationship, SQLModel, col, func, select, update
 from sqlmodel.ext.asyncio.session import AsyncSession
 
@@ -125,6 +126,63 @@ async def release_stale_reservations(
     return released
 
 
+async def prune_dead_api_keys(session: AsyncSession, min_age_seconds: int) -> int:
+    """Delete dead parentless API keys; return the count removed.
+
+    Dead = 0 balance/reservation/spend/requests, older than the grace period,
+    no parent, no children, no pending invoice. Cashu rows are unlinked (not
+    deleted) first to keep the audit trail.
+    """
+    cutoff = int(time.time()) - min_age_seconds
+
+    child = aliased(ApiKey)
+    has_children = (
+        select(child.hashed_key).where(
+            col(child.parent_key_hash) == col(ApiKey.hashed_key)
+        )
+    ).exists()
+    pending_invoice = (
+        select(LightningInvoice.id)
+        .where(col(LightningInvoice.api_key_hash) == col(ApiKey.hashed_key))
+        .where(col(LightningInvoice.status) == "pending")
+    ).exists()
+
+    eligible_hashes = (
+        select(ApiKey.hashed_key)
+        .where(col(ApiKey.balance) == 0)
+        .where(col(ApiKey.reserved_balance) == 0)
+        .where(col(ApiKey.total_spent) == 0)
+        .where(col(ApiKey.total_requests) == 0)
+        .where(col(ApiKey.parent_key_hash).is_(None))
+        .where(
+            (col(ApiKey.created_at).is_(None)) | (col(ApiKey.created_at) < cutoff)
+        )
+        .where(~pending_invoice)
+        .where(~has_children)
+    )
+
+    # Unlink transactions rather than cascade-deleting them, so the financial
+    # audit trail survives. The eligibility predicate is re-evaluated inside both
+    # statements so a key that gained balance mid-run is left untouched.
+    await session.exec(  # type: ignore[call-overload]
+        update(CashuTransaction)
+        .where(col(CashuTransaction.api_key_hashed_key).in_(eligible_hashes))
+        .values(api_key_hashed_key=None)
+    )
+
+    result = await session.exec(  # type: ignore[call-overload]
+        delete(ApiKey).where(col(ApiKey.hashed_key).in_(eligible_hashes))
+    )
+    await session.commit()
+
+    pruned = int(result.rowcount or 0)
+    logger.info(
+        "Pruned dead API keys",
+        extra={"pruned_keys": pruned, "min_age_seconds": min_age_seconds},
+    )
+    return pruned
+
+
 class ModelRow(SQLModel, table=True):  # type: ignore
     __tablename__ = "models"
     id: str = Field(primary_key=True)
@@ -229,7 +287,7 @@ async def store_cashu_transaction(
     created_at: int | None = None,
     source: str = "x-cashu",
     api_key_hashed_key: str | None = None,
-) -> None:
+) -> bool:
     try:
         async with create_session() as session:
             tx = CashuTransaction(
@@ -246,11 +304,13 @@ async def store_cashu_transaction(
             )
             session.add(tx)
             await session.commit()
+        return True
     except Exception as e:
         logger.warning(
             f"Failed to store cashu transaction: {e} (type={typ})",
             extra={"error": str(e), "type": typ},
         )
+        return False
 
 
 class UpstreamProviderRow(SQLModel, table=True):  # type: ignore
@@ -261,6 +321,12 @@ class UpstreamProviderRow(SQLModel, table=True):  # type: ignore
         ),
     )
     id: int | None = Field(default=None, primary_key=True)
+    slug: str | None = Field(
+        default=None,
+        unique=True,
+        index=True,
+        description="Stable external slug used for updates via API key.",
+    )
     provider_type: str = Field(
         description="Provider type: custom, openai, anthropic, azure, openrouter, etc."
     )
@@ -288,6 +354,8 @@ class RoutstrFee(SQLModel, table=True):  # type: ignore
     accumulated_msats: int = Field(default=0)
     total_paid_msats: int = Field(default=0)
     last_paid_at: int | None = Field(default=None)
+    payout_in_progress_msats: int = Field(default=0)
+    payout_started_at: int | None = Field(default=None)
 
 
 class CliToken(SQLModel, table=True):  # type: ignore
@@ -328,18 +396,42 @@ async def get_routstr_fee(session: AsyncSession) -> RoutstrFee:
     return fee
 
 
-async def reset_routstr_fee(session: AsyncSession, paid_msats: int) -> None:
+async def reset_routstr_fee(session: AsyncSession, paid_msats: int) -> bool:
+    """Checkpoint a fee payout before making the external payment."""
     stmt = (
         update(RoutstrFee)
         .where(col(RoutstrFee.id) == 1)
+        .where(col(RoutstrFee.payout_in_progress_msats) == 0)
+        .where(col(RoutstrFee.accumulated_msats) >= paid_msats)
         .values(
             accumulated_msats=RoutstrFee.accumulated_msats - paid_msats,
+            payout_in_progress_msats=paid_msats,
+            payout_started_at=int(time.time()),
+        )
+    )
+    result = await session.exec(stmt)  # type: ignore[call-overload]
+    await session.commit()
+    return result.rowcount == 1
+
+
+async def complete_routstr_fee_payout(
+    session: AsyncSession, paid_msats: int
+) -> bool:
+    """Mark a checkpointed payout complete after the external payment succeeds."""
+    stmt = (
+        update(RoutstrFee)
+        .where(col(RoutstrFee.id) == 1)
+        .where(col(RoutstrFee.payout_in_progress_msats) == paid_msats)
+        .values(
+            payout_in_progress_msats=0,
+            payout_started_at=None,
             total_paid_msats=RoutstrFee.total_paid_msats + paid_msats,
             last_paid_at=int(time.time()),
         )
     )
-    await session.exec(stmt)  # type: ignore[call-overload]
+    result = await session.exec(stmt)  # type: ignore[call-overload]
     await session.commit()
+    return result.rowcount == 1
 
 
 async def balances_for_mint_and_unit(

routstr/core/exceptions.py

@@ -7,11 +7,26 @@ logger = get_logger(__name__)
 
 
 class UpstreamError(Exception):
-    """Exception raised when an upstream provider fails."""
+    """Exception raised when an upstream provider fails.
 
-    def __init__(self, message: str, status_code: int = 502):
+    ``code`` carries a stable, machine-readable classification (e.g.
+    ``UPSTREAM_RATE_LIMIT``) so callers can distinguish failure kinds without
+    string-matching the message. ``details`` holds optional structured,
+    redaction-safe context. Both default to ``None`` for backwards
+    compatibility.
+    """
+
+    def __init__(
+        self,
+        message: str,
+        status_code: int = 502,
+        code: str | None = None,
+        details: dict[str, object] | None = None,
+    ):
         self.message = message
         self.status_code = status_code
+        self.code = code
+        self.details = details
         super().__init__(message)
 
 

routstr/core/logging.py

@@ -51,6 +51,8 @@ from pythonjsonlogger import jsonlogger
 from rich.console import Console
 from rich.logging import RichHandler
 
+from .redaction import redact_obj, redact_org_ids
+
 # Only use RichHandler when stdout is a real TTY. In non-TTY contexts
 # (docker logs, pipes, CI) Rich pads every line to width and wraps long
 # records, producing visually-empty trailing whitespace and split records.
@@ -180,6 +182,37 @@ class RequestIdFilter(logging.Filter):
         return True
 
 
+# Standard ``LogRecord`` attributes that are never user-supplied ``extra``
+# fields; skipped when redacting structured extras (``msg``/``message`` are
+# handled separately above).
+_NON_EXTRA_RECORD_ATTRS = frozenset(
+    {
+        "name",
+        "msg",
+        "args",
+        "levelname",
+        "levelno",
+        "pathname",
+        "filename",
+        "module",
+        "exc_info",
+        "exc_text",
+        "stack_info",
+        "lineno",
+        "funcName",
+        "created",
+        "msecs",
+        "relativeCreated",
+        "thread",
+        "threadName",
+        "processName",
+        "process",
+        "taskName",
+        "message",
+    }
+)
+
+
 class SecurityFilter(logging.Filter):
     """Filter to remove sensitive information from logs."""
 
@@ -203,6 +236,7 @@ class SecurityFilter(logging.Filter):
         """Filter out sensitive information from log records."""
         try:
             message = record.getMessage()
+            message = redact_org_ids(message)
             standalone_patterns = [
                 r"Bearer\s+([a-zA-Z0-9_\-\.]{10,})",  # Bearer token (must be 10 characters or more to reduce false-positives)
                 r"cashu[A-Z]+([a-zA-Z0-9_\-\.=/+]+)",  # Cashu tokens
@@ -224,6 +258,16 @@ class SecurityFilter(logging.Filter):
             record.msg = message
             record.args = ()
 
+            # Structured `extra={...}` fields are emitted by the JSON formatter
+            # straight from the record dict and never pass through the message
+            # formatting above. Redact organization IDs from any string-valued
+            # extra so they cannot leak via structured logs.
+            for attr, value in list(record.__dict__.items()):
+                if attr in _NON_EXTRA_RECORD_ATTRS:
+                    continue
+                if isinstance(value, (str, dict, list, tuple)):
+                    record.__dict__[attr] = redact_obj(value)
+
         except Exception:
             pass
 

routstr/core/main.py

@@ -11,7 +11,11 @@ from starlette.exceptions import HTTPException
 from starlette.responses import Response as StarletteResponse
 from starlette.types import Scope
 
-from ..auth import periodic_key_reset, periodic_stale_reservation_sweep
+from ..auth import (
+    periodic_dead_key_prune,
+    periodic_key_reset,
+    periodic_stale_reservation_sweep,
+)
 from ..balance import balance_router, deprecated_wallet_router
 from ..lightning import lightning_router, periodic_invoice_watcher
 from ..nostr import (
@@ -24,6 +28,7 @@ from ..payment.models import models_router, update_sats_pricing
 from ..payment.price import update_prices_periodically
 from ..proxy import initialize_upstreams, proxy_router, refresh_model_maps_periodically
 from ..upstream.auto_topup import periodic_auto_topup
+from ..upstream.deepseek_v4_pricing_shim import register_deepseek_v4_pricing
 from ..upstream.litellm_routing import configure_litellm
 from ..wallet import periodic_payout, periodic_refund_sweep, periodic_routstr_fee_payout
 from .admin import admin_router
@@ -55,6 +60,7 @@ async def lifespan(_: FastAPI) -> AsyncGenerator[None, None]:
     model_maps_refresh_task = None
     key_reset_task = None
     stale_reservation_task = None
+    dead_key_prune_task = None
     auto_topup_task = None
     refund_sweep_task = None
     routstr_fee_task = None
@@ -65,6 +71,11 @@ async def lifespan(_: FastAPI) -> AsyncGenerator[None, None]:
         # debug logging) before any upstream provider dispatches a request.
         configure_litellm()
 
+        # TEMPORARY: backfill DeepSeek V4 pricing missing from litellm's cost
+        # map (BerriAI/litellm#30430). Remove this call and
+        # deepseek_v4_pricing_shim.py once litellm ships these models.
+        register_deepseek_v4_pricing()
+
         # Run database migrations on startup
         run_migrations()
 
@@ -126,6 +137,7 @@ async def lifespan(_: FastAPI) -> AsyncGenerator[None, None]:
         stale_reservation_task = asyncio.create_task(
             periodic_stale_reservation_sweep()
         )
+        dead_key_prune_task = asyncio.create_task(periodic_dead_key_prune())
         auto_topup_task = asyncio.create_task(periodic_auto_topup())
         refund_sweep_task = asyncio.create_task(periodic_refund_sweep())
         routstr_fee_task = asyncio.create_task(periodic_routstr_fee_payout())
@@ -165,6 +177,8 @@ async def lifespan(_: FastAPI) -> AsyncGenerator[None, None]:
             key_reset_task.cancel()
         if stale_reservation_task is not None:
             stale_reservation_task.cancel()
+        if dead_key_prune_task is not None:
+            dead_key_prune_task.cancel()
         if auto_topup_task is not None:
             auto_topup_task.cancel()
         if refund_sweep_task is not None:
@@ -196,6 +210,8 @@ async def lifespan(_: FastAPI) -> AsyncGenerator[None, None]:
                 tasks_to_wait.append(key_reset_task)
             if stale_reservation_task is not None:
                 tasks_to_wait.append(stale_reservation_task)
+            if dead_key_prune_task is not None:
+                tasks_to_wait.append(dead_key_prune_task)
             if auto_topup_task is not None:
                 tasks_to_wait.append(auto_topup_task)
             if refund_sweep_task is not None:
@@ -241,7 +257,20 @@ app.add_middleware(
     allow_credentials=True,
     allow_methods=["*"],
     allow_headers=["*"],
-    expose_headers=["x-routstr-request-id", "x-cashu"],
+    expose_headers=[
+        "x-routstr-request-id",
+        "x-cashu",
+        "x-routstr-cost-msats",
+        "x-routstr-cost-usd",
+        "x-routstr-input-cost-msats",
+        "x-routstr-output-cost-msats",
+        # EHBP (Tinfoil) protocol headers must be exposed so browser clients
+        # can detect and decrypt encrypted responses. Without these, the
+        # browser hides them via CORS and the SDK treats the response as a
+        # plaintext proxy error, returning raw ciphertext.
+        "Ehbp-Response-Nonce",
+        "Ehbp-Encapsulated-Key",
+    ],
 )
 
 # Add logging middleware
@@ -363,7 +392,10 @@ if UI_DIST_PATH.exists() and UI_DIST_PATH.is_dir():
 
 else:
     logger.warning(
-        f"UI dist directory not found at {UI_DIST_PATH}, skipping static file serving"
+        "UI dist directory not found at %s; serving API only. Run `make ui-build` "
+        "to build the static UI served from here, or `make ui-dev` for the Next.js "
+        "dev server with hot reload on :3000 (it targets this backend on :8000).",
+        UI_DIST_PATH,
     )
 
     @app.get("/", include_in_schema=False)

routstr/core/provider_slugs.py

@@ -0,0 +1,65 @@
+from __future__ import annotations
+
+import re
+from itertools import count
+from typing import Collection
+
+from sqlmodel import select
+from sqlmodel.ext.asyncio.session import AsyncSession
+
+from .db import UpstreamProviderRow
+
+_SLUG_BASE_PATTERN = re.compile(r"[^a-z0-9]+")
+_MAX_SLUG_LENGTH = 64
+
+
+def provider_slug_base(provider_type: str) -> str:
+    """Return a deterministic slug base for a provider type."""
+    base = _SLUG_BASE_PATTERN.sub("-", provider_type.lower()).strip("-")
+    if not base:
+        base = "provider"
+    elif base.isdigit():
+        base = f"provider-{base}"
+    elif len(base) < 3:
+        base = f"{base}-provider"
+
+    if len(base) > _MAX_SLUG_LENGTH:
+        base = base[:_MAX_SLUG_LENGTH].rstrip("-") or "provider"
+    return base
+
+
+def provider_slug_candidate(base: str, suffix_number: int) -> str:
+    if suffix_number == 1:
+        return base
+
+    suffix = f"-{suffix_number}"
+    max_base_length = _MAX_SLUG_LENGTH - len(suffix)
+    return f"{base[:max_base_length].rstrip('-')}{suffix}"
+
+
+async def allocate_unique_provider_slug(
+    session: AsyncSession,
+    provider_type: str,
+    reserved_slugs: Collection[str] = (),
+) -> str:
+    """Allocate a stable, deterministic provider slug.
+
+    The first provider of a type gets ``openai``; later collisions get
+    ``openai-2``, ``openai-3``, etc. ``reserved_slugs`` covers rows staged in
+    memory but not flushed yet, such as settings/env seeding.
+    """
+    base = provider_slug_base(provider_type)
+    reserved = {slug.lower() for slug in reserved_slugs}
+
+    for suffix_number in count(1):
+        candidate = provider_slug_candidate(base, suffix_number)
+        if candidate in reserved:
+            continue
+
+        result = await session.exec(
+            select(UpstreamProviderRow).where(UpstreamProviderRow.slug == candidate)
+        )
+        if result.first() is None:
+            return candidate
+
+    raise RuntimeError("unreachable")

routstr/core/redaction.py

@@ -0,0 +1,51 @@
+"""Redaction helpers for sensitive provider identifiers.
+
+Single source of truth for stripping account-scoped identifiers (e.g. OpenAI
+organization IDs) from any text before it is logged, returned to a caller, or
+written to an audit entry.
+"""
+
+from __future__ import annotations
+
+import re
+from typing import Any
+
+# OpenAI-style organization identifiers look like ``org-<base62>``. Require at
+# least 6 trailing chars so the already-redacted literal ``org-[REDACTED]`` is
+# never re-matched (``[`` is not in the character class).
+_ORG_ID_PATTERN = re.compile(r"\borg-[A-Za-z0-9]{6,}\b")
+
+ORG_ID_PLACEHOLDER = "org-[REDACTED]"
+
+
+def redact_org_ids(text: str) -> str:
+    """Replace OpenAI-style organization IDs with ``org-[REDACTED]``.
+
+    Args:
+        text: Arbitrary text that may embed an ``org-*`` identifier.
+
+    Returns:
+        The text with every organization ID replaced. Non-string input is
+        returned unchanged after coercion to ``str``.
+    """
+    if not text:
+        return text
+    return _ORG_ID_PATTERN.sub(ORG_ID_PLACEHOLDER, text)
+
+
+def redact_obj(obj: Any) -> Any:
+    """Recursively redact organization IDs in arbitrary nested structures.
+
+    Strings are redacted in place; dicts and lists/tuples are walked so that
+    identifiers nested inside structured payloads (e.g. log ``extra`` fields or
+    error ``details``) are also stripped. Other types are returned unchanged.
+    """
+    if isinstance(obj, str):
+        return redact_org_ids(obj)
+    if isinstance(obj, dict):
+        return {key: redact_obj(value) for key, value in obj.items()}
+    if isinstance(obj, list):
+        return [redact_obj(value) for value in obj]
+    if isinstance(obj, tuple):
+        return tuple(redact_obj(value) for value in obj)
+    return obj

routstr/core/settings.py

@@ -73,6 +73,14 @@ class Settings(BaseSettings):
     stale_reservation_timeout_seconds: int = Field(
         default=300, env="STALE_RESERVATION_TIMEOUT_SECONDS"
     )
+    # Background prune of dead (zero balance, never used) API keys.
+    # Interval 0 disables it; min-age is a grace period (default 1 week).
+    dead_key_prune_interval_seconds: int = Field(
+        default=3600, env="DEAD_KEY_PRUNE_INTERVAL_SECONDS"
+    )
+    dead_key_min_age_seconds: int = Field(
+        default=604_800, env="DEAD_KEY_MIN_AGE_SECONDS"
+    )
 
     # Network
     cors_origins: list[str] = Field(default_factory=lambda: ["*"], env="CORS_ORIGINS")

routstr/core/version.py

@@ -20,7 +20,7 @@ import subprocess
 from functools import lru_cache
 from pathlib import Path
 
-BASE_VERSION = "0.4.3"
+BASE_VERSION = "0.4.4"
 
 _REPO_ROOT = Path(__file__).resolve().parents[2]
 _GIT_TIMEOUT_SECONDS = 2.0

routstr/payment/cost_calculation.py

@@ -3,9 +3,17 @@ import math
 from pydantic.v1 import BaseModel
 
 from ..core import get_logger
-from ..core.db import AsyncSession
 from ..core.settings import settings
 from .price import sats_usd_price
+from .usage import normalize_usage, parse_token_count
+
+__all__ = [
+    "CostData",
+    "CostDataError",
+    "MaxCostData",
+    "calculate_cost",
+    "parse_token_count",
+]
 
 logger = get_logger(__name__)
 
@@ -33,8 +41,31 @@ class CostDataError(BaseModel):
     code: str
 
 
+def _empty_cost(cls: type[CostData] = CostData) -> CostData:
+    """Build an all-zero cost object — a full refund for an empty response.
+
+    Shared by the two paths that must not bill: an upstream response with no
+    usage data at all, and one that reports a USD cost but carries zero tokens
+    in every bucket.
+    """
+    return cls(
+        base_msats=0,
+        input_msats=0,
+        output_msats=0,
+        total_msats=0,
+        total_usd=0.0,
+        input_tokens=0,
+        output_tokens=0,
+        cache_read_input_tokens=0,
+        cache_creation_input_tokens=0,
+        cache_read_msats=0,
+        cache_creation_msats=0,
+    )
+
+
 async def calculate_cost(
-    response_data: dict, max_cost: int, session: AsyncSession
+    response_data: dict,
+    max_cost: int,
 ) -> CostData | MaxCostData | CostDataError:
     """Calculate the cost of an API request based on token usage.
 
@@ -44,6 +75,9 @@ async def calculate_cost(
 
     Returns:
         Cost data or error information
+
+    The response's usage object is normalized with the default union parser;
+    this function holds no vendor-dialect knowledge of its own.
     """
     logger.debug(
         "Starting cost calculation",
@@ -54,8 +88,9 @@ async def calculate_cost(
         },
     )
 
-    # Check for usage data
-    if "usage" not in response_data or response_data["usage"] is None:
+    usage = normalize_usage(response_data.get("usage"))
+
+    if usage is None:
         logger.warning(
             "No usage data in response — billing at MaxCostData with zero "
             "tokens. Dashboard will show this request as `(0+0)`. Most "
@@ -70,34 +105,41 @@ async def calculate_cost(
                 else None,
             },
         )
-        return MaxCostData(
-            base_msats=0,
-            input_msats=0,
-            output_msats=0,
-            total_msats=0,
-            total_usd=0.0,
-            input_tokens=0,
-            output_tokens=0,
-            cache_read_input_tokens=0,
-            cache_creation_input_tokens=0,
-            cache_read_msats=0,
-            cache_creation_msats=0,
-        )
+        return _empty_cost(MaxCostData)
 
-    usage_data = response_data["usage"]
+    usage_data = response_data.get("usage") or {}
+    if not isinstance(usage_data, dict):
+        usage_data = {}
 
-    # Extract token counts
-    input_tokens = _extract_token_pair(usage_data, "prompt_tokens", "input_tokens")
-    output_tokens = _extract_token_pair(usage_data, "completion_tokens", "output_tokens")
-
-    # Extract cache tokens (handles OpenAI vs Anthropic formats)
-    cache_read_tokens, cache_creation_tokens, input_tokens = _extract_cache_tokens(
-        usage_data, input_tokens
-    )
+    input_tokens = usage.input_tokens
+    output_tokens = usage.output_tokens
+    cache_read_tokens = usage.cache_read_tokens
+    cache_creation_tokens = usage.cache_write_tokens
 
     # Try USD cost first
     usd_cost = _resolve_usd_cost(usage_data, response_data)
     if usd_cost > 0:
+        truly_empty = (
+            input_tokens == 0
+            and output_tokens == 0
+            and cache_read_tokens == 0
+            and cache_creation_tokens == 0
+        )
+        if truly_empty:
+            logger.warning(
+                "Upstream reported a USD cost but the response carries no "
+                "tokens at all (input, output, cache-read and cache-creation "
+                "are all zero) — refunding in full rather than billing the "
+                "USD-derived cost for an empty response.",
+                extra={
+                    "model": response_data.get("model", "unknown"),
+                    "usd_cost": usd_cost,
+                    "usage_keys": sorted(usage_data.keys())
+                    if isinstance(usage_data, dict)
+                    else None,
+                },
+            )
+            return _empty_cost()
         if input_tokens == 0 and output_tokens == 0:
             logger.warning(
                 "Upstream reported a USD cost but no token counts — "
@@ -115,11 +157,16 @@ async def calculate_cost(
                 },
             )
         try:
+            cost_details = usage_data.get("cost_details", {})
+            if not isinstance(cost_details, dict):
+                cost_details = {}
             input_usd = _coerce_usd(
-                usage_data.get("cost_details", {}).get("input_cost", 0)
+                cost_details.get("input_cost")
+                or cost_details.get("upstream_inference_prompt_cost")
             )
             output_usd = _coerce_usd(
-                usage_data.get("cost_details", {}).get("output_cost", 0)
+                cost_details.get("output_cost")
+                or cost_details.get("upstream_inference_completions_cost")
             )
             return _calculate_from_usd_cost(
                 usd_cost,
@@ -202,22 +249,6 @@ async def calculate_cost(
 # ============================================================================
 
 
-def parse_token_count(value: object) -> int:
-    """Parse a token count from various formats (int, float, str, bool)."""
-    if isinstance(value, bool):
-        return 0
-    if isinstance(value, int):
-        return max(0, value)
-    if isinstance(value, float):
-        return max(0, int(value))
-    if isinstance(value, str):
-        try:
-            return max(0, int(float(value)))
-        except ValueError:
-            return 0
-    return 0
-
-
 def _coerce_usd(value: object) -> float:
     """Coerce a value to USD float, handling various formats safely."""
     if value is None or isinstance(value, bool):
@@ -230,41 +261,20 @@ def _coerce_usd(value: object) -> float:
         return 0.0
 
 
-def _extract_token_pair(
-    usage_data: dict, standard_field: str, alt_field: str
-) -> int:
-    """Extract token count trying two field names in order."""
-    value = parse_token_count(usage_data.get(standard_field, 0))
-    if value > 0:
-        return value
-    return parse_token_count(usage_data.get(alt_field, 0))
-
-
-def _extract_cache_tokens(usage_data: dict, input_tokens: int) -> tuple[int, int, int]:
-    """Extract cache tokens, handling OpenAI vs Anthropic formats.
-
-    Returns: (cache_read_tokens, cache_creation_tokens, adjusted_input_tokens)
-    """
-    cache_read = parse_token_count(usage_data.get("cache_read_input_tokens", 0))
-    cache_creation = parse_token_count(
-        usage_data.get("cache_creation_input_tokens", 0)
-    )
-
-    # OpenAI: cache is included in input_tokens, subtract it
-    prompt_details = usage_data.get("prompt_tokens_details")
-    if isinstance(prompt_details, dict) and not cache_read:
-        openai_cached = parse_token_count(prompt_details.get("cached_tokens", 0))
-        if openai_cached:
-            cache_read = openai_cached
-            input_tokens = max(0, input_tokens - cache_read)
-
-    return cache_read, cache_creation, input_tokens
-
-
 def _resolve_usd_cost(usage_data: dict, response_data: dict) -> float:
     """Resolve USD cost with clear priority order.
 
-    Priority: cost_details.total_cost → total_cost → cost (in both usage and response).
+    Priority:
+
+    1. ``cost_details.total_cost``
+    2. ``cost_details.upstream_inference_cost`` (BYOK — see below)
+    3. ``total_cost`` → ``cost`` (in both usage and response)
+
+    **BYOK path (PPQ.AI):** when ``is_byok`` is true the ``usage.cost`` field
+    is only a small (~5 %) routing fee, not the inference cost.  The real cost
+    lives in ``cost_details.upstream_inference_cost`` and the provider's
+    balance is debited by ``upstream_inference_cost + byok_fee``.  Billing just
+    the fee under-charges by ~20×.
     """
     cost_details = usage_data.get("cost_details")
     if isinstance(cost_details, dict):
@@ -272,6 +282,18 @@ def _resolve_usd_cost(usage_data: dict, response_data: dict) -> float:
         if cost > 0:
             return cost
 
+        # PPQ.AI BYOK: upstream_inference_cost is the real inference cost;
+        # usage.cost is only a ~5 % BYOK routing fee.  Bill the sum — what PPQ
+        # actually deducts from the balance.  For non-BYOK providers (e.g.
+        # OpenRouter) usage.cost already equals upstream_inference_cost, so we
+        # fall through to the normal ``cost`` lookup below.
+        upstream_cost = _coerce_usd(
+            cost_details.get("upstream_inference_cost")
+        )
+        if upstream_cost > 0 and usage_data.get("is_byok"):
+            byok_fee = _coerce_usd(usage_data.get("cost"))
+            return upstream_cost + byok_fee
+
     for source in [usage_data, response_data]:
         if not isinstance(source, dict):
             continue
@@ -286,54 +308,92 @@ def _resolve_usd_cost(usage_data: dict, response_data: dict) -> float:
 def _get_pricing_rates(
     response_data: dict,
 ) -> tuple[float, float, float, float] | None:
-    """Get model-based pricing rates or None if using fixed pricing.
+    """Get configured rates, falling back to LiteLLM's model cost map.
 
-    Returns: (input_rate, output_rate, cache_read_rate, cache_write_rate)
+    Returns: (input_rate, output_rate, cache_read_rate, cache_write_rate).
+    ``None`` means configured fixed pricing should be used by the caller.
     """
-    if settings.fixed_pricing:
+    if settings.fixed_pricing and (
+        settings.fixed_per_1k_input_tokens
+        or settings.fixed_per_1k_output_tokens
+    ):
         return None
 
     from ..proxy import get_model_instance
+    from .models import litellm_cost_entry
 
     response_model = response_data.get("model", "")
     model_obj = get_model_instance(response_model)
 
-    if not model_obj:
-        logger.error("Invalid model in response", extra={"response_model": response_model})
-        raise ValueError(f"Invalid model: {response_model}")
+    if model_obj and model_obj.sats_pricing:
+        try:
+            mspp = float(model_obj.sats_pricing.prompt)
+            mspc = float(model_obj.sats_pricing.completion)
+            mscr = float(model_obj.sats_pricing.input_cache_read or 0)
+            mscw = float(model_obj.sats_pricing.input_cache_write or 0)
 
-    if not model_obj.sats_pricing:
-        logger.error(
-            "Model pricing not defined",
-            extra={"model": response_model, "model_id": response_model},
+            mspp_1k = mspp * 1_000_000.0
+            mspc_1k = mspc * 1_000_000.0
+            mscr_1k = mscr * 1_000_000.0 if mscr > 0 else mspp_1k
+            mscw_1k = mscw * 1_000_000.0 if mscw > 0 else mspp_1k
+            source = "configured"
+        except Exception as e:
+            logger.error("Invalid pricing data", extra={"error": str(e)})
+            raise ValueError("Invalid pricing data") from e
+    else:
+        pricing_model = (
+            model_obj.forwarded_model_id if model_obj else None
+        ) or response_model
+        pricing = litellm_cost_entry(pricing_model)
+        if pricing is None:
+            logger.error(
+                "Model pricing not found in configured models or LiteLLM",
+                extra={
+                    "response_model": response_model,
+                    "pricing_model": pricing_model,
+                },
+            )
+            raise ValueError(f"Pricing not found for model: {response_model}")
+
+        input_usd = _coerce_usd(pricing.get("input_cost_per_token"))
+        output_usd = _coerce_usd(pricing.get("output_cost_per_token"))
+        if input_usd <= 0 or output_usd <= 0:
+            raise ValueError(f"Incomplete LiteLLM pricing for model: {pricing_model}")
+
+        provider_fee = _resolve_provider_fee(response_model)
+        usd_per_sat = sats_usd_price()
+        mspp_1k = input_usd * provider_fee * 1_000_000.0 / usd_per_sat
+        mspc_1k = output_usd * provider_fee * 1_000_000.0 / usd_per_sat
+        cache_read_usd = _coerce_usd(
+            pricing.get("cache_read_input_token_cost")
         )
-        raise ValueError("Model pricing not defined")
-
-    try:
-        mspp = float(model_obj.sats_pricing.prompt)
-        mspc = float(model_obj.sats_pricing.completion)
-        mscr = float(model_obj.sats_pricing.input_cache_read or 0)
-        mscw = float(model_obj.sats_pricing.input_cache_write or 0)
-
-        mspp_1k = mspp * 1_000_000.0
-        mspc_1k = mspc * 1_000_000.0
-        mscr_1k = mscr * 1_000_000.0 if mscr > 0 else mspp_1k
-        mscw_1k = mscw * 1_000_000.0 if mscw > 0 else mspp_1k
-
-        logger.info(
-            "Applied model-specific pricing",
-            extra={
-                "model": response_model,
-                "input_price_msats_per_1k": mspp_1k,
-                "output_price_msats_per_1k": mspc_1k,
-                "cache_read_price_msats_per_1k": mscr_1k,
-                "cache_write_price_msats_per_1k": mscw_1k,
-            },
+        cache_write_usd = _coerce_usd(
+            pricing.get("cache_creation_input_token_cost")
         )
-        return mspp_1k, mspc_1k, mscr_1k, mscw_1k
-    except Exception as e:
-        logger.error("Invalid pricing data", extra={"error": str(e)})
-        raise ValueError("Invalid pricing data") from e
+        mscr_1k = (
+            cache_read_usd * provider_fee * 1_000_000.0 / usd_per_sat
+            if cache_read_usd > 0
+            else mspp_1k
+        )
+        mscw_1k = (
+            cache_write_usd * provider_fee * 1_000_000.0 / usd_per_sat
+            if cache_write_usd > 0
+            else mspp_1k
+        )
+        source = "litellm"
+
+    logger.info(
+        "Applied model-specific pricing",
+        extra={
+            "model": response_model,
+            "pricing_source": source,
+            "input_price_msats_per_1k": mspp_1k,
+            "output_price_msats_per_1k": mspc_1k,
+            "cache_read_price_msats_per_1k": mscr_1k,
+            "cache_write_price_msats_per_1k": mscw_1k,
+        },
+    )
+    return mspp_1k, mspc_1k, mscr_1k, mscw_1k
 
 
 def _resolve_provider_fee(model_id: str) -> float:
@@ -372,8 +432,12 @@ def _calculate_from_usd_cost(
     cost_in_msats = math.ceil(cost_in_sats * 1000)
 
     if input_usd > 0 or output_usd > 0:
-        input_msats = int((input_usd * sats_per_usd) * 1000)
-        output_msats = int((output_usd * sats_per_usd) * 1000)
+        # The total is the authoritative billed amount. Allocating that integer
+        # total proportionally avoids losing sub-millisatoshi remainders when
+        # input and output components are each truncated independently.
+        component_usd = input_usd + output_usd
+        input_msats = math.floor(cost_in_msats * input_usd / component_usd)
+        output_msats = cost_in_msats - input_msats
     else:
         effective_input_tokens = (
             input_tokens + cache_read_tokens + cache_creation_tokens
@@ -454,10 +518,21 @@ def _calculate_from_tokens(
         },
     )
 
+    # Fold the cache-read/write cost into the visible ``input_msats`` so a
+    # dashboard that renders I / O / T sees ``input + output == total``
+    # exactly. This mirrors ``_fold_cache_into_input_tokens`` (which rolls the
+    # cache token counts into the visible prompt total). The standalone
+    # ``cache_read_msats`` / ``cache_creation_msats`` fields stay populated for
+    # clients that want the breakdown; nothing sums the components to derive
+    # ``total_msats`` (it is computed independently above), so this is
+    # display-only and does not change what is billed.
+    visible_output_msats = int(calc_output_msats)
+    visible_input_msats = token_based_cost - visible_output_msats
+
     return CostData(
         base_msats=0,
-        input_msats=int(calc_input_msats),
-        output_msats=int(calc_output_msats),
+        input_msats=visible_input_msats,
+        output_msats=visible_output_msats,
         total_msats=token_based_cost,
         total_usd=total_usd,
         input_tokens=input_tokens,

routstr/payment/helpers.py

@@ -11,6 +11,8 @@ from PIL import Image
 from sqlmodel.ext.asyncio.session import AsyncSession
 
 from ..core import get_logger
+from ..core.exceptions import UpstreamError
+from ..core.redaction import redact_org_ids
 from ..core.settings import settings
 from ..wallet import deserialize_token_from_string
 
@@ -418,16 +420,27 @@ def create_error_response(
     status_code: int,
     request: Request,
     token: str | None = None,
+    code: str | int | None = None,
+    details: dict[str, object] | None = None,
 ) -> Response:
-    """Create a standardized error response."""
+    """Create a standardized error response.
+
+    ``code`` is a stable, machine-readable classification (e.g.
+    ``UPSTREAM_RATE_LIMIT``); when omitted it defaults to the HTTP status code
+    for backwards compatibility. ``details`` carries optional structured,
+    redaction-safe context.
+    """
+    error_obj: dict[str, object] = {
+        "message": redact_org_ids(message),
+        "type": error_type,
+        "code": code if code is not None else status_code,
+    }
+    if details is not None:
+        error_obj["details"] = details
     return Response(
         content=json.dumps(
             {
-                "error": {
-                    "message": message,
-                    "type": error_type,
-                    "code": status_code,
-                },
+                "error": error_obj,
                 "request_id": getattr(request.state, "request_id", "unknown"),
             }
         ),
@@ -435,3 +448,20 @@ def create_error_response(
         media_type="application/json",
         headers={"X-Cashu": token} if token else {},
     )
+
+
+def create_upstream_error_response(
+    error: UpstreamError,
+    request: Request,
+    fallback_status: int = 502,
+) -> Response:
+    """Build an error response from an :class:`UpstreamError`, preserving its
+    structured ``code``, ``details``, and original ``status_code``."""
+    return create_error_response(
+        "upstream_error",
+        str(error),
+        error.status_code or fallback_status,
+        request=request,
+        code=getattr(error, "code", None),
+        details=getattr(error, "details", None),
+    )

routstr/payment/lnurl.py

@@ -1,11 +1,17 @@
 from __future__ import annotations
 
+import asyncio
 import math
 from typing import TypedDict
 
 import httpx
 from cashu.wallet.wallet import Proof, Wallet
 
+# The Cashu library issues POST /v1/melt/bolt11 with timeout=None, so a hung or
+# very slow mint can block a melt (and any caller, e.g. the payout loop)
+# indefinitely. Bound it here so callers fail instead of hanging forever.
+MELT_TIMEOUT_SECONDS = 60
+
 try:
     from bech32 import bech32_decode, convertbits  # type: ignore
 except ModuleNotFoundError:  # pragma: no cover – allow runtime miss
@@ -220,10 +226,18 @@ async def raw_send_to_lnurl(
     if amount:
         proofs, _ = await wallet.select_to_send(proofs, amount, set_reserved=True)
 
-    _ = await wallet.melt(
-        proofs=proofs,
-        invoice=bolt11_invoice,
-        fee_reserve_sat=melt_quote_resp.fee_reserve,
-        quote_id=melt_quote_resp.quote,
-    )
+    try:
+        _ = await asyncio.wait_for(
+            wallet.melt(
+                proofs=proofs,
+                invoice=bolt11_invoice,
+                fee_reserve_sat=melt_quote_resp.fee_reserve,
+                quote_id=melt_quote_resp.quote,
+            ),
+            timeout=MELT_TIMEOUT_SECONDS,
+        )
+    except asyncio.TimeoutError as e:
+        raise LNURLError(
+            f"Melt timed out after {MELT_TIMEOUT_SECONDS}s (mint unresponsive)"
+        ) from e
     return final_amount

routstr/payment/models.py

@@ -85,6 +85,64 @@ class Model(BaseModel):
         return hash(self.id)
 
 
+def litellm_cost_entry(model_id: str) -> dict | None:
+    """Look up ``model_id`` in litellm's bundled cost map.
+
+    litellm ships per-model USD rates keyed by the exact OpenRouter id
+    (``deepseek/deepseek-chat``) or the bare model name (``gpt-4o``,
+    ``claude-sonnet-4-5``), so both spellings are tried. Keys are lowercase, so
+    a mixed-case upstream id (``deepseek-ai/DeepSeek-V4-Flash``) is retried via
+    a case-insensitive scan. Returns the matched cost dict, or ``None``.
+    """
+    import litellm
+
+    candidates = (model_id, model_id.split("/", 1)[-1])
+    for key in candidates:
+        info = litellm.model_cost.get(key)
+        if isinstance(info, dict):
+            return info
+
+    lowered = {c.lower() for c in candidates}
+    for key, info in litellm.model_cost.items():
+        if isinstance(key, str) and key.lower() in lowered and isinstance(info, dict):
+            return info
+    return None
+
+
+def backfill_cache_pricing(model_id: str, pricing: Pricing) -> Pricing:
+    """Fill missing cache rates from litellm's bundled cost map.
+
+    The OpenRouter model feed omits ``input_cache_read``/``input_cache_write``
+    for many models (most DeepSeek entries, openai/gpt-4o, ...). Without a
+    cache rate, billing falls back to the full input rate, which overcharges
+    cache reads (DeepSeek hits are 10x cheaper) and undercharges Anthropic
+    cache writes (1.25x). The lookup (see ``litellm_cost_entry``) tries both
+    id spellings and a case-insensitive fallback.
+
+    Rates already present (e.g. provided by OpenRouter) are authoritative and
+    never overwritten. Unknown models are returned unchanged.
+    """
+    needs_read = (pricing.input_cache_read or 0.0) <= 0.0
+    needs_write = (pricing.input_cache_write or 0.0) <= 0.0
+    if not (needs_read or needs_write):
+        return pricing
+
+    info = litellm_cost_entry(model_id)
+    if info is None:
+        return pricing
+
+    updated = Pricing.parse_obj(pricing.dict())
+    if needs_read:
+        read_rate = info.get("cache_read_input_token_cost")
+        if isinstance(read_rate, (int, float)) and read_rate > 0:
+            updated.input_cache_read = float(read_rate)
+    if needs_write:
+        write_rate = info.get("cache_creation_input_token_cost")
+        if isinstance(write_rate, (int, float)) and write_rate > 0:
+            updated.input_cache_write = float(write_rate)
+    return updated
+
+
 def _has_valid_pricing(model: dict) -> bool:
     """Check if model has valid pricing (not free, no negative values)."""
     pricing = model.get("pricing", {})
@@ -173,13 +231,29 @@ def _row_to_model(
     )
     top_provider_dict = json.loads(row.top_provider) if row.top_provider else None
 
-    if apply_provider_fee and isinstance(pricing, dict):
-        pricing = {k: float(v) * provider_fee for k, v in pricing.items()}
-
     if isinstance(pricing, dict) and float(pricing.get("request", 0.0)) <= 0.0:
         pricing["request"] = max(pricing.get("request", 0.0), 0.0)
 
     parsed_pricing = Pricing.parse_obj(pricing)
+
+    # Fill missing cache-read/write rates from litellm's cost map BEFORE applying
+    # the provider fee, so they carry the same markup as every other component.
+    # DB-stored override pricing (e.g. generic providers) omits cache rates;
+    # without this, ``_row_to_model`` bills cache reads at the full input rate —
+    # the ``_apply_provider_fee_to_model`` path backfills, but the override path
+    # used for admin-configured providers did not.
+    #
+    # Key on ``forwarded_model_id`` (the actual upstream model name litellm
+    # prices) when set: an alias row (id="local-alias",
+    # forwarded_model_id="deepseek-v4-flash") would otherwise look up the alias
+    # and miss the cache rate.
+    pricing_model_id = getattr(row, "forwarded_model_id", None) or row.id
+    parsed_pricing = backfill_cache_pricing(pricing_model_id, parsed_pricing)
+
+    if apply_provider_fee:
+        parsed_pricing = Pricing.parse_obj(
+            {k: float(v) * provider_fee for k, v in parsed_pricing.dict().items()}
+        )
     model = Model(
         id=row.id,
         name=row.name,

routstr/payment/usage.py

@@ -0,0 +1,131 @@
+"""Vendor-agnostic normalization of upstream usage objects.
+
+Upstream providers report token usage in vendor dialects that differ in field
+names and in whether cached tokens are included in the input count:
+
+* OpenAI / Azure / xAI / Groq / Moonshot / Qwen / Gemini-compat: cache reads in
+  ``prompt_tokens_details.cached_tokens``, included in ``prompt_tokens``.
+* OpenRouter: same as OpenAI plus cache *writes* in
+  ``prompt_tokens_details.cache_write_tokens``, also included in
+  ``prompt_tokens``.
+* litellm-normalized: same nesting, but names the write field
+  ``prompt_tokens_details.cache_creation_tokens`` (and additionally mirrors the
+  Anthropic top-level fields), with ``prompt_tokens`` as the grand total.
+* Anthropic native: ``cache_read_input_tokens`` / ``cache_creation_input_tokens``
+  top-level, additive to (not included in) ``input_tokens``.
+* DeepSeek: ``prompt_cache_hit_tokens`` / ``prompt_cache_miss_tokens``, with
+  ``prompt_tokens = hit + miss``.
+
+What decides whether cached tokens must be subtracted out of the input count is
+**which prompt field the vendor uses**, not which cache field appears:
+
+* ``prompt_tokens`` present -> cached + cache-write tokens are *included* in it
+  (OpenAI family, DeepSeek, OpenRouter, litellm); subtract both so
+  ``input_tokens`` holds only the regular-rate portion.
+* only ``input_tokens`` (Anthropic native) -> cached tokens are *additive*;
+  leave ``input_tokens`` untouched.
+
+``normalize_usage`` maps all of them onto one canonical ``NormalizedUsage``
+shape so billing code needs no vendor knowledge. The known dialects' field
+names do not collide, so a single union parser is safe; a vendor whose fields
+would genuinely conflict needs a dedicated branch here.
+"""
+
+from pydantic.v1 import BaseModel
+
+
+class NormalizedUsage(BaseModel):
+    """Canonical token usage: input_tokens never includes cached tokens."""
+
+    input_tokens: int = 0
+    output_tokens: int = 0
+    cache_read_tokens: int = 0
+    cache_write_tokens: int = 0
+
+
+def parse_token_count(value: object) -> int:
+    """Parse a token count from various formats (int, float, str, bool)."""
+    if isinstance(value, bool):
+        return 0
+    if isinstance(value, int):
+        return max(0, value)
+    if isinstance(value, float):
+        return max(0, int(value))
+    if isinstance(value, str):
+        try:
+            return max(0, int(float(value)))
+        except ValueError:
+            return 0
+    return 0
+
+
+def _first_token_count(usage_data: dict, *fields: str) -> int:
+    """Return the first positive token count among the given fields."""
+    for field in fields:
+        value = parse_token_count(usage_data.get(field, 0))
+        if value > 0:
+            return value
+    return 0
+
+
+def _extract_cache_tokens(usage_data: dict) -> tuple[int, int]:
+    """Pull (cache_read, cache_write) across all known dialects.
+
+    Precedence (highest first), independent for reads and writes:
+
+    * Anthropic top-level: ``cache_read_input_tokens`` /
+      ``cache_creation_input_tokens``.
+    * Nested ``prompt_tokens_details``: ``cached_tokens`` for reads;
+      ``cache_creation_tokens`` (litellm) or ``cache_write_tokens``
+      (OpenRouter) for writes.
+    * DeepSeek: ``prompt_cache_hit_tokens`` for reads (no write concept).
+    """
+    cache_read = parse_token_count(usage_data.get("cache_read_input_tokens", 0))
+    cache_write = parse_token_count(usage_data.get("cache_creation_input_tokens", 0))
+
+    prompt_details = usage_data.get("prompt_tokens_details")
+    if isinstance(prompt_details, dict):
+        if not cache_read:
+            cache_read = parse_token_count(prompt_details.get("cached_tokens", 0))
+        if not cache_write:
+            cache_write = _first_token_count(
+                prompt_details, "cache_creation_tokens", "cache_write_tokens"
+            )
+
+    if not cache_read:
+        # DeepSeek: prompt_tokens = prompt_cache_hit_tokens + prompt_cache_miss_tokens
+        cache_read = parse_token_count(usage_data.get("prompt_cache_hit_tokens", 0))
+
+    return cache_read, cache_write
+
+
+def normalize_usage(usage_data: object) -> NormalizedUsage | None:
+    """Map a vendor usage dict onto the canonical shape, or None if absent.
+
+    Cached reads and writes are subtracted from the input count exactly once,
+    only for dialects that report a ``prompt_tokens`` grand total that already
+    includes them (OpenAI family, DeepSeek, OpenRouter, litellm). Anthropic
+    native reports them additively under ``input_tokens`` and is left untouched.
+    """
+    if not isinstance(usage_data, dict):
+        return None
+
+    output_tokens = _first_token_count(
+        usage_data, "completion_tokens", "output_tokens"
+    )
+    cache_read, cache_write = _extract_cache_tokens(usage_data)
+
+    # ``prompt_tokens`` is the inclusive grand total; ``input_tokens`` (Anthropic
+    # native) excludes cached tokens. The field chosen decides whether to subtract.
+    if "prompt_tokens" in usage_data:
+        input_tokens = parse_token_count(usage_data.get("prompt_tokens", 0))
+        input_tokens = max(0, input_tokens - cache_read - cache_write)
+    else:
+        input_tokens = parse_token_count(usage_data.get("input_tokens", 0))
+
+    return NormalizedUsage(
+        input_tokens=input_tokens,
+        output_tokens=output_tokens,
+        cache_read_tokens=cache_read,
+        cache_write_tokens=cache_write,
+    )

routstr/proxy.py

@@ -24,10 +24,12 @@ from .payment.helpers import (
     calculate_discounted_max_cost,
     check_token_balance,
     create_error_response,
+    create_upstream_error_response,
     get_max_cost_for_model,
 )
 from .payment.models import Model
 from .upstream import BaseUpstreamProvider
+from .upstream.ehbp import forward_ehbp_request, forward_ehbp_x_cashu_request
 from .upstream.helpers import init_upstreams
 from .upstream.request_correction import correct_request, extract_error_message
 
@@ -103,6 +105,34 @@ def get_unique_models() -> list[Model]:
     return list(_unique_models.values())
 
 
+def _is_tinfoil_attestation_path(path: str) -> bool:
+    """Return True for exact Tinfoil attestation routes, with optional slash."""
+    return path in {
+        "attestation",
+        "attestation/",
+        "tee/attestation",
+        "tee/attestation/",
+    }
+
+
+def _select_unauthenticated_get_upstreams(
+    path: str, upstreams: list[BaseUpstreamProvider]
+) -> list[BaseUpstreamProvider]:
+    """Select upstream candidates for unauthenticated GET bypass paths.
+
+    Tinfoil attestation endpoints are provider-specific. Trying every enabled
+    upstream can return an unrelated provider's 404 before Tinfoil is reached,
+    so route those paths only to Tinfoil providers.
+    """
+    if _is_tinfoil_attestation_path(path):
+        return [
+            upstream
+            for upstream in upstreams
+            if getattr(upstream, "provider_type", None) == "tinfoil"
+        ]
+    return upstreams
+
+
 async def refresh_model_maps() -> None:
     """Refresh global model and provider maps using the cost-based algorithm."""
     from sqlalchemy.orm import selectinload
@@ -117,22 +147,23 @@ async def refresh_model_maps() -> None:
         result = await session.exec(query)
         provider_rows = result.all()
 
-    overrides_by_id: dict[str, tuple[ModelRow, float]] = {}
-    disabled_model_ids: set[str] = set()
+    overrides_by_key: dict[tuple[str, int], tuple[ModelRow, float]] = {}
+    disabled_model_keys: set[tuple[str, int]] = set()
 
     for provider in provider_rows:
         if not provider.enabled:
             continue
         for model in provider.models:
+            model_key = (model.id.lower(), model.upstream_provider_id)
             if model.enabled:
-                overrides_by_id[model.id] = (model, provider.provider_fee)
+                overrides_by_key[model_key] = (model, provider.provider_fee)
             else:
-                disabled_model_ids.add(model.id)
+                disabled_model_keys.add(model_key)
 
     _model_instances, _provider_map, _unique_models = create_model_mappings(
         upstreams=_upstreams,
-        overrides_by_id=overrides_by_id,
-        disabled_model_ids=disabled_model_ids,
+        overrides_by_key=overrides_by_key,
+        disabled_model_keys=disabled_model_keys,
     )
 
 
@@ -165,6 +196,7 @@ _API_PATH_PREFIXES = (
     "moderations",
     "providers",
     "tee/",
+    "attestation",
 )
 
 
@@ -183,20 +215,54 @@ async def proxy(
 
     is_responses_api = path.startswith("v1/responses") or path.startswith("responses")
     request_body = await request.body()
-    request_body_dict = parse_request_body_json(request_body, path)
 
-    # /tee/* GET requests (e.g. attestation) don't map to models — just
-    # forward to all enabled upstreams without model/cost/auth lookups.
-    if request.method == "GET" and path.startswith("tee/"):
-        all_upstreams = _upstreams
+    # EHBP (Encrypted HTTP Body Protocol) requests carry an Ehbp-Encapsulated-Key
+    # header and a binary HPKE-sealed body. The proxy cannot parse the body to
+    # extract the model id, so the SDK sends it in X-Routstr-Model. Forward the
+    # raw encrypted body to the upstream's /private/ endpoint and stream the
+    # encrypted response back untouched — the SDK's SecureClient decrypts it.
+    is_ehbp = "ehbp-encapsulated-key" in headers
+    if is_ehbp:
+        request_body_dict = {}
+        model_id = headers.get("x-routstr-model", "")
+        if not model_id:
+            return create_error_response(
+                "invalid_request",
+                "EHBP request missing X-Routstr-Model header",
+                400,
+                request=request,
+            )
+    else:
+        request_body_dict = parse_request_body_json(request_body, path)
+        if is_responses_api:
+            model_id = extract_model_from_responses_request(request_body_dict)
+        else:
+            model_id = request_body_dict.get("model", "unknown")
+
+    # Exact Tinfoil attestation GET routes don't map to models — forward
+    # without model/cost/auth lookups. Do not prefix-match here: paths such as
+    # /attestationjunk must continue through normal authentication.
+    if request.method == "GET" and _is_tinfoil_attestation_path(path):
+        selected_upstreams = _select_unauthenticated_get_upstreams(path, _upstreams)
+        if not selected_upstreams:
+            return create_error_response(
+                "upstream_error",
+                "No upstream available for unauthenticated GET path",
+                502,
+                request=request,
+            )
+
         last_error_response = None
-        for i, upstream in enumerate(all_upstreams):
+        for i, upstream in enumerate(selected_upstreams):
             try:
                 headers = upstream.prepare_headers(dict(request.headers))
                 response = await upstream.forward_get_request(request, path, headers)
-                if response.status_code in [502, 429] and i < len(all_upstreams) - 1:
+                if (
+                    response.status_code in [502, 429]
+                    and i < len(selected_upstreams) - 1
+                ):
                     logger.warning(
-                        "Upstream %s returned %s for tee GET %s, trying next",
+                        "Upstream %s returned %s for unauthenticated GET %s, trying next",
                         upstream.provider_type,
                         response.status_code,
                         path,
@@ -205,25 +271,18 @@ async def proxy(
                 return response
             except UpstreamError as e:
                 logger.warning(
-                    "Upstream %s failed for tee GET %s: %s",
+                    "Upstream %s failed for unauthenticated GET %s: %s",
                     upstream.provider_type,
                     path,
                     e,
                 )
-                if i == len(all_upstreams) - 1:
-                    last_error_response = create_error_response(
-                        "upstream_error", str(e), 502, request=request
-                    )
+                if i == len(selected_upstreams) - 1:
+                    last_error_response = create_upstream_error_response(e, request)
                 continue
         return last_error_response or create_error_response(
             "upstream_error", "All upstreams failed", 502, request=request
         )
 
-    if is_responses_api:
-        model_id = extract_model_from_responses_request(request_body_dict)
-    else:
-        model_id = request_body_dict.get("model", "unknown")
-
     model_obj = get_model_instance(model_id)
 
     if not model_obj:
@@ -240,6 +299,16 @@ async def proxy(
             request=request,
         )
 
+    if is_ehbp:
+        upstreams = [upstream for upstream in upstreams if upstream.supports_ehbp]
+        if not upstreams:
+            return create_error_response(
+                "unsupported_request",
+                f"No EHBP-capable provider found for model '{model_id}'",
+                400,
+                request=request,
+            )
+
     # todo figure out cost calculation since fallback provider is usually not the same price
     # Use first provider for initial checks/cost calculation
     # primary_upstream = upstreams[0]
@@ -259,7 +328,23 @@ async def proxy(
         last_error = None
         for i, upstream in enumerate(upstreams):
             try:
-                if is_responses_api:
+                if is_ehbp:
+                    if not upstream.supports_ehbp:
+                        logger.warning(
+                            "Upstream %s does not support EHBP for model=%s",
+                            upstream.provider_type,
+                            model_id,
+                        )
+                        continue
+                    return await forward_ehbp_x_cashu_request(
+                        request=request,
+                        x_cashu_token=x_cashu,
+                        path=path,
+                        max_cost_for_model=max_cost_for_model,
+                        model_obj=model_obj,
+                        upstream=upstream,
+                    )
+                elif is_responses_api:
                     return await upstream.handle_x_cashu_responses(
                         request, x_cashu, path, max_cost_for_model, model_obj
                     )
@@ -283,11 +368,10 @@ async def proxy(
                     last_error = e
                 continue
 
+        if last_error is not None:
+            return create_upstream_error_response(last_error, request)
         return create_error_response(
-            "upstream_error",
-            str(last_error) if last_error else "All upstreams failed",
-            502,
-            request=request,
+            "upstream_error", "All upstreams failed", 502, request=request
         )
 
     elif auth := headers.get("authorization", None):
@@ -343,15 +427,13 @@ async def proxy(
             except UpstreamError as e:
                 logger.warning(f"Upstream {upstream.provider_type} failed (GET): {e}")
                 if i == len(upstreams) - 1:
-                    last_error_response = create_error_response(
-                        "upstream_error", str(e), 502, request=request
-                    )
+                    last_error_response = create_upstream_error_response(e, request)
                 continue
         return last_error_response or create_error_response(
             "upstream_error", "All upstreams failed", 502, request=request
         )
 
-    if request_body_dict:
+    if is_ehbp or request_body_dict:
         await pay_for_request(key, max_cost_for_model, session)
 
     # Tracks request params already removed in response to upstream rejections,
@@ -365,7 +447,29 @@ async def proxy(
         try:
             while True:
                 try:
-                    if is_responses_api:
+                    if is_ehbp:
+                        if not upstream.supports_ehbp:
+                            logger.warning(
+                                "Upstream %s does not support EHBP for model=%s",
+                                upstream.provider_type,
+                                model_id,
+                            )
+                            raise UpstreamError(
+                                f"Provider {upstream.provider_type} does not support EHBP",
+                                status_code=400,
+                            )
+                        response = await forward_ehbp_request(
+                            request=request,
+                            path=path,
+                            headers=headers,
+                            request_body=request_body,
+                            upstream=upstream,
+                            key=key,
+                            max_cost_for_model=max_cost_for_model,
+                            session=session,
+                            model_obj=model_obj,
+                        )
+                    elif is_responses_api:
                         response = await upstream.forward_responses_request(
                             request,
                             path,
@@ -410,7 +514,7 @@ async def proxy(
                 # When the upstream 400s naming such a param, strip it from the
                 # body and retry the SAME upstream. ``already_stripped`` bounds
                 # this to one retry per distinct param so it always terminates.
-                if response.status_code == 400:
+                if response.status_code == 400 and not is_ehbp:
                     correction = correct_request(
                         request_body,
                         extract_error_message(response),
@@ -525,9 +629,7 @@ async def proxy(
             # If this was the last provider
             if i == len(upstreams) - 1:
                 await revert_pay_for_request(key, session, max_cost_for_model)
-                return create_error_response(
-                    "upstream_error", str(e), 502, request=request
-                )
+                return create_upstream_error_response(e, request)
 
             # Otherwise loop continues to next provider
             continue

routstr/upstream/__init__.py

@@ -11,6 +11,7 @@ from .openrouter import OpenRouterUpstreamProvider
 from .perplexity import PerplexityUpstreamProvider
 from .ppqai import PPQAIUpstreamProvider
 from .routstr import RoutstrUpstreamProvider
+from .tinfoil import TinfoilUpstreamProvider
 from .xai import XAIUpstreamProvider
 
 upstream_provider_classes: list[type[BaseUpstreamProvider]] = [
@@ -26,6 +27,7 @@ upstream_provider_classes: list[type[BaseUpstreamProvider]] = [
     PerplexityUpstreamProvider,
     PPQAIUpstreamProvider,
     RoutstrUpstreamProvider,
+    TinfoilUpstreamProvider,
     XAIUpstreamProvider,
 ]
 """List of all upstream classes"""

routstr/upstream/anthropic.py

@@ -24,7 +24,7 @@ class AnthropicUpstreamProvider(BaseUpstreamProvider):
         )
 
     @classmethod
-    def from_db_row(
+    def _build_from_row(
         cls, provider_row: "UpstreamProviderRow"
     ) -> "AnthropicUpstreamProvider":
         return cls(

routstr/upstream/auto_topup.py

@@ -4,7 +4,12 @@ import json
 from sqlmodel import select
 
 from ..core import get_logger
-from ..core.db import UpstreamProviderRow, create_session
+from ..core.db import (
+    CashuTransaction,
+    UpstreamProviderRow,
+    create_session,
+    store_cashu_transaction,
+)
 from ..wallet import send_token
 from .routstr import RoutstrUpstreamProvider
 
@@ -97,6 +102,8 @@ async def _check_and_topup(row: UpstreamProviderRow) -> None:
 
     # Instantiate provider and check balance
     provider = RoutstrUpstreamProvider.from_db_row(row)
+    if provider is None:
+        return
     balance = await provider.get_balance()
 
     if balance is None:
@@ -121,7 +128,6 @@ async def _check_and_topup(row: UpstreamProviderRow) -> None:
         },
     )
 
-    print(amount, mint_url)
     try:
         token = await send_token(amount, "sat", mint_url)
     except Exception as e:
@@ -136,6 +142,22 @@ async def _check_and_topup(row: UpstreamProviderRow) -> None:
         )
         return
 
+    stored = await store_cashu_transaction(
+        token=token,
+        amount=amount,
+        unit="sat",
+        mint_url=mint_url,
+        typ="out",
+        collected=False,
+        source="auto_topup",
+    )
+    if not stored:
+        logger.critical(
+            "Aborting auto top-up because its cashu token could not be persisted",
+            extra={"provider_id": row.id, "mint_url": mint_url},
+        )
+        return
+
     result = await provider.topup(token)
 
     if "error" in result:
@@ -147,6 +169,26 @@ async def _check_and_topup(row: UpstreamProviderRow) -> None:
             },
         )
     else:
+        async with create_session() as session:
+            transaction = (
+                await session.exec(
+                    select(CashuTransaction).where(
+                        CashuTransaction.token == token,
+                        CashuTransaction.type == "out",
+                        CashuTransaction.source == "auto_topup",
+                    )
+                )
+            ).first()
+            if transaction is None:
+                logger.critical(
+                    "Completed auto top-up transaction is missing from the database",
+                    extra={"provider_id": row.id, "mint_url": mint_url},
+                )
+            else:
+                transaction.collected = True
+                session.add(transaction)
+                await session.commit()
+
         logger.info(
             "Auto top-up completed successfully",
             extra={

routstr/upstream/azure.py

@@ -38,7 +38,7 @@ class AzureUpstreamProvider(BaseUpstreamProvider):
         self.api_version = api_version
 
     @classmethod
-    def from_db_row(
+    def _build_from_row(
         cls, provider_row: "UpstreamProviderRow"
     ) -> "AzureUpstreamProvider | None":
         if not provider_row.api_version:

routstr/upstream/base.py

@@ -2,16 +2,17 @@ from __future__ import annotations
 
 import asyncio
 import json
+import math
 import traceback
+import typing
 import uuid
 from collections.abc import AsyncGenerator, AsyncIterator, Iterator
-from typing import Any, Mapping, cast
+from typing import Any, Mapping, Self, cast
 
 import httpx
 from fastapi import BackgroundTasks, HTTPException, Request
 from fastapi.responses import Response, StreamingResponse
 from pydantic.v1 import BaseModel
-from sqlmodel import select
 
 from ..auth import adjust_payment_for_tokens
 from ..core import get_logger
@@ -23,6 +24,7 @@ from ..core.db import (
     store_cashu_transaction,
 )
 from ..core.exceptions import UpstreamError
+from ..core.redaction import redact_org_ids
 from ..payment.cost_calculation import (
     CostData,
     CostDataError,
@@ -35,13 +37,27 @@ from ..payment.models import (
     Pricing,
     _calculate_usd_max_costs,
     _update_model_sats_pricing,
+    backfill_cache_pricing,
     list_models,
 )
 from ..payment.price import sats_usd_price
-from ..wallet import recieve_token, send_token
+from ..wallet import (
+    SPENT_TOKEN_CODES,
+    classify_redemption_error,
+    recieve_token,
+    send_token,
+)
 from . import messages_dispatch
+from .cache_breakpoints import (
+    inject_anthropic_cache_breakpoints,
+    is_explicit_cache_model,
+)
 from .count_tokens import count_tokens_locally
 from .litellm_routing import detect_litellm_prefix
+from .rate_limit import UPSTREAM_RATE_LIMIT, classify_rate_limit
+
+if typing.TYPE_CHECKING:
+    from .ehbp import ConfidentialInferenceProfile, EHBPForwardingTarget
 
 logger = get_logger(__name__)
 
@@ -84,6 +100,11 @@ class BaseUpstreamProvider:
     base_url: str
     api_key: str
     provider_fee: float = 1.05
+    # Primary key of the ``upstream_providers`` row this instance was built
+    # from. Set by ``from_db_row`` so a live provider can re-find its own row by
+    # stable identity instead of its rotatable ``api_key``. ``None`` for
+    # instances not sourced from a row.
+    db_id: int | None = None
     _models_cache: list[Model] = []
     _models_by_id: dict[str, Model] = {}
 
@@ -98,6 +119,7 @@ class BaseUpstreamProvider:
         self.base_url = base_url
         self.api_key = api_key
         self.provider_fee = provider_fee
+        self.db_id = None
         self._models_cache = []
         self._models_by_id = {}
 
@@ -115,10 +137,13 @@ class BaseUpstreamProvider:
         return detect_litellm_prefix(self.base_url)
 
     @classmethod
-    def from_db_row(
-        cls, provider_row: "UpstreamProviderRow"
-    ) -> "BaseUpstreamProvider | None":
-        """Factory method to instantiate provider from database row.
+    def from_db_row(cls, provider_row: "UpstreamProviderRow") -> "Self | None":
+        """Instantiate a provider from a database row, carrying its identity.
+
+        Construction itself is delegated to the ``_build_from_row`` hook (which
+        subclasses override to match their constructor); this wrapper stamps the
+        row's primary key onto the instance as ``db_id`` so the provider can
+        later re-find its own row by identity rather than by its ``api_key``.
 
         Args:
             provider_row: Database row containing provider configuration
@@ -126,6 +151,19 @@ class BaseUpstreamProvider:
         Returns:
             Instantiated provider or None if instantiation fails
         """
+        provider = cls._build_from_row(provider_row)
+        if provider is not None:
+            provider.db_id = provider_row.id
+        return provider
+
+    @classmethod
+    def _build_from_row(cls, provider_row: "UpstreamProviderRow") -> "Self | None":
+        """Construct the provider instance from a row (no identity stamping).
+
+        Overridden by subclasses whose constructors differ from the base
+        ``(base_url, api_key, provider_fee)`` shape. Callers should use
+        ``from_db_row`` instead, which also attaches ``db_id``.
+        """
         return cls(
             base_url=provider_row.base_url,
             api_key=provider_row.api_key,
@@ -432,6 +470,19 @@ class BaseUpstreamProvider:
 
         return body
 
+    def _upstream_accepts_cache_control(self) -> bool:
+        """True when this upstream accepts explicit ``cache_control`` markers.
+
+        Only OpenRouter (documents Anthropic + Alibaba explicit caching) and the
+        native Anthropic API accept the markers. Stamping them toward an
+        automatic-cache or non-supporting upstream risks a 400, so injection is
+        confined to these. Base URL is also checked so an OpenRouter endpoint
+        configured through the generic provider is still recognised.
+        """
+        if self.provider_type in ("openrouter", "anthropic"):
+            return True
+        return "openrouter.ai" in (self.base_url or "")
+
     def prepare_request_body(
         self, body: bytes | None, model_obj: Model
     ) -> bytes | None:
@@ -501,6 +552,27 @@ class BaseUpstreamProvider:
                 data["stream_options"] = merged
                 changed = True
 
+        # Explicit-cache models (Anthropic Claude, Alibaba Qwen / deepseek-v3.2)
+        # cache nothing without ``cache_control`` markers in the body. Clients
+        # that don't recognise a routstr URL as one of these never send them, so
+        # caching silently never engages over routstr even though it works
+        # against OpenRouter directly. Stamp the standard breakpoints so caching
+        # works by default, deferring to any client-set markers. Gated to
+        # upstreams that accept the markers (OpenRouter / Anthropic) so they
+        # never leak to an automatic-cache provider that would reject them.
+        if (
+            "messages" in data
+            and isinstance(data.get("messages"), list)
+            and self._upstream_accepts_cache_control()
+            and is_explicit_cache_model(
+                model_obj.id,
+                model_obj.forwarded_model_id,
+                model_obj.canonical_slug,
+            )
+        ):
+            if inject_anthropic_cache_breakpoints(data):
+                changed = True
+
         if changed:
             return json.dumps(data).encode()
         return body
@@ -543,7 +615,7 @@ class BaseUpstreamProvider:
             preview = body_bytes.decode("utf-8", errors="ignore").strip()
             if preview:
                 message = preview[:500]
-        return message, upstream_code
+        return redact_org_ids(message), upstream_code
 
     async def on_upstream_error_redirect(
         self, status_code: int, error_message: str
@@ -586,10 +658,23 @@ class BaseUpstreamProvider:
             body_bytes = b""
             body_read_error = f"{type(exc).__name__}: {exc}"
 
+        # ``message`` is already redacted by ``_extract_upstream_error_message``;
+        # the raw body preview is redacted here before it reaches logs or the
+        # forwarded envelope so provider account identifiers never leak.
         message, upstream_code = self._extract_upstream_error_message(body_bytes)
-        body_preview = body_bytes.decode("utf-8", errors="ignore").strip()[:500]
+        body_preview = redact_org_ids(
+            body_bytes.decode("utf-8", errors="ignore").strip()[:500]
+        )
         is_json_body = _is_json_content_type(content_type)
 
+        # Classify upstream rate-limit failures into a stable, structured error.
+        rate_limit = classify_rate_limit(status_code, message, headers)
+        error_code: str | int = upstream_code or status_code
+        error_details: dict[str, object] | None = None
+        if rate_limit is not None:
+            error_code = UPSTREAM_RATE_LIMIT
+            error_details = rate_limit.as_details()
+
         logger.warning(
             "Upstream %s returned %s for model=%s path=%s: %s",
             self.provider_type,
@@ -603,6 +688,7 @@ class BaseUpstreamProvider:
                 "model": model_id or "unknown",
                 "upstream_status": status_code,
                 "upstream_code": upstream_code,
+                "error_code": error_code,
                 "upstream_content_type": content_type,
                 "upstream_request_id": upstream_request_id,
                 "message_preview": message[:200],
@@ -637,13 +723,41 @@ class BaseUpstreamProvider:
         ):
             headers.pop(header_name, None)
 
+        # Propagate a usable retry hint to the caller when the upstream supplied
+        # one but did not echo a ``Retry-After`` header. RFC 7231 delta-seconds
+        # is an integer, so round sub-second hints up to a usable ``1``.
+        if (
+            rate_limit is not None
+            and rate_limit.retry_after_seconds is not None
+            and "retry-after" not in {k.lower() for k in headers}
+        ):
+            headers["Retry-After"] = str(max(1, math.ceil(rate_limit.retry_after_seconds)))
+
         if is_json_body:
             if not content_type:
                 headers.pop("content-type", None)
                 headers.pop("Content-Type", None)
             media_type = content_type or None
+            # Re-serialise the body with organization IDs stripped. The narrow
+            # ``org-*`` regex preserves the surrounding JSON structure.
+            redacted_text = redact_org_ids(body_bytes.decode("utf-8", errors="ignore"))
+            redacted_body = redacted_text.encode()
+            # Surface the stable rate-limit classification on the forwarded
+            # body so callers can switch on ``error.code`` without parsing the
+            # provider-specific message. Fall back to the redacted bytes if the
+            # body is not a JSON object with an ``error`` mapping.
+            if rate_limit is not None:
+                try:
+                    parsed = json.loads(redacted_text)
+                    err = parsed.get("error") if isinstance(parsed, dict) else None
+                    if isinstance(err, dict):
+                        err["code"] = UPSTREAM_RATE_LIMIT
+                        err["details"] = error_details
+                        redacted_body = json.dumps(parsed).encode()
+                except (ValueError, AttributeError):
+                    pass
             return Response(
-                content=body_bytes,
+                content=redacted_body,
                 status_code=status_code,
                 headers=headers,
                 media_type=media_type,
@@ -654,15 +768,18 @@ class BaseUpstreamProvider:
         for header_name in ("content-type", "Content-Type"):
             headers.pop(header_name, None)
 
+        error_obj: dict[str, object] = {
+            "message": message or "Upstream returned a non-JSON error response",
+            "type": "upstream_error",
+            "code": error_code,
+            "upstream_status": status_code,
+            "upstream_content_type": content_type or None,
+            "upstream_body_preview": body_preview or None,
+        }
+        if error_details is not None:
+            error_obj["details"] = error_details
         envelope = {
-            "error": {
-                "message": message or "Upstream returned a non-JSON error response",
-                "type": "upstream_error",
-                "code": upstream_code or status_code,
-                "upstream_status": status_code,
-                "upstream_content_type": content_type or None,
-                "upstream_body_preview": body_preview or None,
-            },
+            "error": error_obj,
             "request_id": getattr(request.state, "request_id", None),
         }
 
@@ -727,7 +844,9 @@ class BaseUpstreamProvider:
                     except Exception:
                         pass
 
-            def _process_event(raw_event: bytes) -> Iterator[bytes]:
+            def _process_event(
+                raw_event: bytes, final: bool = False
+            ) -> Iterator[bytes]:
                 """Process one complete SSE event block (lines up to a blank line).
 
                 Handles arbitrary upstream framing across every supported
@@ -833,6 +952,12 @@ class BaseUpstreamProvider:
                         return
                     yield prefix + b"data: " + json.dumps(obj).encode() + b"\n\n"
                 else:
+                    if final:
+                        # Final flush of a truncated tail: the upstream closed
+                        # mid-event, so ``data`` is incomplete JSON. Emitting it
+                        # as a ``data:`` frame would hand the client invalid
+                        # JSON (the "unexpected token" parse error). Drop it.
+                        return
                     # Non-JSON data payload (partial fragment already reassembled
                     # by buffering, or a provider control string). Re-prefix each
                     # line so multi-line ``data`` stays valid SSE framing - a bare
@@ -851,7 +976,13 @@ class BaseUpstreamProvider:
                 # boundary-independent for every provider.
                 buffer = b""
                 async for chunk in response.aiter_bytes():
-                    buffer += chunk.replace(b"\r\n", b"\n")
+                    # Normalize the *joined* buffer, not each chunk in
+                    # isolation: a CRLF event delimiter can straddle two
+                    # ``aiter_bytes`` chunks (``...\r`` then ``\n...``). A
+                    # per-chunk replace would leave a stray ``\r`` and the
+                    # ``\n\n`` split would miss the delimiter, merging two
+                    # events into one frame and breaking SSE clients.
+                    buffer = (buffer + chunk).replace(b"\r\n", b"\n")
                     while b"\n\n" in buffer:
                         raw_event, buffer = buffer.split(b"\n\n", 1)
                         for out in _process_event(raw_event):
@@ -859,7 +990,7 @@ class BaseUpstreamProvider:
 
                 # Flush any trailing event that lacked a final blank line.
                 if buffer.strip():
-                    for out in _process_event(buffer):
+                    for out in _process_event(buffer, final=True):
                         yield out
 
                 async with create_session() as session:
@@ -1018,7 +1149,10 @@ class BaseUpstreamProvider:
                 response_json["id"] = f"chatcmpl-{uuid.uuid4()}"
 
             cost_data = await adjust_payment_for_tokens(
-                key, response_json, session, deducted_max_cost
+                key,
+                response_json,
+                session,
+                deducted_max_cost,
             )
 
             await session.refresh(key)
@@ -1162,7 +1296,9 @@ class BaseUpstreamProvider:
                     except Exception:
                         pass
 
-            def _process_event(raw_event: bytes) -> Iterator[bytes]:
+            def _process_event(
+                raw_event: bytes, final: bool = False
+            ) -> Iterator[bytes]:
                 """Process one complete SSE event block for the Responses API.
 
                 Buffers full events (delimited by a blank line) so parsing is
@@ -1232,6 +1368,11 @@ class BaseUpstreamProvider:
 
                     yield prefix + b"data: " + json.dumps(obj).encode() + b"\n\n"
                 else:
+                    if final:
+                        # Final flush of a truncated tail: upstream closed
+                        # mid-event, so ``data`` is incomplete JSON. Dropping it
+                        # avoids handing the client an invalid ``data:`` frame.
+                        return
                     # Re-prefix each line so multi-line ``data`` stays valid SSE
                     # framing for the client.
                     body = b"".join(
@@ -1244,14 +1385,20 @@ class BaseUpstreamProvider:
                 # delimiter so parsing is independent of byte boundaries.
                 buffer = b""
                 async for chunk in response.aiter_bytes():
-                    buffer += chunk.replace(b"\r\n", b"\n")
+                    # Normalize the *joined* buffer, not each chunk in
+                    # isolation: a CRLF event delimiter can straddle two
+                    # ``aiter_bytes`` chunks (``...\r`` then ``\n...``). A
+                    # per-chunk replace would leave a stray ``\r`` and the
+                    # ``\n\n`` split would miss the delimiter, merging two
+                    # events into one frame and breaking SSE clients.
+                    buffer = (buffer + chunk).replace(b"\r\n", b"\n")
                     while b"\n\n" in buffer:
                         raw_event, buffer = buffer.split(b"\n\n", 1)
                         for out in _process_event(raw_event):
                             yield out
 
                 if buffer.strip():
-                    for out in _process_event(buffer):
+                    for out in _process_event(buffer, final=True):
                         yield out
 
                 # Always emit a cost-bearing data chunk
@@ -1436,7 +1583,10 @@ class BaseUpstreamProvider:
                 response_json["id"] = f"chatcmpl-{uuid.uuid4()}"
 
             cost_data = await adjust_payment_for_tokens(
-                key, response_json, session, deducted_max_cost
+                key,
+                response_json,
+                session,
+                deducted_max_cost,
             )
 
             await session.refresh(key)
@@ -1631,7 +1781,10 @@ class BaseUpstreamProvider:
                             "usage": None,
                         }
                         cost_data = await adjust_payment_for_tokens(
-                            fresh_key, fallback, new_session, max_cost_for_model
+                            fresh_key,
+                            fallback,
+                            new_session,
+                            max_cost_for_model,
                         )
                         usage_finalized = True
                         return f"event: cost\ndata: {json.dumps({'cost': cost_data})}\n\n".encode()
@@ -1850,7 +2003,10 @@ class BaseUpstreamProvider:
                 response_json["usage"] = {"input_tokens": input_tokens}
 
             cost_data = await adjust_payment_for_tokens(
-                key, response_json, session, deducted_max_cost
+                key,
+                response_json,
+                session,
+                deducted_max_cost,
             )
 
             self.inject_cost_metadata(response_json, cost_data, key)
@@ -1952,7 +2108,10 @@ class BaseUpstreamProvider:
             response_json["model"] = requested_model
 
         cost_data = await adjust_payment_for_tokens(
-            key, response_json, session, max_cost_for_model
+            key,
+            response_json,
+            session,
+            max_cost_for_model,
         )
         self.inject_cost_metadata(response_json, cost_data, key)
 
@@ -2440,9 +2599,16 @@ class BaseUpstreamProvider:
                         body_bytes = await response.aread()
                     except Exception:
                         body_bytes = b""
-                    body_preview = body_bytes.decode(
-                        "utf-8", errors="ignore"
-                    ).strip()[:500]
+                    # Redact provider account identifiers before the body text
+                    # reaches logs or the raised error.
+                    body_preview = redact_org_ids(
+                        body_bytes.decode("utf-8", errors="ignore").strip()[:500]
+                    )
+                    rate_limit = classify_rate_limit(
+                        response.status_code,
+                        body_preview,
+                        dict(response.headers),
+                    )
                     logger.error(
                         "Upstream %s returned %s for model=%s path=%s: %s",
                         self.provider_type,
@@ -2454,6 +2620,7 @@ class BaseUpstreamProvider:
                             "provider": self.provider_type,
                             "model": original_model_id or "unknown",
                             "status_code": response.status_code,
+                            "error_code": rate_limit.code if rate_limit else None,
                             "reason_phrase": response.reason_phrase,
                             "path": path,
                             "body_preview": body_preview,
@@ -2466,6 +2633,8 @@ class BaseUpstreamProvider:
                         f"for model {original_model_id or 'unknown'}: "
                         f"{body_preview[:200] or '<empty>'}",
                         status_code=response.status_code,
+                        code=rate_limit.code if rate_limit else None,
+                        details=rate_limit.as_details() if rate_limit else None,
                     )
 
                 try:
@@ -2680,6 +2849,26 @@ class BaseUpstreamProvider:
             # Don't revert here — proxy.py owns payment revert to avoid double-revert
             raise UpstreamError("An unexpected server error occurred", status_code=500)
 
+    supports_ehbp: bool = False
+
+    def get_confidential_inference_profile(self) -> "ConfidentialInferenceProfile | None":
+        """Return provider policy for encrypted/confidential inference forwarding."""
+        return None
+
+    def get_ehbp_forwarding_target(
+        self, path: str, model_obj: Model
+    ) -> "EHBPForwardingTarget":
+        """Return the EHBP forwarding target for this provider.
+
+        Providers must explicitly opt in by setting ``supports_ehbp = True``
+        and overriding this method. Most upstreams do not accept EHBP-encrypted
+        request bodies, so the base provider intentionally does not provide a
+        default endpoint.
+        """
+        raise NotImplementedError(
+            f"Provider {self.provider_type} does not support EHBP forwarding"
+        )
+
     async def forward_responses_request(
         self,
         request: Request,
@@ -2762,9 +2951,16 @@ class BaseUpstreamProvider:
                         body_bytes = await response.aread()
                     except Exception:
                         body_bytes = b""
-                    body_preview = body_bytes.decode(
-                        "utf-8", errors="ignore"
-                    ).strip()[:500]
+                    # Redact provider account identifiers before the body text
+                    # reaches logs or the raised error.
+                    body_preview = redact_org_ids(
+                        body_bytes.decode("utf-8", errors="ignore").strip()[:500]
+                    )
+                    rate_limit = classify_rate_limit(
+                        response.status_code,
+                        body_preview,
+                        dict(response.headers),
+                    )
                     logger.error(
                         "Upstream %s returned %s for model=%s path=%s: %s",
                         self.provider_type,
@@ -2776,6 +2972,7 @@ class BaseUpstreamProvider:
                             "provider": self.provider_type,
                             "model": original_model_id or "unknown",
                             "status_code": response.status_code,
+                            "error_code": rate_limit.code if rate_limit else None,
                             "path": path,
                             "body_preview": body_preview,
                         },
@@ -2787,6 +2984,8 @@ class BaseUpstreamProvider:
                         f"for model {original_model_id or 'unknown'}: "
                         f"{body_preview[:200] or '<empty>'}",
                         status_code=response.status_code,
+                        code=rate_limit.code if rate_limit else None,
+                        details=rate_limit.as_details() if rate_limit else None,
                     )
 
                 try:
@@ -3025,44 +3224,46 @@ class BaseUpstreamProvider:
             extra={"model": model, "has_usage": "usage" in response_data},
         )
 
-        async with create_session() as session:
-            match await calculate_cost(response_data, max_cost_for_model, session):
-                case MaxCostData() as cost:
-                    logger.debug(
-                        "Using max cost pricing",
-                        extra={"model": model, "max_cost_msats": cost.total_msats},
-                    )
-                    return cost
-                case CostData() as cost:
-                    logger.debug(
-                        "Using token-based pricing",
-                        extra={
-                            "model": model,
-                            "total_cost_msats": cost.total_msats,
-                            "input_msats": cost.input_msats,
-                            "output_msats": cost.output_msats,
-                        },
-                    )
-                    return cost
-                case CostDataError() as error:
-                    logger.error(
-                        "Cost calculation error",
-                        extra={
-                            "model": model,
-                            "error_message": error.message,
-                            "error_code": error.code,
-                        },
-                    )
-                    raise HTTPException(
-                        status_code=400,
-                        detail={
-                            "error": {
-                                "message": error.message,
-                                "type": "invalid_request_error",
-                                "code": error.code,
-                            }
-                        },
-                    )
+        match await calculate_cost(
+            response_data,
+            max_cost_for_model,
+        ):
+            case MaxCostData() as cost:
+                logger.debug(
+                    "Using max cost pricing",
+                    extra={"model": model, "max_cost_msats": cost.total_msats},
+                )
+                return cost
+            case CostData() as cost:
+                logger.debug(
+                    "Using token-based pricing",
+                    extra={
+                        "model": model,
+                        "total_cost_msats": cost.total_msats,
+                        "input_msats": cost.input_msats,
+                        "output_msats": cost.output_msats,
+                    },
+                )
+                return cost
+            case CostDataError() as error:
+                logger.error(
+                    "Cost calculation error",
+                    extra={
+                        "model": model,
+                        "error_message": error.message,
+                        "error_code": error.code,
+                    },
+                )
+                raise HTTPException(
+                    status_code=400,
+                    detail={
+                        "error": {
+                            "message": error.message,
+                            "type": "invalid_request_error",
+                            "code": error.code,
+                        }
+                    },
+                )
         return None
 
     async def send_refund(
@@ -3809,9 +4010,19 @@ class BaseUpstreamProvider:
             },
         )
 
+        redeemed = False
         try:
             headers = dict(request.headers)
             amount, unit, mint = await recieve_token(x_cashu_token)
+            # Reject a zero/negative redemption (empty/dust token, or a value
+            # fully consumed by fees) before marking the token redeemed, so it
+            # classifies as cashu_token_zero_value like the bearer/top-up paths
+            # rather than being forwarded as a free request.
+            if amount <= 0:
+                raise ValueError(
+                    f"Redeemed token amount must be positive, got {amount} {unit}"
+                )
+            redeemed = True
             headers = self.prepare_headers(dict(request.headers))
 
             request_id = getattr(request.state, "request_id", None)
@@ -3855,40 +4066,37 @@ class BaseUpstreamProvider:
                 },
             )
 
-            # Use same error handling as regular X-Cashu
-            if "already spent" in error_message.lower():
+            # Post-redemption the token is spent; a forwarding failure must not
+            # be reported as a retryable redemption error (see handle_x_cashu).
+            if redeemed:
                 return create_error_response(
-                    "token_already_spent",
-                    "The provided CASHU token has already been spent",
-                    400,
+                    "upstream_error",
+                    "Payment succeeded but the upstream request failed",
+                    502,
                     request=request,
-                    token=x_cashu_token,
+                    code="upstream_request_failed",
                 )
 
-            if "invalid token" in error_message.lower():
+            classified = classify_redemption_error(e)
+            if classified is None:
                 return create_error_response(
-                    "invalid_token",
-                    "The provided CASHU token is invalid",
-                    400,
+                    "api_error",
+                    "Internal error during token redemption",
+                    500,
                     request=request,
-                    token=x_cashu_token,
+                    code="internal_error",
                 )
-
-            if "mint error" in error_message.lower():
-                return create_error_response(
-                    "mint_error",
-                    f"CASHU mint error: {error_message}",
-                    422,
-                    request=request,
-                    token=x_cashu_token,
-                )
-
+            error_type, status_code, message, error_code = classified
+            # Echo the token back only when it is still spendable, so clients
+            # can recover it; a spent/consumed token is never re-offered.
+            echo_token = None if error_code in SPENT_TOKEN_CODES else x_cashu_token
             return create_error_response(
-                "cashu_error",
-                f"CASHU token processing failed: {error_message}",
-                400,
+                error_type,
+                message,
+                status_code,
                 request=request,
-                token=x_cashu_token,
+                token=echo_token,
+                code=error_code,
             )
 
     async def forward_x_cashu_responses_request(
@@ -4478,9 +4686,19 @@ class BaseUpstreamProvider:
             },
         )
 
+        redeemed = False
         try:
             headers = dict(request.headers)
             amount, unit, mint = await recieve_token(x_cashu_token)
+            # Reject a zero/negative redemption (empty/dust token, or a value
+            # fully consumed by fees) before marking the token redeemed, so it
+            # classifies as cashu_token_zero_value like the bearer/top-up paths
+            # rather than being forwarded as a free request.
+            if amount <= 0:
+                raise ValueError(
+                    f"Redeemed token amount must be positive, got {amount} {unit}"
+                )
+            redeemed = True
             headers = self.prepare_headers(dict(request.headers))
 
             request_id = getattr(request.state, "request_id", None)
@@ -4524,52 +4742,56 @@ class BaseUpstreamProvider:
                 },
             )
 
-            if "already spent" in error_message.lower():
+            # Once redeemed the token is spent, so a later forwarding failure
+            # must not surface as a retryable mint_unreachable (spent-token retry
+            # bait). Redemption classification only applies while not redeemed.
+            if redeemed:
                 return create_error_response(
-                    "token_already_spent",
-                    "The provided CASHU token has already been spent",
-                    400,
+                    "upstream_error",
+                    "Payment succeeded but the upstream request failed",
+                    502,
                     request=request,
-                    token=x_cashu_token,
+                    code="upstream_request_failed",
                 )
 
-            if "invalid token" in error_message.lower():
+            classified = classify_redemption_error(e)
+            if classified is None:
                 return create_error_response(
-                    "invalid_token",
-                    "The provided CASHU token is invalid",
-                    400,
+                    "api_error",
+                    "Internal error during token redemption",
+                    500,
                     request=request,
-                    token=x_cashu_token,
+                    code="internal_error",
                 )
-
-            if "mint error" in error_message.lower():
-                return create_error_response(
-                    "mint_error",
-                    f"CASHU mint error: {error_message}",
-                    422,
-                    request=request,
-                    token=x_cashu_token,
-                )
-
+            error_type, status_code, message, error_code = classified
+            # Echo the token back only when it is still spendable, so clients
+            # can recover it; a spent/consumed token is never re-offered.
+            echo_token = None if error_code in SPENT_TOKEN_CODES else x_cashu_token
             return create_error_response(
-                "cashu_error",
-                f"CASHU token processing failed: {error_message}",
-                400,
+                error_type,
+                message,
+                status_code,
                 request=request,
-                token=x_cashu_token,
+                token=echo_token,
+                code=error_code,
             )
 
     def _apply_provider_fee_to_model(self, model: Model) -> Model:
         """Apply provider fee to model's USD pricing and calculate max costs.
 
+        Cache rates missing from the upstream pricing feed are backfilled from
+        litellm's cost map first, so they carry the provider fee like every
+        other price component.
+
         Args:
             model: Model object to update
 
         Returns:
             Model with provider fee applied to pricing and max costs calculated
         """
+        base_pricing = backfill_cache_pricing(model.id, model.pricing)
         adjusted_pricing = Pricing.parse_obj(
-            {k: v * self.provider_fee for k, v in model.pricing.dict().items()}
+            {k: v * self.provider_fee for k, v in base_pricing.dict().items()}
         )
 
         temp_model = Model(
@@ -4722,14 +4944,11 @@ class BaseUpstreamProvider:
         """Refresh the in-memory models cache from upstream API."""
         try:
             async with create_session() as session:
-                stmt = select(UpstreamProviderRow).where(
-                    UpstreamProviderRow.base_url == self.base_url,
-                    UpstreamProviderRow.api_key == self.api_key,
+                provider = (
+                    await session.get(UpstreamProviderRow, self.db_id)
+                    if self.db_id is not None
+                    else None
                 )
-                result = await session.exec(stmt)
-
-                # .first() returns the object or None if not found
-                provider = result.first()
                 if not provider or not provider.id:
                     raise HTTPException(status_code=404, detail="Provider not found")
 

routstr/upstream/cache_breakpoints.py

@@ -0,0 +1,157 @@
+"""Inject explicit prompt-cache breakpoints into OpenAI-shaped requests.
+
+Some upstreams cache *explicitly*: the request must carry
+``cache_control: {"type": "ephemeral"}`` markers on the content blocks that
+should be cached. Two model families use this identical wire format:
+
+* **Anthropic Claude** — direct or via OpenRouter's ``anthropic/*`` models.
+* **Alibaba's explicit-cache models on OpenRouter** — ``qwen/qwen3-max``,
+  ``qwen/qwen-plus``, ``qwen/qwen3.6-plus``, ``qwen/qwen3-coder-plus``,
+  ``qwen/qwen3-coder-flash`` and ``deepseek/deepseek-v3.2`` — which OpenRouter
+  documents as using "the same syntax as Anthropic explicit caching".
+
+Every other provider routstr proxies (OpenAI, Azure, xAI/Grok, Groq, Moonshot,
+default DeepSeek, Gemini implicit, Fireworks) caches *automatically* and needs
+no markers — they are left untouched.
+
+A client that doesn't know it is talking to one of these models *through*
+routstr (e.g. an OpenAI-compatible coding agent pointed at a routstr URL) never
+emits the markers — it only adds them when it recognises the provider as
+OpenRouter. So caching silently never engages over routstr even though the same
+client caches fine talking to OpenRouter directly.
+
+This module restores caching by stamping the standard breakpoints onto the
+forwarded body — the system prompt, the last tool, and the last conversation
+message (the format allows up to four; we use three, matching the common
+agent convention) — but only when the client supplied none of its own, so
+explicit client control always wins. The caller is responsible for only
+applying this toward an upstream that accepts the markers (OpenRouter /
+Anthropic), so they never leak to a provider that would reject them.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+# The single ephemeral marker stamped onto each chosen breakpoint. A 5-minute
+# TTL (the default for ``ephemeral``) — deliberately not the 1h tier, which
+# carries a higher cache-write premium and should stay opt-in.
+EPHEMERAL_CACHE_CONTROL: dict[str, str] = {"type": "ephemeral"}
+
+# Alibaba's explicit-cache models on OpenRouter. Matched as substrings of the
+# model id (any spelling routstr carries). Snapshot endpoints that OpenRouter
+# documents as *not* supporting explicit caching (e.g. ``qwen3.5-plus-02-15``)
+# are different families and deliberately absent from this list.
+_ALIBABA_EXPLICIT_CACHE_SLUGS: tuple[str, ...] = (
+    "qwen3-max",
+    "qwen-plus",
+    "qwen3.6-plus",
+    "qwen3-coder-plus",
+    "qwen3-coder-flash",
+    "deepseek-v3.2",
+)
+
+
+def is_explicit_cache_model(model_id: str | None, *fallbacks: str | None) -> bool:
+    """True when the target model uses the explicit ``cache_control`` dialect.
+
+    Covers the Claude family (broadly — every Claude model supports it) and
+    Alibaba's documented explicit-cache models, across the id spellings routstr
+    carries: the OpenRouter id (``anthropic/claude-...``, ``qwen/qwen3-max``),
+    the bare upstream id, and any forwarded/canonical alias.
+    """
+    for candidate in (model_id, *fallbacks):
+        if not candidate:
+            continue
+        lowered = candidate.lower()
+        if "claude" in lowered or "anthropic/" in lowered:
+            return True
+        if any(slug in lowered for slug in _ALIBABA_EXPLICIT_CACHE_SLUGS):
+            return True
+    return False
+
+
+def _has_cache_control(obj: Any) -> bool:
+    """Recursively detect any client-supplied ``cache_control`` marker."""
+    if isinstance(obj, dict):
+        if "cache_control" in obj:
+            return True
+        return any(_has_cache_control(v) for v in obj.values())
+    if isinstance(obj, list):
+        return any(_has_cache_control(v) for v in obj)
+    return False
+
+
+def body_has_cache_control(data: dict) -> bool:
+    """True when the request already carries cache_control on messages/tools."""
+    return _has_cache_control(data.get("messages")) or _has_cache_control(
+        data.get("tools")
+    )
+
+
+def _stamp_text_content(message: dict) -> bool:
+    """Add the ephemeral marker to a message's last text block.
+
+    A string content is promoted to the array form Anthropic requires for
+    cache markers; an existing array gets the marker on its last text part.
+    Returns True when a marker was placed.
+    """
+    content = message.get("content")
+    if isinstance(content, str):
+        if not content:
+            return False
+        message["content"] = [
+            {
+                "type": "text",
+                "text": content,
+                "cache_control": dict(EPHEMERAL_CACHE_CONTROL),
+            }
+        ]
+        return True
+    if isinstance(content, list):
+        for part in reversed(content):
+            if isinstance(part, dict) and part.get("type") == "text":
+                part["cache_control"] = dict(EPHEMERAL_CACHE_CONTROL)
+                return True
+    return False
+
+
+def _stamp_system_prompt(messages: list) -> None:
+    for message in messages:
+        if isinstance(message, dict) and message.get("role") in (
+            "system",
+            "developer",
+        ):
+            _stamp_text_content(message)
+            return
+
+
+def _stamp_last_tool(tools: Any) -> None:
+    if isinstance(tools, list) and tools and isinstance(tools[-1], dict):
+        tools[-1]["cache_control"] = dict(EPHEMERAL_CACHE_CONTROL)
+
+
+def _stamp_last_conversation_message(messages: list) -> None:
+    for message in reversed(messages):
+        if isinstance(message, dict) and message.get("role") in ("user", "assistant"):
+            if _stamp_text_content(message):
+                return
+
+
+def inject_anthropic_cache_breakpoints(data: dict) -> bool:
+    """Stamp ephemeral cache breakpoints onto an OpenAI-shaped chat body.
+
+    Mutates ``data`` in place (the established convention in
+    ``prepare_request_body``) and returns True when anything changed. No-ops
+    when the body isn't chat-shaped or the client already set cache_control.
+    """
+    messages = data.get("messages")
+    if not isinstance(messages, list) or not messages:
+        return False
+    if body_has_cache_control(data):
+        return False
+
+    _stamp_system_prompt(messages)
+    _stamp_last_tool(data.get("tools"))
+    _stamp_last_conversation_message(messages)
+    return True

routstr/upstream/deepseek_v4_pricing_shim.py

@@ -0,0 +1,73 @@
+"""TEMPORARY: local DeepSeek V4 pricing shim.
+
+litellm's bundled cost map does not yet ship ``deepseek-v4-flash`` /
+``deepseek-v4-pro``. Without an entry, ``backfill_cache_pricing`` cannot find a
+``cache_read_input_token_cost`` and cache reads fall back to the full input
+rate — a large overcharge on cache hits (DeepSeek V4 hits are ~0.008-0.02x
+input, i.e. cached tokens cost 50-120x less than regular input).
+
+This module injects the missing entries into ``litellm.model_cost`` at startup
+so the existing backfill path resolves them. Rates mirror the canonical
+``deepseek`` provider entries now in litellm's ``model_prices`` map
+(``input_cost_per_token`` is the cache-*miss* rate;
+``cache_read_input_token_cost`` is the cache-*hit* rate), sourced from
+https://api-docs.deepseek.com/quick_start/pricing via
+https://github.com/BerriAI/litellm/pull/26380 (issue
+https://github.com/BerriAI/litellm/issues/30430).
+
+=== REMOVAL (once litellm ships these models) ===
+Delete this file and the single ``register_deepseek_v4_pricing()`` call in
+``routstr/core/main.py``. Nothing else depends on it. Entries are only added
+when absent, so a stale shim is harmless after upstream lands — but remove it.
+"""
+
+import litellm
+
+from ..core import get_logger
+
+logger = get_logger(__name__)
+
+# USD per token. Mirrors the canonical ``deepseek`` provider entries in
+# litellm's model_prices map (source: DeepSeek API pricing docs). Keep these in
+# sync with ``litellm.model_cost["deepseek/deepseek-v4-*"]``.
+_DEEPSEEK_V4_RATES: dict[str, dict[str, float]] = {
+    "deepseek-v4-flash": {
+        "input_cost_per_token": 1.4e-07,
+        "output_cost_per_token": 2.8e-07,
+        "cache_read_input_token_cost": 2.8e-09,
+        "cache_creation_input_token_cost": 0.0,
+        "input_cost_per_token_cache_hit": 2.8e-09,
+    },
+    "deepseek-v4-pro": {
+        "input_cost_per_token": 4.35e-07,
+        "output_cost_per_token": 8.7e-07,
+        "cache_read_input_token_cost": 3.625e-09,
+        "cache_creation_input_token_cost": 0.0,
+        "input_cost_per_token_cache_hit": 3.625e-09,
+    },
+}
+
+
+def register_deepseek_v4_pricing() -> None:
+    """Inject DeepSeek V4 pricing into ``litellm.model_cost`` if absent.
+
+    Idempotent and non-destructive: a key already present in the cost map
+    (e.g. once litellm ships it) is left untouched. Registers both the bare
+    (``deepseek-v4-flash``) and prefixed (``deepseek/deepseek-v4-flash``)
+    spellings since ``backfill_cache_pricing`` tries both.
+    """
+    added = []
+    for bare, rates in _DEEPSEEK_V4_RATES.items():
+        for key in (bare, f"deepseek/{bare}"):
+            if key in litellm.model_cost:
+                continue
+            entry: dict[str, object] = dict(rates)
+            entry["litellm_provider"] = "deepseek"
+            entry["mode"] = "chat"
+            litellm.model_cost[key] = entry
+            added.append(key)
+    if added:
+        logger.info(
+            "Registered temporary DeepSeek V4 pricing shim",
+            extra={"models": added},
+        )

routstr/upstream/fireworks.py

@@ -20,7 +20,7 @@ class FireworksUpstreamProvider(BaseUpstreamProvider):
         )
 
     @classmethod
-    def from_db_row(
+    def _build_from_row(
         cls, provider_row: "UpstreamProviderRow"
     ) -> "FireworksUpstreamProvider":
         return cls(

routstr/upstream/gemini.py

@@ -51,7 +51,7 @@ class GeminiUpstreamProvider(BaseUpstreamProvider):
         return self._client
 
     @classmethod
-    def from_db_row(
+    def _build_from_row(
         cls, provider_row: "UpstreamProviderRow"
     ) -> "GeminiUpstreamProvider":
         return cls(

routstr/upstream/generic.py

@@ -5,6 +5,12 @@ from typing import TYPE_CHECKING
 import httpx
 
 from .base import BaseUpstreamProvider
+from .pricing_resolver import (
+    FallbackPricingResolver,
+    ResolvedPricing,
+    _as_float,
+    estimate_context_length,
+)
 
 if TYPE_CHECKING:
     from ..core.db import UpstreamProviderRow
@@ -45,7 +51,7 @@ class GenericUpstreamProvider(BaseUpstreamProvider):
         )
 
     @classmethod
-    def from_db_row(
+    def _build_from_row(
         cls, provider_row: "UpstreamProviderRow"
     ) -> "GenericUpstreamProvider":
         return cls(
@@ -64,6 +70,40 @@ class GenericUpstreamProvider(BaseUpstreamProvider):
             "platform_url": cls.platform_url,
         }
 
+    def _native_pricing(
+        self, model_id: str, model_spec: dict
+    ) -> ResolvedPricing | None:
+        """Read pricing/metadata from Venice's bespoke ``model_spec`` schema.
+
+        Returns ``None`` when the upstream reported no *usable* native price —
+        absent, non-numeric, negative, or both-zero — so the caller falls
+        through to the shared resolution chain instead of fabricating a number
+        or trusting a bogus one. This mirrors the money-safety guards the
+        litellm and OpenRouter rungs already apply: a both-zero price would
+        serve the model free, a negative one would credit the caller, and a
+        non-numeric string would otherwise throw and drop the whole catalog.
+        """
+        pricing_info = model_spec.get("pricing", {})
+        input_usd = _as_float(pricing_info.get("input", {}).get("usd"))
+        output_usd = _as_float(pricing_info.get("output", {}).get("usd"))
+        if input_usd is None or output_usd is None:
+            return None
+        if input_usd < 0 or output_usd < 0 or (input_usd == 0 and output_usd == 0):
+            return None
+
+        capabilities = model_spec.get("capabilities", {})
+        input_modalities = ["text"]
+        if capabilities.get("supportsVision", False):
+            input_modalities.append("image")
+
+        return ResolvedPricing(
+            prompt=input_usd / 1_000_000,
+            completion=output_usd / 1_000_000,
+            context_length=model_spec.get("availableContextTokens"),
+            source="native",
+            input_modalities=input_modalities,
+        )
+
     async def fetch_models(self) -> list[Model]:
         """Fetch models from upstream API using /models endpoint."""
         from ..payment.models import Architecture, Model, Pricing, TopProvider
@@ -78,6 +118,7 @@ class GenericUpstreamProvider(BaseUpstreamProvider):
                 response.raise_for_status()
                 data = response.json()
 
+                resolver = FallbackPricingResolver()
                 models_list = []
                 for model_data in data.get("data", []):
                     model_id = model_data.get("id", "")
@@ -89,41 +130,44 @@ class GenericUpstreamProvider(BaseUpstreamProvider):
                     owned_by = model_data.get("owned_by", "unknown")
                     model_spec = model_data.get("model_spec", {})
 
-                    context_length = 4096
-                    if model_spec.get("availableContextTokens"):
-                        context_length = model_spec["availableContextTokens"]
-                    elif any(
-                        pattern in model_id.lower() for pattern in ["32k", "32000"]
-                    ):
-                        context_length = 32768
-                    elif any(
-                        pattern in model_id.lower() for pattern in ["16k", "16000"]
-                    ):
-                        context_length = 16384
-                    elif any(pattern in model_id.lower() for pattern in ["8k", "8000"]):
-                        context_length = 8192
-                    elif "gpt-4" in model_id.lower():
-                        context_length = 8192
-                    elif "claude" in model_id.lower():
-                        context_length = 200000
+                    resolved = self._native_pricing(model_id, model_spec)
+                    if resolved is None:
+                        resolved = await resolver.resolve(model_id)
 
-                    pricing_info = model_spec.get("pricing", {})
-                    input_pricing = pricing_info.get("input", {})
-                    output_pricing = pricing_info.get("output", {})
+                    if resolved is None:
+                        # Fail closed: never invent a price. Import the model
+                        # disabled with a warning so the operator can price it
+                        # (the admin UI surfaces disabled remote models).
+                        logger.warning(
+                            f"No pricing source resolved for '{model_id}' from "
+                            f"{self.upstream_name}; importing it disabled",
+                            extra={"model_id": model_id, "base_url": self.base_url},
+                        )
+                        resolved = ResolvedPricing(
+                            prompt=0.0,
+                            completion=0.0,
+                            context_length=None,
+                            source="unresolved",
+                        )
+                        enabled = False
+                    else:
+                        enabled = True
 
-                    prompt_price = input_pricing.get("usd", 0.001) / 1000000
-                    completion_price = output_pricing.get("usd", 0.001) / 1000000
+                    # Prefer the source's own modality string (OpenRouter ships
+                    # one, e.g. "text+image->text"); otherwise derive it from the
+                    # captured input/output modalities in the same "in->out" shape
+                    # rather than flattening vision models to "text->text".
+                    modality = resolved.modality or (
+                        f"{'+'.join(resolved.input_modalities)}"
+                        f"->{'+'.join(resolved.output_modalities)}"
+                    )
 
-                    capabilities = model_spec.get("capabilities", {})
-                    input_modalities = ["text"]
-                    output_modalities = ["text"]
-
-                    if capabilities.get("supportsVision", False):
-                        input_modalities.append("image")
-
-                    modality = "text"
-                    if capabilities.get("supportsVision", False):
-                        modality = "text->text"
+                    # A source can carry a price but no context (e.g. a litellm
+                    # entry missing max_input_tokens); fall back to an id-based
+                    # estimate so we never persist a zero-length window.
+                    context_length = resolved.context_length or estimate_context_length(
+                        model_id
+                    )
 
                     spec_name = model_spec.get("name", model_name)
                     description = f"{spec_name}"
@@ -139,30 +183,33 @@ class GenericUpstreamProvider(BaseUpstreamProvider):
                             context_length=context_length,
                             architecture=Architecture(
                                 modality=modality,
-                                input_modalities=input_modalities,
-                                output_modalities=output_modalities,
-                                tokenizer="unknown",
-                                instruct_type=None,
+                                input_modalities=resolved.input_modalities,
+                                output_modalities=resolved.output_modalities,
+                                tokenizer=resolved.tokenizer,
+                                instruct_type=resolved.instruct_type,
                             ),
                             pricing=Pricing(
-                                prompt=prompt_price,
-                                completion=completion_price,
+                                prompt=resolved.prompt,
+                                completion=resolved.completion,
                                 request=0.0,
                                 image=0.0,
                                 web_search=0.0,
                                 internal_reasoning=0.0,
-                                max_prompt_cost=0.001,
-                                max_completion_cost=0.001,
-                                max_cost=0.001,
+                                input_cache_read=resolved.input_cache_read,
+                                input_cache_write=resolved.input_cache_write,
                             ),
                             sats_pricing=None,
                             per_request_limits=None,
                             top_provider=TopProvider(
                                 context_length=context_length,
-                                max_completion_tokens=context_length // 2,
-                                is_moderated=False,
+                                max_completion_tokens=(
+                                    resolved.max_completion_tokens
+                                    if resolved.max_completion_tokens is not None
+                                    else context_length // 2
+                                ),
+                                is_moderated=bool(resolved.is_moderated),
                             ),
-                            enabled=True,
+                            enabled=enabled,
                             upstream_provider_id=None,
                             canonical_slug=None,
                         )

routstr/upstream/groq.py

@@ -20,7 +20,7 @@ class GroqUpstreamProvider(BaseUpstreamProvider):
         )
 
     @classmethod
-    def from_db_row(cls, provider_row: "UpstreamProviderRow") -> "GroqUpstreamProvider":
+    def _build_from_row(cls, provider_row: "UpstreamProviderRow") -> "GroqUpstreamProvider":
         return cls(
             api_key=provider_row.api_key,
             provider_fee=provider_row.provider_fee,

routstr/upstream/helpers.py

@@ -12,6 +12,7 @@ from sqlmodel import select
 
 from ..core import get_logger
 from ..core.db import AsyncSession, ModelRow, UpstreamProviderRow, create_session
+from ..core.provider_slugs import allocate_unique_provider_slug
 from ..payment.models import Model
 from .base import BaseUpstreamProvider
 
@@ -93,12 +94,10 @@ async def get_all_models_with_overrides(
         provider_result = await session.exec(select(UpstreamProviderRow))
         providers_by_id = {p.id: p for p in provider_result.all()}
 
-        overrides_by_id: dict[str, tuple[ModelRow, float]] = {
-            row.id: (
+        overrides_by_key: dict[tuple[str, int], tuple[ModelRow, float]] = {
+            (row.id.lower(), row.upstream_provider_id): (
                 row,
-                providers_by_id[row.upstream_provider_id].provider_fee
-                if row.upstream_provider_id in providers_by_id
-                else 1.01,
+                providers_by_id[row.upstream_provider_id].provider_fee,
             )
             for row in override_rows
             if row.upstream_provider_id is not None
@@ -106,17 +105,28 @@ async def get_all_models_with_overrides(
             and providers_by_id[row.upstream_provider_id].enabled
         }
 
-    all_models: dict[str, Model] = {}
+    all_models: dict[tuple[str, str], Model] = {}
 
     for upstream in upstreams:
+        upstream_db_id = getattr(upstream, "db_id", None)
+        provider_key = (
+            f"db:{upstream_db_id}"
+            if isinstance(upstream_db_id, int)
+            else f"{getattr(upstream, 'provider_type', '')}|{getattr(upstream, 'base_url', '')}"
+        )
         for model in upstream.get_cached_models():
-            if model.id in overrides_by_id:
-                override_row, provider_fee = overrides_by_id[model.id]
-                all_models[model.id] = _row_to_model(
+            model_key = (
+                (model.id.lower(), upstream_db_id)
+                if isinstance(upstream_db_id, int)
+                else None
+            )
+            if model_key is not None and model_key in overrides_by_key:
+                override_row, provider_fee = overrides_by_key[model_key]
+                all_models[(model.id.lower(), provider_key)] = _row_to_model(
                     override_row, apply_provider_fee=True, provider_fee=provider_fee
                 )
             elif model.enabled:
-                all_models[model.id] = model
+                all_models[(model.id.lower(), provider_key)] = model
 
     return list(all_models.values())
 
@@ -215,9 +225,6 @@ async def init_upstreams() -> list[BaseUpstreamProvider]:
 
             provider = _instantiate_provider(provider_row)
             if provider:
-                # Keep provider DB id on runtime instance so model mapping can
-                # bind DB overrides to the correct upstream.
-                setattr(provider, "db_id", provider_row.id)
                 await provider.refresh_models_cache()
                 logger.debug(
                     f"Initialized {provider_row.provider_type} provider",
@@ -250,6 +257,7 @@ async def _seed_providers_from_settings(
 
     providers_to_add: list[UpstreamProviderRow] = []
     seeded_provider_keys: set[tuple[str, str]] = set()
+    reserved_slugs: set[str] = set()
 
     provider_classes_by_type = {
         cls.provider_type: cls
@@ -264,6 +272,7 @@ async def _seed_providers_from_settings(
         ("PERPLEXITY_API_KEY", "perplexity", None, None),
         ("FIREWORKS_API_KEY", "fireworks", None, None),
         ("XAI_API_KEY", "xai", None, None),
+        ("TINFOIL_API_KEY", "tinfoil", None, None),
     ]
 
     for env_key, provider_type, _, _ in env_mappings:
@@ -279,8 +288,13 @@ async def _seed_providers_from_settings(
                     )
                 )
                 if not result.first():
+                    slug = await allocate_unique_provider_slug(
+                        session, provider_type, reserved_slugs
+                    )
+                    reserved_slugs.add(slug)
                     providers_to_add.append(
                         UpstreamProviderRow(
+                            slug=slug,
                             provider_type=provider_type,
                             base_url=base_url,
                             api_key=api_key,
@@ -299,8 +313,13 @@ async def _seed_providers_from_settings(
             )
         )
         if not result.first():
+            slug = await allocate_unique_provider_slug(
+                session, "ollama", reserved_slugs
+            )
+            reserved_slugs.add(slug)
             providers_to_add.append(
                 UpstreamProviderRow(
+                    slug=slug,
                     provider_type="ollama",
                     base_url=ollama_base_url,
                     api_key=ollama_api_key,
@@ -320,8 +339,13 @@ async def _seed_providers_from_settings(
                 )
             )
             if not result.first():
+                slug = await allocate_unique_provider_slug(
+                    session, "azure", reserved_slugs
+                )
+                reserved_slugs.add(slug)
                 providers_to_add.append(
                     UpstreamProviderRow(
+                        slug=slug,
                         provider_type="azure",
                         base_url=base_url,
                         api_key=api_key,
@@ -342,8 +366,13 @@ async def _seed_providers_from_settings(
                 )
             )
             if not result.first():
+                slug = await allocate_unique_provider_slug(
+                    session, "custom", reserved_slugs
+                )
+                reserved_slugs.add(slug)
                 providers_to_add.append(
                     UpstreamProviderRow(
+                        slug=slug,
                         provider_type="custom",
                         base_url=base_url,
                         api_key=api_key,
@@ -356,7 +385,7 @@ async def _seed_providers_from_settings(
         session.add(provider)
         logger.info(
             f"Seeding {provider.provider_type} provider",  # type: ignore[str-format]
-            extra={"base_url": provider.base_url},
+            extra={"base_url": provider.base_url, "slug": provider.slug},
         )
 
 
@@ -391,9 +420,7 @@ def _instantiate_provider(
             return provider
 
         if provider_row.provider_type == "custom":
-            return BaseUpstreamProvider(
-                provider_row.base_url, provider_row.api_key, provider_row.provider_fee
-            )
+            return BaseUpstreamProvider.from_db_row(provider_row)
 
         logger.error(
             f"Unknown provider type: {provider_row.provider_type}",

routstr/upstream/messages_dispatch.py

@@ -29,7 +29,9 @@ import litellm
 
 from ..core import get_logger
 from ..core.exceptions import UpstreamError
+from ..core.redaction import redact_org_ids
 from ..payment.models import Model
+from .rate_limit import classify_rate_limit
 
 logger = get_logger(__name__)
 
@@ -505,23 +507,33 @@ async def dispatch_anthropic_messages(
     try:
         result = await litellm.anthropic.messages.acreate(**kwargs)
     except Exception as exc:
-        exc_message = getattr(exc, "message", None) or str(exc) or repr(exc)
+        raw_message = getattr(exc, "message", None) or str(exc) or repr(exc)
+        # Redact provider account identifiers before the message reaches logs
+        # or the surfaced error.
+        exc_message = redact_org_ids(raw_message)
         exc_status = getattr(exc, "status_code", None)
         exc_response = getattr(exc, "response", None)
         response_text = None
         if exc_response is not None:
             try:
-                response_text = getattr(exc_response, "text", str(exc_response))
+                response_text = redact_org_ids(
+                    getattr(exc_response, "text", str(exc_response))
+                )
             except Exception:
                 response_text = "<unreadable>"
+        status_for_classify = exc_status if isinstance(exc_status, int) else 502
+        rate_limit = classify_rate_limit(
+            status_for_classify, exc_message, getattr(exc, "headers", None)
+        )
         logger.error(
             "litellm dispatch failed",
             extra={
                 "error": exc_message,
                 "error_type": type(exc).__name__,
                 "status_code": exc_status,
+                "error_code": rate_limit.code if rate_limit else None,
                 "llm_provider": getattr(exc, "llm_provider", None),
-                "body": getattr(exc, "body", None),
+                "body": redact_org_ids(str(getattr(exc, "body", "") or "")) or None,
                 "response_text": response_text,
                 "model": litellm_model,
                 "api_base": base_url,
@@ -529,7 +541,9 @@ async def dispatch_anthropic_messages(
         )
         raise UpstreamError(
             f"Upstream error via litellm: {exc_message}",
-            status_code=exc_status if isinstance(exc_status, int) else 502,
+            status_code=status_for_classify,
+            code=rate_limit.code if rate_limit else None,
+            details=rate_limit.as_details() if rate_limit else None,
         ) from exc
 
     if not client_stream and hasattr(result, "__aiter__"):

routstr/upstream/ollama.py

@@ -43,7 +43,7 @@ class OllamaUpstreamProvider(BaseUpstreamProvider):
         )
 
     @classmethod
-    def from_db_row(
+    def _build_from_row(
         cls, provider_row: "UpstreamProviderRow"
     ) -> "OllamaUpstreamProvider":
         return cls(

routstr/upstream/openai.py

@@ -20,7 +20,7 @@ class OpenAIUpstreamProvider(BaseUpstreamProvider):
         )
 
     @classmethod
-    def from_db_row(
+    def _build_from_row(
         cls, provider_row: "UpstreamProviderRow"
     ) -> "OpenAIUpstreamProvider":
         return cls(

routstr/upstream/openrouter.py

@@ -57,7 +57,7 @@ class OpenRouterUpstreamProvider(BaseUpstreamProvider):
         )
 
     @classmethod
-    def from_db_row(
+    def _build_from_row(
         cls, provider_row: "UpstreamProviderRow"
     ) -> "OpenRouterUpstreamProvider":
         return cls(

routstr/upstream/perplexity.py

@@ -23,7 +23,7 @@ class PerplexityUpstreamProvider(BaseUpstreamProvider):
         )
 
     @classmethod
-    def from_db_row(
+    def _build_from_row(
         cls, provider_row: "UpstreamProviderRow"
     ) -> "PerplexityUpstreamProvider":
         return cls(

routstr/upstream/ppqai.py

@@ -8,6 +8,7 @@ from pydantic.v1 import BaseModel, Field
 from ..core.logging import get_logger
 from ..payment.models import Architecture, Model, Pricing, async_fetch_openrouter_models
 from .base import BaseUpstreamProvider, TopupData
+from .ehbp import EHBPForwardingTarget
 
 if TYPE_CHECKING:
     from ..core.db import UpstreamProviderRow
@@ -39,6 +40,10 @@ class PPQAIUpstreamProvider(BaseUpstreamProvider):
     default_base_url = "https://api.ppq.ai"
     platform_url = "https://ppq.ai/api-docs"
     IGNORED_MODEL_IDS: list[str] = ["auto"]
+    # PPQ.AI has a private encrypted endpoint, but this proxy currently has no
+    # provider-attested usage extractor/model binding for it. Keep EHBP disabled
+    # until a ConfidentialInferenceProfile can bill it without max-cost fallback.
+    supports_ehbp = False
 
     def __init__(self, api_key: str, provider_fee: float = 1.0):
         super().__init__(
@@ -46,7 +51,7 @@ class PPQAIUpstreamProvider(BaseUpstreamProvider):
         )
 
     @classmethod
-    def from_db_row(
+    def _build_from_row(
         cls, provider_row: "UpstreamProviderRow"
     ) -> "PPQAIUpstreamProvider":
         return cls(
@@ -70,6 +75,20 @@ class PPQAIUpstreamProvider(BaseUpstreamProvider):
     def transform_model_name(self, model_id: str) -> str:
         return model_id
 
+    def get_ehbp_forwarding_target(
+        self, path: str, model_obj: Model
+    ) -> EHBPForwardingTarget:
+        """Return the PPQ.AI private enclave target for EHBP requests.
+
+        PPQ.AI exposes EHBP-aware inference under /private/v1/... separate
+        from the public /v1/... endpoint. The encrypted body remains opaque to
+        Routstr, so PPQ.AI also needs X-Private-Model for routing/billing.
+        """
+        return EHBPForwardingTarget(
+            url=f"{self.base_url.rstrip('/')}/private/{path.lstrip('/')}",
+            headers={"X-Private-Model": model_obj.forwarded_model_id or model_obj.id},
+        )
+
     @classmethod
     async def create_account_static(cls) -> dict[str, object]:
         """Create a new PPQ.AI account without requiring an instance.
@@ -229,17 +248,14 @@ class PPQAIUpstreamProvider(BaseUpstreamProvider):
                 f"Disabling PPQ.AI provider ({self.base_url}) due to insufficient balance",
                 extra={"error": error_message},
             )
-            from sqlmodel import select
-
             from ..core.db import UpstreamProviderRow, create_session
 
             async with create_session() as session:
-                statement = select(UpstreamProviderRow).where(
-                    UpstreamProviderRow.base_url == self.base_url,
-                    UpstreamProviderRow.api_key == self.api_key,
+                provider = (
+                    await session.get(UpstreamProviderRow, self.db_id)
+                    if self.db_id is not None
+                    else None
                 )
-                result = await session.exec(statement)
-                provider = result.first()
 
                 if provider:
                     provider.enabled = False

routstr/upstream/pricing_resolver.py

@@ -0,0 +1,203 @@
+"""Shared price/metadata resolution chain for upstream model discovery.
+
+Most OpenAI-compatible ``/models`` responses carry no pricing. Rather than let
+a provider fabricate one, this module resolves a model through decreasingly
+trustworthy sources — litellm's bundled cost map (curated list prices, mirrors
+provider docs), then the OpenRouter feed (resale prices, broader coverage) —
+and returns ``None`` when none of them know the model, so the caller can fail
+closed instead of inventing a number.
+
+Provider-native pricing (a gateway's own ``/models`` schema, e.g. Venice's
+``model_spec``) is authoritative and handled by the provider before this chain
+is consulted; only the shared fallback lives here so a later refactor can hoist
+it into the base provider unchanged.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+
+
+@dataclass
+class ResolvedPricing:
+    """Per-token pricing plus whatever metadata the answering source carried.
+
+    Prices are USD per token. ``source`` records provenance
+    (``native``/``litellm``/``openrouter``/``unresolved``) so later work can
+    surface where each price came from.
+    """
+
+    prompt: float
+    completion: float
+    context_length: int | None
+    source: str
+    modality: str | None = None
+    max_completion_tokens: int | None = None
+    input_cache_read: float = 0.0
+    input_cache_write: float = 0.0
+    input_modalities: list[str] = field(default_factory=lambda: ["text"])
+    output_modalities: list[str] = field(default_factory=lambda: ["text"])
+    tokenizer: str = "unknown"
+    instruct_type: str | None = None
+    is_moderated: bool | None = None
+
+
+def estimate_context_length(model_id: str) -> int:
+    """Best-effort context window from a model id when no source reports one.
+
+    The last rung of the fallback chain, reached only for a model whose price
+    resolved but whose context did not (or that imported disabled). Context is
+    not a billing input, so a rough id-based guess is acceptable here where a
+    guessed *price* never would be.
+    """
+    lowered = model_id.lower()
+    if any(pattern in lowered for pattern in ["32k", "32000"]):
+        return 32768
+    if any(pattern in lowered for pattern in ["16k", "16000"]):
+        return 16384
+    if any(pattern in lowered for pattern in ["8k", "8000"]):
+        return 8192
+    if "gpt-4" in lowered:
+        return 8192
+    if "claude" in lowered:
+        return 200000
+    return 4096
+
+
+def _as_float(value: object) -> float | None:
+    """OpenRouter reports prices as strings; coerce, ``None`` if unparseable."""
+    try:
+        return float(value)  # type: ignore[arg-type]
+    except (TypeError, ValueError):
+        return None
+
+
+def _as_int(value: object) -> int | None:
+    """Coerce an already-numeric token count to ``int``, else ``None``."""
+    return int(value) if isinstance(value, (int, float)) else None
+
+
+def _from_litellm(model_id: str) -> ResolvedPricing | None:
+    # Lazy import so the resolver stays import-light and shares the exact
+    # lookup semantics used by cache-rate backfill.
+    from ..payment.models import litellm_cost_entry
+
+    info = litellm_cost_entry(model_id)
+    if info is None:
+        return None
+
+    prompt = info.get("input_cost_per_token")
+    completion = info.get("output_cost_per_token")
+    if not isinstance(prompt, (int, float)) or not isinstance(completion, (int, float)):
+        return None
+    # A both-zero entry is litellm listing a model without a real price (free
+    # moderation/rerank tiers do this) — treating 0/0 as resolved would serve
+    # the model for free. Reject it (and any negative) so the caller falls
+    # through, mirroring async_fetch_openrouter_models' _has_valid_pricing.
+    if prompt < 0 or completion < 0 or (prompt == 0 and completion == 0):
+        return None
+
+    input_modalities = ["text"]
+    if info.get("supports_vision"):
+        input_modalities.append("image")
+
+    return ResolvedPricing(
+        prompt=float(prompt),
+        completion=float(completion),
+        # max_input_tokens is the context window; max_tokens is litellm's
+        # completion cap (it tracks max_output_tokens for ~94% of models), so
+        # it is never a context source. A missing window falls to the id-based
+        # estimate downstream rather than borrowing the output cap.
+        context_length=_as_int(info.get("max_input_tokens")),
+        source="litellm",
+        max_completion_tokens=_as_int(info.get("max_output_tokens")),
+        input_cache_read=float(info.get("cache_read_input_token_cost") or 0.0),
+        input_cache_write=float(info.get("cache_creation_input_token_cost") or 0.0),
+        input_modalities=input_modalities,
+    )
+
+
+def _match_openrouter(model_id: str, feed: list[dict]) -> dict | None:
+    """Find ``model_id`` in the OpenRouter feed, exact id before bare tail.
+
+    Bare-tail matching (``deepseek-chat`` ↔ ``deepseek/deepseek-chat``) is a
+    looser, lower-trust match — OpenRouter fans a model out across resellers —
+    so an exact id match always wins first. When several entries share the bare
+    tail, the one with the highest *combined* (prompt + completion) per-token
+    cost wins: the choice must be deterministic (not feed-order-dependent) and
+    money-safe whichever way traffic leans, since undercharging is the hazard.
+    Ranking on prompt alone could pick an entry that is cheap on input but dear
+    on output. The live feed has no such collisions today; this only governs
+    the latent case.
+    """
+    bare = model_id.split("/", 1)[-1]
+    exact = next((m for m in feed if m.get("id") == model_id), None)
+    if exact is not None:
+        return exact
+    matches = [m for m in feed if m.get("id", "").split("/", 1)[-1] == bare]
+    if not matches:
+        return None
+
+    def _combined_cost(m: dict) -> float:
+        pricing = m.get("pricing", {})
+        return (_as_float(pricing.get("prompt")) or 0.0) + (
+            _as_float(pricing.get("completion")) or 0.0
+        )
+
+    return max(matches, key=_combined_cost)
+
+
+def _from_openrouter(model_id: str, feed: list[dict]) -> ResolvedPricing | None:
+    entry = _match_openrouter(model_id, feed)
+    if entry is None:
+        return None
+
+    pricing = entry.get("pricing", {})
+    prompt = _as_float(pricing.get("prompt"))
+    completion = _as_float(pricing.get("completion"))
+    if prompt is None or completion is None:
+        return None
+
+    architecture = entry.get("architecture", {})
+    top_provider = entry.get("top_provider", {})
+
+    return ResolvedPricing(
+        prompt=prompt,
+        completion=completion,
+        context_length=_as_int(entry.get("context_length")),
+        source="openrouter",
+        modality=architecture.get("modality"),
+        max_completion_tokens=_as_int(top_provider.get("max_completion_tokens")),
+        input_cache_read=_as_float(pricing.get("input_cache_read")) or 0.0,
+        input_cache_write=_as_float(pricing.get("input_cache_write")) or 0.0,
+        input_modalities=architecture.get("input_modalities") or ["text"],
+        output_modalities=architecture.get("output_modalities") or ["text"],
+        tokenizer=architecture.get("tokenizer") or "unknown",
+        instruct_type=architecture.get("instruct_type"),
+        is_moderated=top_provider.get("is_moderated"),
+    )
+
+
+class FallbackPricingResolver:
+    """Resolves models via litellm → OpenRouter for one discovery pass.
+
+    The OpenRouter catalog is fetched at most once and only when a model
+    actually misses litellm, so a provider full of litellm-known models never
+    touches the network. Instantiate one per ``fetch_models`` call.
+    """
+
+    def __init__(self) -> None:
+        self._openrouter_feed: list[dict] | None = None
+
+    async def resolve(self, model_id: str) -> ResolvedPricing | None:
+        """Resolve ``model_id``; ``None`` if no source knows it."""
+        resolved = _from_litellm(model_id)
+        if resolved is not None:
+            return resolved
+
+        if self._openrouter_feed is None:
+            # Lazy import so tests can patch the feed at its source.
+            from ..payment.models import async_fetch_openrouter_models
+
+            self._openrouter_feed = await async_fetch_openrouter_models()
+        return _from_openrouter(model_id, self._openrouter_feed)

routstr/upstream/rate_limit.py

@@ -0,0 +1,136 @@
+"""Detection and parsing of upstream provider rate-limit errors.
+
+Upstream OpenAI-compatible providers signal rate limits via HTTP 429 and/or a
+human-readable message such as::
+
+    Rate limit reached for gpt-5.5-2026-04-23 (for limit gpt-5.5) in organization
+    org-XXXX on tokens per min (TPM): Limit 180000000, Used 180000000,
+    Requested 8929. Please try again in 2ms.
+
+This module classifies those failures into a stable :data:`UPSTREAM_RATE_LIMIT`
+code and extracts useful debugging fields. All retained text is redacted of
+organization IDs first.
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import asdict, dataclass
+
+from ..core.redaction import redact_org_ids
+
+# Stable error code callers can switch on to distinguish upstream rate limits
+# from generic request failures. The literal value matches the identifier named
+# in issue #555 ("UPSTREAM_RATE_LIMIT") so the public API contract is exact.
+UPSTREAM_RATE_LIMIT = "UPSTREAM_RATE_LIMIT"
+
+# Message fragments that indicate a rate-limit even when the status code is not
+# 429 (some providers wrap it in a 400/500 envelope).
+_RATE_LIMIT_MARKERS = (
+    "rate limit reached",
+    "rate_limit_exceeded",
+    "rate limit exceeded",
+    "too many requests",
+)
+
+_MODEL_RE = re.compile(r"Rate limit reached for ([^\s(]+)", re.IGNORECASE)
+_LIMIT_NAME_RE = re.compile(r"\(for limit ([^)]+)\)", re.IGNORECASE)
+_METRIC_RE = re.compile(r"on ([a-z ]+\((?:TPM|RPM|TPD|RPD|IPM)\))", re.IGNORECASE)
+_LIMIT_RE = re.compile(r"Limit (\d+)", re.IGNORECASE)
+_USED_RE = re.compile(r"Used (\d+)", re.IGNORECASE)
+_REQUESTED_RE = re.compile(r"Requested (\d+)", re.IGNORECASE)
+_RETRY_RE = re.compile(r"try again in ([\d.]+)\s*(ms|s)", re.IGNORECASE)
+
+
+@dataclass
+class RateLimitInfo:
+    """Structured, redaction-safe view of an upstream rate-limit error."""
+
+    code: str
+    message: str
+    model: str | None = None
+    limit_name: str | None = None
+    metric: str | None = None
+    limit: int | None = None
+    used: int | None = None
+    requested: int | None = None
+    retry_after_seconds: float | None = None
+
+    def as_details(self) -> dict[str, object]:
+        """Return a JSON-serialisable dict for embedding in an error envelope."""
+        return {k: v for k, v in asdict(self).items() if v is not None}
+
+
+def _looks_like_rate_limit(status_code: int, message: str) -> bool:
+    if status_code == 429:
+        return True
+    lowered = message.lower()
+    return any(marker in lowered for marker in _RATE_LIMIT_MARKERS)
+
+
+def _parse_retry_after_header(headers: dict[str, str] | None) -> float | None:
+    """Parse a ``Retry-After`` header (delta-seconds form) into seconds."""
+    if not headers:
+        return None
+    raw = headers.get("retry-after") or headers.get("Retry-After")
+    if raw is None:
+        return None
+    try:
+        return float(str(raw).strip())
+    except (TypeError, ValueError):
+        return None
+
+
+def _int_or_none(match: re.Match[str] | None) -> int | None:
+    if match is None:
+        return None
+    try:
+        return int(match.group(1))
+    except (TypeError, ValueError):
+        return None
+
+
+def classify_rate_limit(
+    status_code: int,
+    message: str,
+    headers: dict[str, str] | None = None,
+) -> RateLimitInfo | None:
+    """Classify an upstream error as a rate-limit and extract its fields.
+
+    Args:
+        status_code: HTTP status code from the upstream response.
+        message: Upstream error message (may contain sensitive identifiers).
+        headers: Optional upstream response headers, used for ``Retry-After``.
+
+    Returns:
+        A :class:`RateLimitInfo` when the error is a rate-limit, else ``None``.
+    """
+    message = message or ""
+    if not _looks_like_rate_limit(status_code, message):
+        return None
+
+    redacted = redact_org_ids(message)
+
+    model_match = _MODEL_RE.search(redacted)
+    metric_match = _METRIC_RE.search(redacted)
+
+    retry_after = _parse_retry_after_header(headers)
+    if retry_after is None:
+        retry_match = _RETRY_RE.search(redacted)
+        if retry_match is not None:
+            value = float(retry_match.group(1))
+            retry_after = value / 1000.0 if retry_match.group(2).lower() == "ms" else value
+
+    limit_name_match = _LIMIT_NAME_RE.search(redacted)
+
+    return RateLimitInfo(
+        code=UPSTREAM_RATE_LIMIT,
+        message=redacted,
+        model=model_match.group(1) if model_match else None,
+        limit_name=limit_name_match.group(1).strip() if limit_name_match else None,
+        metric=metric_match.group(1).strip() if metric_match else None,
+        limit=_int_or_none(_LIMIT_RE.search(redacted)),
+        used=_int_or_none(_USED_RE.search(redacted)),
+        requested=_int_or_none(_REQUESTED_RE.search(redacted)),
+        retry_after_seconds=retry_after,
+    )

routstr/upstream/routstr.py

@@ -55,7 +55,7 @@ class RoutstrUpstreamProvider(BaseUpstreamProvider):
         return path.lstrip("/")
 
     @classmethod
-    def from_db_row(
+    def _build_from_row(
         cls, provider_row: "UpstreamProviderRow"
     ) -> "RoutstrUpstreamProvider":
         import json

routstr/upstream/tinfoil.py

@@ -0,0 +1,236 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import httpx
+from fastapi import Request
+from fastapi.responses import Response, StreamingResponse
+from pydantic.v1 import BaseModel
+
+from ..core.exceptions import UpstreamError
+from ..core.logging import get_logger
+from ..payment.models import Architecture, Model, Pricing
+from .base import BaseUpstreamProvider
+from .ehbp import (
+    _ENCLAVE_URL_HEADER,
+    _PROXY_ONLY_HEADERS,
+    _RESPONSE_USAGE_HEADER,
+    ConfidentialInferenceProfile,
+    EHBPForwardingTarget,
+)
+
+if TYPE_CHECKING:
+    from ..core.db import UpstreamProviderRow
+
+logger = get_logger(__name__)
+
+
+class TinfoilModelPricing(BaseModel):
+    inputTokenPricePer1M: float = 0.0
+    outputTokenPricePer1M: float = 0.0
+    requestPrice: float = 0.0
+
+
+class TinfoilModel(BaseModel):
+    id: str
+    context_window: int = 0
+    created: int = 0
+    multimodal: bool = False
+    reasoning: bool = False
+    tool_calling: bool = False
+    type: str = "chat"
+    pricing: TinfoilModelPricing = TinfoilModelPricing()
+    endpoints: list[str] = []
+
+
+class TinfoilUpstreamProvider(BaseUpstreamProvider):
+    """Direct upstream provider for the Tinfoil inference API.
+
+    Tinfoil hosts open-source models inside attested secure enclaves and exposes
+    an OpenAI-compatible API at ``https://inference.tinfoil.sh``. Request and
+    response bodies are encrypted end-to-end with EHBP (HPKE), so Routstr acts
+    as a blind relay: it forwards the opaque encrypted body, never sees
+    plaintext, and bills from the ``X-Tinfoil-Usage-Metrics`` header that
+    Tinfoil returns outside the encrypted body when
+    ``X-Tinfoil-Request-Usage-Metrics: true`` is set.
+    """
+
+    provider_type = "tinfoil"
+    default_base_url = "https://inference.tinfoil.sh"
+    platform_url = "https://docs.tinfoil.sh"
+    supports_ehbp = True
+    confidential_inference_profile = ConfidentialInferenceProfile(
+        usage_response_header=_RESPONSE_USAGE_HEADER,
+        client_target_url_header=_ENCLAVE_URL_HEADER,
+        allow_client_target_override=True,
+        proxy_only_headers=_PROXY_ONLY_HEADERS,
+    )
+
+    def __init__(self, api_key: str, provider_fee: float = 1.0):
+        super().__init__(
+            base_url=self.default_base_url,
+            api_key=api_key,
+            provider_fee=provider_fee,
+        )
+
+    @classmethod
+    def _build_from_row(
+        cls, provider_row: "UpstreamProviderRow"
+    ) -> "TinfoilUpstreamProvider":
+        return cls(
+            api_key=provider_row.api_key,
+            provider_fee=provider_row.provider_fee,
+        )
+
+    @classmethod
+    def get_provider_metadata(cls) -> dict[str, object]:
+        return {
+            "id": cls.provider_type,
+            "name": "Tinfoil",
+            "default_base_url": cls.default_base_url,
+            "fixed_base_url": True,
+            "platform_url": cls.platform_url,
+            "can_create_account": False,
+            "can_topup": False,
+            "can_show_balance": False,
+        }
+
+    def transform_model_name(self, model_id: str) -> str:
+        return model_id.removeprefix("tinfoil/")
+
+    def get_confidential_inference_profile(self) -> ConfidentialInferenceProfile:
+        return self.confidential_inference_profile
+
+    async def forward_get_request(
+        self,
+        request: Request,
+        path: str,
+        headers: dict,
+    ) -> Response | StreamingResponse:
+        """Handle Tinfoil-specific GET endpoints.
+
+        * ``/attestation`` (or ``/tee/attestation``): proxy to the Tinfoil ATC
+          (attestation bundle proxy) at ``https://atc.tinfoil.sh/attestation``.
+        * Other GETs: forward to the provider base URL
+          (``https://inference.tinfoil.sh``). ``X-Tinfoil-Enclave-Url`` is an
+          EHBP-only header used for encrypted POST requests and is not honored
+          for unencrypted GET requests.
+        """
+        clean_path = path.removeprefix("tee/").rstrip("/")
+        if clean_path == "attestation":
+            return await self._proxy_attestation(headers)
+        return await super().forward_get_request(request, path, headers)
+
+    async def _proxy_attestation(self, headers: dict) -> Response:
+        url = "https://atc.tinfoil.sh/attestation"
+        async with httpx.AsyncClient(
+            transport=httpx.AsyncHTTPTransport(retries=1),
+            timeout=30.0,
+        ) as client:
+            try:
+                resp = await client.get(
+                    url,
+                    headers={
+                        "Accept": headers.get("accept", "application/json"),
+                    },
+                )
+                response_headers = dict(resp.headers)
+                response_headers.pop("content-encoding", None)
+                response_headers.pop("content-length", None)
+                return Response(
+                    content=resp.content,
+                    status_code=resp.status_code,
+                    headers=response_headers,
+                )
+            except Exception as exc:
+                raise UpstreamError(
+                    f"Error fetching Tinfoil attestation: {type(exc).__name__}",
+                    status_code=502,
+                ) from exc
+
+    def get_ehbp_forwarding_target(
+        self, path: str, model_obj: Model
+    ) -> EHBPForwardingTarget:
+        """Return the Tinfoil enclave target for EHBP requests.
+
+        Requests usage metrics from the enclave so Routstr can bill exactly
+        without decrypting the response body. The actual forwarding URL is
+        overridden at dispatch time by ``X-Tinfoil-Enclave-Url`` when the SDK
+        sends it (see ``routstr/upstream/ehbp.py``).
+        """
+        return EHBPForwardingTarget(
+            url=f"{self.base_url.rstrip('/')}/{path.lstrip('/')}",
+            headers={"X-Tinfoil-Request-Usage-Metrics": "true"},
+            profile=self.confidential_inference_profile,
+        )
+
+    async def fetch_models(self) -> list[Model]:
+        """Fetch models from the public Tinfoil models endpoint.
+
+        ``GET /v1/models`` is unauthenticated and returns all available models
+        with their pricing in USD per 1M tokens.
+        """
+        url = f"{self.base_url}/v1/models"
+        try:
+            async with httpx.AsyncClient(timeout=30.0) as client:
+                response = await client.get(url)
+                response.raise_for_status()
+                data = response.json()
+                models_data = data.get("data", [])
+
+                models: list[Model] = []
+                for model_data in models_data:
+                    try:
+                        tf = TinfoilModel.parse_obj(model_data)
+                        input_price = tf.pricing.inputTokenPricePer1M
+                        output_price = tf.pricing.outputTokenPricePer1M
+                        request_price = tf.pricing.requestPrice
+
+                        modality = "text->text"
+                        input_modalities = ["text"]
+                        output_modalities = ["text"]
+                        if tf.multimodal:
+                            modality = "text->text+image"
+                            input_modalities = ["text", "image"]
+
+                        models.append(
+                            Model(
+                                id=tf.id,
+                                name=tf.id,
+                                created=tf.created,
+                                description=f"Tinfoil {tf.type} model",
+                                context_length=tf.context_window,
+                                architecture=Architecture(
+                                    modality=modality,
+                                    input_modalities=input_modalities,
+                                    output_modalities=output_modalities,
+                                    tokenizer="Unknown",
+                                    instruct_type=None,
+                                ),
+                                pricing=Pricing(
+                                    prompt=input_price / 1_000_000,
+                                    completion=output_price / 1_000_000,
+                                    request=request_price,
+                                    image=0.0,
+                                    web_search=0.0,
+                                    internal_reasoning=0.0,
+                                ),
+                            )
+                        )
+                    except Exception as e:
+                        logger.warning(
+                            "Failed to parse Tinfoil model",
+                            extra={
+                                "model_id": model_data.get("id", "unknown"),
+                                "error": str(e),
+                                "error_type": type(e).__name__,
+                            },
+                        )
+
+                return models
+        except Exception as e:
+            logger.error(
+                "Error fetching models from Tinfoil",
+                extra={"error": str(e), "error_type": type(e).__name__},
+            )
+            return []

routstr/upstream/tinfoil_trailer.py

@@ -0,0 +1,189 @@
+"""h11-based HTTP client for EHBP requests that captures HTTP trailers.
+
+httpx/httpcore silently discard HTTP trailers during chunked transfer
+decoding. Tinfoil returns ``X-Tinfoil-Usage-Metrics`` as a trailer on
+streaming responses, so we need a lower-level HTTP client that preserves
+trailers from the h11 ``EndOfMessage`` event.
+
+Because EHBP response bodies are opaque encrypted blobs, buffering the full
+response is acceptable — the client decrypts the complete body regardless of
+whether it arrived streamed or buffered.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import ssl
+from dataclasses import dataclass, field
+from urllib.parse import urlsplit
+
+import h11
+
+from ..core import get_logger
+
+logger = get_logger(__name__)
+
+_READ_BUFSIZE = 65536
+_DEFAULT_TIMEOUT_SECONDS = 30.0
+_DEFAULT_CLOSE_TIMEOUT_SECONDS = 1.0
+_DEFAULT_MAX_RESPONSE_BYTES = 25 * 1024 * 1024
+_HOP_BY_HOP_HEADERS = {
+    "connection",
+    "keep-alive",
+    "proxy-authenticate",
+    "proxy-authorization",
+    "te",
+    "trailer",
+    "transfer-encoding",
+    "upgrade",
+}
+
+
+@dataclass
+class TrailerResponse:
+    """Buffered HTTP response with optional trailer headers."""
+
+    status_code: int
+    headers: list[tuple[str, str]]
+    body: bytes
+    trailers: list[tuple[str, str]] = field(default_factory=list)
+
+
+def _get_header(headers: list[tuple[str, str]], name: str) -> str | None:
+    name_lower = name.lower()
+    for k, v in headers:
+        if k.lower() == name_lower:
+            return v
+    return None
+
+
+def _strip_hop_by_hop_headers(headers: dict[str, str]) -> dict[str, str]:
+    """Remove connection-specific headers before serializing a new request."""
+    connection_tokens: set[str] = set()
+    for key, value in headers.items():
+        if key.lower() == "connection":
+            connection_tokens.update(
+                token.strip().lower() for token in value.split(",") if token.strip()
+            )
+
+    excluded = _HOP_BY_HOP_HEADERS | connection_tokens
+    return {key: value for key, value in headers.items() if key.lower() not in excluded}
+
+
+async def forward_with_trailer(
+    *,
+    method: str,
+    url: str,
+    headers: dict[str, str],
+    body: bytes,
+    timeout_seconds: float = _DEFAULT_TIMEOUT_SECONDS,
+    max_response_bytes: int = _DEFAULT_MAX_RESPONSE_BYTES,
+    close_timeout_seconds: float = _DEFAULT_CLOSE_TIMEOUT_SECONDS,
+) -> TrailerResponse:
+    """Send an HTTP/1.1 request via h11 and capture HTTP trailers.
+
+    Returns a :class:`TrailerResponse` with the full buffered body and any
+    trailer headers from the ``EndOfMessage`` event.
+    """
+    parsed = urlsplit(url)
+    host = parsed.hostname
+    if not host:
+        raise ValueError(f"Invalid URL (no hostname): {url}")
+    port = parsed.port or 443
+    path = parsed.path or "/"
+    if parsed.query:
+        path = f"{path}?{parsed.query}"
+
+    # FastAPI has already decoded the incoming request body. Do not carry the
+    # original connection's framing or other hop-by-hop metadata into the new
+    # upstream connection.
+    headers = _strip_hop_by_hop_headers(headers)
+
+    ssl_ctx = ssl.create_default_context()
+    reader, writer = await asyncio.wait_for(
+        asyncio.open_connection(host, port, ssl=ssl_ctx),
+        timeout=timeout_seconds,
+    )
+
+    try:
+        # Build HTTP/1.1 request
+        header_lines = [f"{method} {path} HTTP/1.1"]
+        has_host = any(k.lower() == "host" for k in headers)
+        if not has_host:
+            header_lines.append(f"Host: {host}")
+        header_lines.append("Connection: close")
+
+        for key, value in headers.items():
+            if key.lower() == "host":
+                continue
+            header_lines.append(f"{key}: {value}")
+
+        if body and not any(k.lower() == "content-length" for k in headers):
+            header_lines.append(f"Content-Length: {len(body)}")
+
+        request_data = "\r\n".join(header_lines).encode() + b"\r\n\r\n"
+        if body:
+            request_data += body
+
+        writer.write(request_data)
+        await asyncio.wait_for(writer.drain(), timeout=timeout_seconds)
+
+        # Parse response with h11
+        conn = h11.Connection(h11.CLIENT)
+        status_code = 0
+        resp_headers: list[tuple[str, str]] = []
+        body_chunks: list[bytes] = []
+        body_size = 0
+        trailers: list[tuple[str, str]] = []
+
+        while True:
+            event = conn.next_event()
+
+            if event is h11.NEED_DATA:
+                data = await asyncio.wait_for(
+                    reader.read(_READ_BUFSIZE),
+                    timeout=timeout_seconds,
+                )
+                conn.receive_data(data if data else b"")
+                continue
+
+            if isinstance(event, h11.Response):
+                status_code = event.status_code
+                resp_headers = [(k.decode(), v.decode()) for k, v in event.headers]
+
+            elif isinstance(event, h11.Data):
+                body_size += len(event.data)
+                if body_size > max_response_bytes:
+                    raise ValueError(
+                        f"EHBP response exceeded {max_response_bytes} bytes"
+                    )
+                body_chunks.append(event.data)
+
+            elif isinstance(event, h11.EndOfMessage):
+                for k, v in event.headers:
+                    trailers.append((k.decode(), v.decode()))
+                break
+
+            elif isinstance(event, h11.PAUSED):
+                # Shouldn't happen for simple request/response, but break safely
+                logger.warning("h11 PAUSED event during EHBP response parsing")
+                break
+
+            elif isinstance(event, h11.ConnectionClosed):
+                break
+
+        return TrailerResponse(
+            status_code=status_code,
+            headers=resp_headers,
+            body=b"".join(body_chunks),
+            trailers=trailers,
+        )
+    finally:
+        writer.close()
+        if close_timeout_seconds > 0:
+            try:
+                await asyncio.wait_for(
+                    writer.wait_closed(), timeout=close_timeout_seconds
+                )
+            except Exception:
+                pass

routstr/upstream/xai.py

@@ -21,7 +21,7 @@ class XAIUpstreamProvider(BaseUpstreamProvider):
         )
 
     @classmethod
-    def from_db_row(cls, provider_row: "UpstreamProviderRow") -> "XAIUpstreamProvider":
+    def _build_from_row(cls, provider_row: "UpstreamProviderRow") -> "XAIUpstreamProvider":
         return cls(
             api_key=provider_row.api_key,
             provider_fee=provider_row.provider_fee,

routstr/wallet.py

@@ -1,8 +1,11 @@
 import asyncio
+import re
+import socket
 import time
 import typing
 from typing import TypedDict
 
+import httpx
 from cashu.core.base import Proof, Token
 from cashu.core.mint_info import MintInfo as _CashuMintInfo
 from cashu.wallet.helpers import deserialize_token_from_string
@@ -30,6 +33,149 @@ _CashuMintInfo.model_rebuild(force=True)
 logger = get_logger(__name__)
 
 
+class MintConnectionError(Exception):
+    """The mint could not be reached (network transport failure).
+
+    Maps to a 503, not a 4xx: the token is fine, the mint is just unavailable.
+    """
+
+
+class TokenConsumedError(Exception):
+    """A failure that happened AFTER the token's proofs were spent (melt
+    succeeded, or redemption already returned) — e.g. minting on the primary
+    mint or the DB credit then failed.
+
+    Non-retryable: the same token will not work again. Seals the cause chain so
+    a transport error underneath is never re-surfaced as a retryable
+    mint_unreachable.
+    """
+
+
+# httpx base classes cover their subclasses. HTTPStatusError is excluded on
+# purpose — that means the mint answered, just with an error status.
+_TRANSPORT_EXC_TYPES: tuple[type[BaseException], ...] = (
+    httpx.NetworkError,
+    httpx.TimeoutException,
+    ConnectionError,  # refused/reset/aborted
+    socket.gaierror,  # DNS failure
+    asyncio.TimeoutError,
+)
+
+
+def is_mint_connection_error(error: BaseException) -> bool:
+    """True if ``error`` (or anything in its cause/context chain) is a mint
+    transport failure. Walks the chain because some sites re-raise transport
+    errors wrapped in ValueError/MintConnectionError; matches on TYPE, not text.
+    """
+    seen: set[int] = set()
+    current: BaseException | None = error
+    while current is not None and id(current) not in seen:
+        seen.add(id(current))
+        if isinstance(current, TokenConsumedError):
+            # Sealed: the token was already spent, so whatever transport error
+            # sits underneath must not make this look retryable.
+            return False
+        if isinstance(current, MintConnectionError):
+            return True
+        if isinstance(current, _TRANSPORT_EXC_TYPES):
+            return True
+        current = current.__cause__ or current.__context__
+    return False
+
+
+# Redemption ``code`` values whose token is spent/consumed/unusable — the
+# X-Cashu path must NOT echo the original token for these (echoing invites a
+# retry with a token that can never succeed again).
+SPENT_TOKEN_CODES: frozenset[str] = frozenset(
+    {
+        "cashu_token_already_spent",
+        "cashu_token_consumed",
+        "cashu_token_zero_value",
+        "internal_error",
+    }
+)
+
+
+def classify_redemption_error(
+    error: Exception,
+) -> tuple[str, int, str, str] | None:
+    """Map a token-redemption failure to ``(type, status, message, code)``.
+
+    Single source of truth for every endpoint that redeems a token (bearer,
+    X-Cashu, top-up) so the same failure yields the same taxonomy everywhere.
+    ``type`` and ``code`` are stable client contract; ``message`` is sanitized
+    (raw error text stays in logs). Returns None for an unclassified internal
+    fault — the caller emits a generic 500.
+    """
+    if isinstance(error, TokenConsumedError):
+        return (
+            "token_consumed",
+            500,
+            "Token was redeemed but could not be credited; do not retry",
+            "cashu_token_consumed",
+        )
+    if is_mint_connection_error(error):
+        return (
+            "mint_unreachable",
+            503,
+            "Cashu mint is unreachable",
+            "cashu_mint_unreachable",
+        )
+    lowered = str(error).lower()
+    if "already spent" in lowered:
+        return (
+            "token_already_spent",
+            400,
+            "Cashu token already spent",
+            "cashu_token_already_spent",
+        )
+    if (
+        "insufficient" in lowered
+        or "melt fee" in lowered
+        or "exceed token amount" in lowered
+        or "estimate fees" in lowered
+    ):
+        return (
+            "mint_error",
+            422,
+            "Token value is too small to cover swap fees",
+            "cashu_token_swap_fees_exceed_amount",
+        )
+    if "failed to melt" in lowered:
+        return (
+            "mint_error",
+            422,
+            "Failed to swap token from foreign mint",
+            "cashu_foreign_mint_swap_failed",
+        )
+    if ("invalid" in lowered or "decode" in lowered) and "token" in lowered:
+        # Anchored to "token" so internal faults whose text merely contains
+        # "invalid"/"decode" fall through to the 500 branch, not a token error.
+        return (
+            "invalid_token",
+            400,
+            "Invalid Cashu token",
+            "invalid_cashu_token",
+        )
+    if "must be positive" in lowered or "yielded no value" in lowered:
+        # Redeemed to <= 0 (empty/dust token, or value fully consumed by fees).
+        # Consumed, so non-retryable, but its own code — not the generic bucket.
+        return (
+            "cashu_error",
+            400,
+            "Failed to redeem Cashu token: token yielded no value",
+            "cashu_token_zero_value",
+        )
+    if isinstance(error, ValueError):
+        return (
+            "cashu_error",
+            400,
+            "Failed to redeem Cashu token",
+            "cashu_token_redemption_failed",
+        )
+    return None
+
+
 async def get_balance(unit: str) -> int:
     wallet = await get_wallet(settings.primary_mint, unit)
     return wallet.available_balance.amount
@@ -129,6 +275,72 @@ async def send_token(amount: int, unit: str, mint_url: str | None = None) -> str
     return token
 
 
+# A foreign mint's fee_reserve is a non-binding estimate (NUT-05): the mint may
+# demand more when re-quoting or at melt execution. Instead of padding the
+# estimate with a safety buffer (which strands the margin at the foreign mint
+# on every swap), the swap retries with the amount recomputed from the fees the
+# mint actually demands, up to this many attempts.
+_MAX_SWAP_ATTEMPTS = 3
+
+_MINT_ERROR_CODE_RE = re.compile(r"\(Code: (\d+)\)")
+_MELT_SHORTFALL_RE = re.compile(r"Provided: (\d+), needed: (\d+)")
+
+# Insufficient-melt-inputs failures differ across mint implementations. 11005 is
+# the registered "Transaction is not balanced" code (cdk), specific enough to
+# trust on the code alone. 11000 is nutshell's generic, unregistered
+# TransactionError covering many unrelated failures, so it only counts as a fee
+# shortfall alongside the "not enough inputs" detail text. With no code suffix at
+# all, that same text is the only signal.
+
+
+def _net_minted_amount(amount_msat: int, token_unit: str, fees: int) -> int:
+    """
+    Convert the token value minus fees (given in the token unit) into an
+    amount in the primary mint's unit.
+    """
+    fee_msat = fees * 1000 if token_unit == "sat" else fees
+    remaining_msat = amount_msat - fee_msat
+    if settings.primary_mint_unit == "sat":
+        return int(remaining_msat // 1000)
+    return int(remaining_msat)
+
+
+def _melt_insufficient_shortfall(error: Exception) -> int | None:
+    """
+    Classify a melt failure: return the observed shortfall (in the token unit)
+    when the mint rejected the inputs as insufficient, or None when the failure
+    is unrelated to fees and must not be retried (e.g. a Lightning payment
+    failure, where a smaller invoice would not help).
+
+    Cashu errors carry no structured amounts (NUT-00 defines only detail/code,
+    flattened to "Mint Error: <detail> (Code: <code>)" by cashu-py), so the
+    classification uses the code and the shortfall must be inferred: the
+    "Provided: X, needed: Y" amounts are nutshell-specific free text and only
+    refine the shortfall when present; otherwise shrink one unit at a time.
+    """
+    message = str(error)
+    code_match = _MINT_ERROR_CODE_RE.search(message)
+    code = code_match.group(1) if code_match is not None else None
+    has_shortfall_text = "not enough inputs" in message.lower()
+
+    match code:
+        case "11005":  # registered TransactionUnbalanced: trust the code
+            pass
+        case "11000" if has_shortfall_text:  # generic nutshell error: needs the text
+            pass
+        case None if has_shortfall_text:  # no code suffix: text is the only signal
+            pass
+        case _:  # other codes, a bare 11000, or no signal: must not retry
+            return None
+
+    amounts = _MELT_SHORTFALL_RE.search(message)
+    if amounts is not None:
+        provided, needed = int(amounts.group(1)), int(amounts.group(2))
+        if needed > provided:
+            return needed - provided
+    return 1
+
+
 async def _calculate_swap_amount(
     amount_msat: int,
     token_unit: str,
@@ -167,26 +379,18 @@ async def _calculate_swap_amount(
 
         fee_reserve = dummy_melt_quote.fee_reserve
         input_fees = token_wallet.get_fees_for_proofs(proofs)
-        if token_unit == "sat":
-            fee_msat = (fee_reserve + input_fees) * 1000
-        else:
-            fee_msat = fee_reserve + input_fees
-
-        amount_msat_after_fee = amount_msat - fee_msat
-
-        if settings.primary_mint_unit == "sat":
-            minted_amount = int(amount_msat_after_fee // 1000)
-        else:
-            minted_amount = int(amount_msat_after_fee)
+        total_fees = fee_reserve + input_fees
+        minted_amount = _net_minted_amount(amount_msat, token_unit, total_fees)
 
         if minted_amount <= 0:
-            raise ValueError(f"Fees ({fee_reserve + input_fees} {token_unit}) exceed token amount")
+            raise ValueError(f"Fees ({total_fees} {token_unit}) exceed token amount")
 
         logger.info(
             "swap_to_primary_mint: fee estimation result",
             extra={
                 "token_amount_sat": amount_msat // 1000,
-                "estimated_fee_sat": fee_msat // 1000,
+                "estimated_fee": total_fees,
+                "estimated_fee_unit": token_unit,
                 "input_fees": input_fees,
                 "minted_amount": minted_amount,
                 "minted_unit": settings.primary_mint_unit,
@@ -199,6 +403,8 @@ async def _calculate_swap_amount(
             "swap_to_primary_mint: fee estimation failed",
             extra={"error": str(e)},
         )
+        if is_mint_connection_error(e):
+            raise MintConnectionError("Cashu mint is unreachable") from e
         raise ValueError(f"Failed to estimate fees: {e}") from e
 
 
@@ -251,67 +457,123 @@ async def swap_to_primary_mint(
         token_obj.proofs,
     )
 
-    mint_quote = await primary_wallet.request_mint(minted_amount)
-    logger.info(
-        "swap_to_primary_mint: mint quote received",
-        extra={"mint_quote_id": mint_quote.quote},
-    )
+    # The estimate above is non-binding: the mint may demand a higher fee on the
+    # real quote or reject the melt outright. Retry the quote/melt cycle with the
+    # amount recomputed from the fees the mint actually demands.
+    observed_extra_fee = 0
+    attempt = 0
+    while True:
+        attempt += 1
+        mint_quote = await primary_wallet.request_mint(minted_amount)
+        logger.info(
+            "swap_to_primary_mint: mint quote received",
+            extra={"mint_quote_id": mint_quote.quote, "attempt": attempt},
+        )
 
-    melt_quote = await token_wallet.melt_quote(mint_quote.request)
-    input_fees = token_wallet.get_fees_for_proofs(token_obj.proofs)
-    total_needed = melt_quote.amount + melt_quote.fee_reserve + input_fees
-    logger.info(
-        "swap_to_primary_mint: melt quote received",
-        extra={
-            "melt_quote_id": melt_quote.quote,
-            "melt_amount": melt_quote.amount,
-            "melt_fee_reserve": melt_quote.fee_reserve,
-            "input_fees": input_fees,
-            "total_needed": total_needed,
-            "token_amount": token_amount,
-        },
-    )
-
-    if total_needed > token_amount:
-        logger.warning(
-            "swap_to_primary_mint: insufficient token amount for melt fees",
+        melt_quote = await token_wallet.melt_quote(mint_quote.request)
+        input_fees = token_wallet.get_fees_for_proofs(token_obj.proofs)
+        total_needed = melt_quote.amount + melt_quote.fee_reserve + input_fees
+        logger.info(
+            "swap_to_primary_mint: melt quote received",
             extra={
-                "token_amount": token_amount,
+                "melt_quote_id": melt_quote.quote,
                 "melt_amount": melt_quote.amount,
                 "melt_fee_reserve": melt_quote.fee_reserve,
                 "input_fees": input_fees,
                 "total_needed": total_needed,
-                "shortfall": total_needed - token_amount,
+                "token_amount": token_amount,
+                "attempt": attempt,
             },
         )
-        raise ValueError(
-            f"Token amount ({token_amount} {token_obj.unit}) is insufficient to cover "
-            f"melt fees. Needed: {total_needed} {token_obj.unit} "
-            f"(amount: {melt_quote.amount} + fee: {melt_quote.fee_reserve} + input_fees: {input_fees})"
-        )
 
-    try:
-        _ = await token_wallet.melt(
-            proofs=token_obj.proofs,
-            invoice=mint_quote.request,
-            fee_reserve_sat=melt_quote.fee_reserve,
-            quote_id=melt_quote.quote,
-        )
-    except Exception as e:
-        logger.error(
-            "swap_to_primary_mint: melt failed",
-            extra={
-                "error": str(e),
-                "error_type": type(e).__name__,
-                "foreign_mint": token_obj.mint,
-                "token_amount": token_amount,
-                "melt_quote_id": melt_quote.quote,
-                "total_needed": total_needed,
-            },
-        )
-        raise ValueError(
-            f"Failed to melt token from foreign mint {token_obj.mint}: {e}"
-        ) from e
+        if total_needed > token_amount:
+            recomputed = _net_minted_amount(
+                amount_msat,
+                token_obj.unit,
+                melt_quote.fee_reserve + input_fees + observed_extra_fee,
+            )
+            if attempt >= _MAX_SWAP_ATTEMPTS or recomputed <= 0:
+                logger.warning(
+                    "swap_to_primary_mint: insufficient token amount for melt fees",
+                    extra={
+                        "token_amount": token_amount,
+                        "melt_amount": melt_quote.amount,
+                        "melt_fee_reserve": melt_quote.fee_reserve,
+                        "input_fees": input_fees,
+                        "total_needed": total_needed,
+                        "shortfall": total_needed - token_amount,
+                        "attempts": attempt,
+                    },
+                )
+                raise ValueError(
+                    f"Token amount ({token_amount} {token_obj.unit}) is insufficient to cover "
+                    f"melt fees. Needed: {total_needed} {token_obj.unit} "
+                    f"(amount: {melt_quote.amount} + fee: {melt_quote.fee_reserve} + input_fees: {input_fees})"
+                )
+            logger.warning(
+                "swap_to_primary_mint: melt quote exceeds token amount, retrying",
+                extra={
+                    "total_needed": total_needed,
+                    "token_amount": token_amount,
+                    "retry_minted_amount": recomputed,
+                    "attempt": attempt,
+                },
+            )
+            minted_amount = recomputed
+            continue
+
+        try:
+            _ = await token_wallet.melt(
+                proofs=token_obj.proofs,
+                invoice=mint_quote.request,
+                fee_reserve_sat=melt_quote.fee_reserve,
+                quote_id=melt_quote.quote,
+            )
+        except Exception as e:
+            # A down mint won't fix itself by retrying with a smaller amount.
+            if is_mint_connection_error(e):
+                logger.error(
+                    "swap_to_primary_mint: melt failed — mint unreachable",
+                    extra={"error": str(e), "foreign_mint": token_obj.mint},
+                )
+                raise MintConnectionError("Cashu mint is unreachable") from e
+            shortfall = _melt_insufficient_shortfall(e)
+            recomputed = 0
+            if shortfall is not None:
+                observed_extra_fee += shortfall
+                recomputed = _net_minted_amount(
+                    amount_msat,
+                    token_obj.unit,
+                    melt_quote.fee_reserve + input_fees + observed_extra_fee,
+                )
+            if shortfall is None or attempt >= _MAX_SWAP_ATTEMPTS or recomputed <= 0:
+                logger.error(
+                    "swap_to_primary_mint: melt failed",
+                    extra={
+                        "error": str(e),
+                        "error_type": type(e).__name__,
+                        "foreign_mint": token_obj.mint,
+                        "token_amount": token_amount,
+                        "melt_quote_id": melt_quote.quote,
+                        "total_needed": total_needed,
+                        "attempts": attempt,
+                    },
+                )
+                raise ValueError(
+                    f"Failed to melt token from foreign mint {token_obj.mint}: {e}"
+                ) from e
+            logger.warning(
+                "swap_to_primary_mint: mint demanded more than quoted at melt, retrying",
+                extra={
+                    "shortfall": shortfall,
+                    "retry_minted_amount": recomputed,
+                    "attempt": attempt,
+                },
+            )
+            minted_amount = recomputed
+            continue
+
+        break
 
     logger.info(
         "swap_to_primary_mint: melt succeeded, minting on primary",
@@ -350,21 +612,21 @@ async def swap_to_primary_mint(
                     # Recovery scan ran but did NOT restore the orphaned proofs
                     # (mint reports them as spent — they're stuck). Refuse to
                     # credit the API key balance for proofs we don't actually hold.
-                    raise ValueError(
+                    raise TokenConsumedError(
                         f"Swap recovery failed: mint signed outputs but proofs are "
                         f"unrecoverable (mint reports them spent). "
                         f"Expected {minted_amount}, recovered {balance_gained}. "
                         f"Local wallet DB ('.wallet/') state is corrupted — "
                         f"the counter for keyset is stuck at a bad index range."
                     )
-            except ValueError:
+            except TokenConsumedError:
                 raise
             except Exception as recovery_err:
                 logger.error(
                     "swap_to_primary_mint: recovery failed",
                     extra={"error": str(recovery_err)},
                 )
-                raise ValueError(
+                raise TokenConsumedError(
                     f"Mint on primary failed and recovery unsuccessful: {e}"
                 ) from e
         else:
@@ -377,7 +639,10 @@ async def swap_to_primary_mint(
                     "mint_quote_id": mint_quote.quote,
                 },
             )
-            raise
+            # Foreign proofs already melted (spent) — non-retryable.
+            raise TokenConsumedError(
+                "Mint on primary failed after successful melt"
+            ) from e
 
     logger.info(
         "swap_to_primary_mint: completed successfully",
@@ -435,15 +700,33 @@ async def credit_balance(
             extra={"old_balance": key.balance, "credit_amount": amount},
         )
 
-        # Use atomic SQL UPDATE to prevent race conditions during concurrent topups
-        stmt = (
-            update(db.ApiKey)
-            .where(col(db.ApiKey.hashed_key) == key.hashed_key)
-            .values(balance=(db.ApiKey.balance) + amount)
-        )
-        await session.exec(stmt)  # type: ignore[call-overload]
-        await session.commit()
-        await session.refresh(key)
+        # The token is already redeemed (spent) here, so any crediting failure
+        # is post-redemption and non-retryable — surface it as TokenConsumedError
+        # (a key that vanished mid-flight, or an unexpected DB fault), never a
+        # retryable/token-error taxonomy.
+        try:
+            # Atomic UPDATE to prevent race conditions during concurrent topups.
+            stmt = (
+                update(db.ApiKey)
+                .where(col(db.ApiKey.hashed_key) == key.hashed_key)
+                .values(balance=(db.ApiKey.balance) + amount)
+            )
+            result = await session.exec(stmt)  # type: ignore[call-overload]
+            # If pruning removed this key after redemption, do not commit a no-op
+            # balance update and pretend the top-up succeeded.
+            if (getattr(result, "rowcount", 0) or 0) == 0:
+                raise TokenConsumedError(
+                    "Token redeemed but the API key disappeared before the "
+                    "credit could be recorded"
+                )
+            await session.commit()
+            await session.refresh(key)
+        except TokenConsumedError:
+            raise
+        except Exception as db_error:
+            raise TokenConsumedError(
+                "Token redeemed but crediting the balance failed"
+            ) from db_error
 
         logger.info(
             "credit_balance: Balance updated successfully",
@@ -580,11 +863,19 @@ async def fetch_all_balances(
             }
             return error_result
 
+    # Build the set of mints to inspect. Received tokens are stored against
+    # ``primary_mint`` (which defaults to a real mint even when ``cashu_mints``
+    # is empty), so include it as a fallback — otherwise a node that accepts
+    # payments would still report empty balances when ``cashu_mints`` is unset.
+    mint_urls: list[str] = list(settings.cashu_mints)
+    if settings.primary_mint and settings.primary_mint not in mint_urls:
+        mint_urls.append(settings.primary_mint)
+
     # Create tasks for all mint/unit combinations
     async with db.create_session() as session:
         tasks = [
             fetch_balance(session, mint_url, unit)
-            for mint_url in settings.cashu_mints
+            for mint_url in mint_urls
             for unit in units
         ]
 
@@ -628,104 +919,132 @@ async def fetch_all_balances(
 async def periodic_payout() -> None:
     while True:
         await asyncio.sleep(settings.payout_interval_seconds)
-        print(settings.payout_interval_seconds)
-        if not settings.receive_ln_address:
-            continue
         try:
+            if not settings.receive_ln_address:
+                continue
+
+            # Include the primary mint even if it is not listed in cashu_mints,
+            # matching fetch_all_balances(); otherwise primary-mint funds never
+            # auto-payout.
+            mint_urls: list[str] = list(settings.cashu_mints)
+            if settings.primary_mint and settings.primary_mint not in mint_urls:
+                mint_urls.append(settings.primary_mint)
+
             async with db.create_session() as session:
-                for mint_url in settings.cashu_mints:
+                for mint_url in mint_urls:
                     for unit in ["sat", "msat"]:
-                        wallet = await get_wallet(mint_url, unit)
-                        proofs = get_proofs_per_mint_and_unit(
-                            wallet, mint_url, unit, not_reserved=True
-                        )
-                        proofs = await slow_filter_spend_proofs(proofs, wallet)
-                        await asyncio.sleep(5)
-                        user_balance = await db.balances_for_mint_and_unit(
-                            session, mint_url, unit
-                        )
-                        if unit == "sat":
-                            user_balance = user_balance // 1000
-                        proofs_balance = sum(proof.amount for proof in proofs)
-                        available_balance = proofs_balance - user_balance
-                        # Threshold is configured in sats; convert for msat wallets.
-                        min_amount = (
-                            settings.min_payout_sat
-                            if unit == "sat"
-                            else settings.min_payout_sat * 1000
-                        )
-                        if available_balance > min_amount:
-                            amount_received = await raw_send_to_lnurl(
-                                wallet,
-                                proofs,
-                                settings.receive_ln_address,
-                                unit,
-                                amount=available_balance,
+                        # Isolate failures per mint/unit so one slow or failing
+                        # mint does not abort payout for every other mint/unit.
+                        try:
+                            wallet = await get_wallet(mint_url, unit)
+                            proofs = get_proofs_per_mint_and_unit(
+                                wallet, mint_url, unit, not_reserved=True
                             )
-                            logger.info(
-                                "Payout sent successfully",
+                            proofs = await slow_filter_spend_proofs(proofs, wallet)
+                            await asyncio.sleep(5)
+                            user_balance = await db.balances_for_mint_and_unit(
+                                session, mint_url, unit
+                            )
+                            if unit == "sat":
+                                user_balance = user_balance // 1000
+                            proofs_balance = sum(proof.amount for proof in proofs)
+                            available_balance = proofs_balance - user_balance
+                            # Threshold is configured in sats; convert for msat wallets.
+                            min_amount = (
+                                settings.min_payout_sat
+                                if unit == "sat"
+                                else settings.min_payout_sat * 1000
+                            )
+                            if available_balance > min_amount:
+                                amount_received = await raw_send_to_lnurl(
+                                    wallet,
+                                    proofs,
+                                    settings.receive_ln_address,
+                                    unit,
+                                    amount=available_balance,
+                                )
+                                logger.info(
+                                    "Payout sent successfully",
+                                    extra={
+                                        "mint_url": mint_url,
+                                        "unit": unit,
+                                        "balance": available_balance,
+                                        "amount_received": amount_received,
+                                    },
+                                )
+                        except Exception as e:
+                            logger.error(
+                                f"Error sending payout: {type(e).__name__}",
                                 extra={
+                                    "error": str(e),
                                     "mint_url": mint_url,
                                     "unit": unit,
-                                    "balance": available_balance,
-                                    "amount_received": amount_received,
                                 },
                             )
         except Exception as e:
             logger.error(
-                f"Error sending payout: {type(e).__name__}",
+                f"Error in periodic payout cycle: {type(e).__name__}",
                 extra={"error": str(e)},
             )
 
 
+async def _refund_sweep_once(cutoff: int) -> None:
+    async with db.create_session() as session:
+        stmt = select(db.CashuTransaction).where(
+            db.CashuTransaction.type == "out",
+            db.CashuTransaction.collected == False,  # noqa: E712
+            db.CashuTransaction.swept == False,  # noqa: E712
+            db.CashuTransaction.created_at < cutoff,
+        )
+        results = await session.exec(stmt)
+        refunds = results.all()
+
+        for refund in refunds:
+            try:
+                await recieve_token(refund.token)
+                refund.swept = True
+                session.add(refund)
+                logger.info(
+                    "Swept uncollected refund",
+                    extra={
+                        "id": refund.id,
+                        "amount": refund.amount,
+                        "unit": refund.unit,
+                    },
+                )
+            except Exception as e:
+                error_msg = str(e).lower()
+                if "already spent" in error_msg:
+                    refund.collected = True
+                    session.add(refund)
+                    logger.info(
+                        "Refund already spent (client collected), marking swept",
+                        extra={
+                            "id": refund.id,
+                        },
+                    )
+                else:
+                    logger.warning(
+                        "Failed to sweep refund",
+                        extra={
+                            "id": refund.id,
+                            "error": str(e),
+                        },
+                    )
+        await session.commit()
+
+
+async def refund_sweep_once() -> None:
+    """Sweep eligible uncollected refund tokens once."""
+    cutoff = int(time.time()) - settings.refund_sweep_ttl_seconds
+    await _refund_sweep_once(cutoff)
+
+
 async def periodic_refund_sweep() -> None:
     while True:
         await asyncio.sleep(60 * 60)  # every hour
         try:
-            cutoff = int(time.time()) - settings.refund_sweep_ttl_seconds
-            async with db.create_session() as session:
-                stmt = select(db.CashuTransaction).where(
-                    db.CashuTransaction.type == "out",
-                    db.CashuTransaction.collected == False,  # noqa: E712
-                    db.CashuTransaction.swept == False,  # noqa: E712
-                    db.CashuTransaction.created_at < cutoff,
-                )
-                results = await session.exec(stmt)
-                refunds = results.all()
-
-                for refund in refunds:
-                    try:
-                        await recieve_token(refund.token)
-                        refund.swept = True
-                        session.add(refund)
-                        logger.info(
-                            "Swept uncollected refund",
-                            extra={
-                                "id": refund.id,
-                                "amount": refund.amount,
-                                "unit": refund.unit,
-                            },
-                        )
-                    except Exception as e:
-                        error_msg = str(e).lower()
-                        if "already spent" in error_msg:
-                            refund.collected = True
-                            session.add(refund)
-                            logger.info(
-                                "Refund already spent (client collected), marking swept",
-                                extra={
-                                    "id": refund.id,
-                                },
-                            )
-                        else:
-                            logger.warning(
-                                "Failed to sweep refund",
-                                extra={
-                                    "id": refund.id,
-                                    "error": str(e),
-                                },
-                            )
-                await session.commit()
+            await refund_sweep_once()
         except Exception as e:
             logger.error(
                 "Error in periodic refund sweep",
@@ -748,17 +1067,56 @@ async def periodic_routstr_fee_payout() -> None:
         try:
             async with db.create_session() as session:
                 fee = await db.get_routstr_fee(session)
+                if fee.payout_in_progress_msats:
+                    logger.critical(
+                        "Routstr fee payout requires manual reconciliation",
+                        extra={
+                            "payout_in_progress_msats": fee.payout_in_progress_msats,
+                            "payout_started_at": fee.payout_started_at,
+                        },
+                    )
+                    continue
+
                 accumulated_sats = fee.accumulated_msats // 1000
                 if accumulated_sats >= ROUTSTR_FEE_DEFAULT_PAYOUT:
                     wallet = await get_wallet(settings.primary_mint, "sat")
                     proofs = get_proofs_per_mint_and_unit(
                         wallet, settings.primary_mint, "sat", not_reserved=True
                     )
-                    amount_received = await raw_send_to_lnurl(
-                        wallet, proofs, ROUTSTR_LN_ADDRESS, "sat", amount=accumulated_sats
-                    )
                     paid_msats = accumulated_sats * 1000
-                    await db.reset_routstr_fee(session, paid_msats)
+                    payout_checkpointed = await db.reset_routstr_fee(
+                        session, paid_msats
+                    )
+                    if not payout_checkpointed:
+                        logger.warning("Routstr fee payout was already claimed")
+                        continue
+
+                    try:
+                        amount_received = await raw_send_to_lnurl(
+                            wallet,
+                            proofs,
+                            ROUTSTR_LN_ADDRESS,
+                            "sat",
+                            amount=accumulated_sats,
+                        )
+                    except Exception:
+                        logger.critical(
+                            "Routstr fee payout outcome is unknown; manual reconciliation required",
+                            extra={"payout_in_progress_msats": paid_msats},
+                            exc_info=True,
+                        )
+                        continue
+
+                    payout_completed = await db.complete_routstr_fee_payout(
+                        session, paid_msats
+                    )
+                    if not payout_completed:
+                        logger.critical(
+                            "Routstr fee payout sent but checkpoint was not completed",
+                            extra={"payout_in_progress_msats": paid_msats},
+                        )
+                        continue
+
                     logger.info(
                         "Routstr fee payout sent",
                         extra={

tests/integration/test_admin_model_cache_pricing.py

@@ -0,0 +1,229 @@
+import json
+from datetime import datetime, timedelta, timezone
+from unittest.mock import patch
+
+import pytest
+from httpx import AsyncClient
+from sqlmodel.ext.asyncio.session import AsyncSession
+
+from routstr.core.admin import admin_sessions
+from routstr.core.db import ModelRow, UpstreamProviderRow
+from routstr.payment.cost_calculation import CostData, calculate_cost
+from routstr.proxy import get_model_instance, reinitialize_upstreams
+
+
+def _admin_headers() -> dict[str, str]:
+    token = "test-admin-cache-pricing-token"
+    admin_sessions[token] = int(
+        (datetime.now(timezone.utc) + timedelta(minutes=5)).timestamp()
+    )
+    return {"Authorization": f"Bearer {token}"}
+
+
+def _model_payload(
+    provider_id: int,
+    *,
+    cache_read: float,
+    cache_write: float,
+) -> dict[str, object]:
+    return {
+        "id": "custom-cache-model",
+        "name": "Custom Cache Model",
+        "description": "custom model with explicit cache pricing",
+        "created": 0,
+        "context_length": 128000,
+        "architecture": {
+            "modality": "text",
+            "input_modalities": ["text"],
+            "output_modalities": ["text"],
+            "tokenizer": "unknown",
+            "instruct_type": None,
+        },
+        "pricing": {
+            "prompt": 1.4e-7,
+            "completion": 2.8e-7,
+            "input_cache_read": cache_read,
+            "input_cache_write": cache_write,
+            "request": 0.0,
+            "image": 0.0,
+            "web_search": 0.0,
+            "internal_reasoning": 0.0,
+        },
+        "per_request_limits": None,
+        "top_provider": None,
+        "upstream_provider_id": provider_id,
+        "canonical_slug": None,
+        "alias_ids": [],
+        "enabled": True,
+        "forwarded_model_id": "custom-cache-model",
+    }
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_admin_provider_model_api_persists_cache_pricing_on_create_and_update(
+    integration_client: AsyncClient,
+    integration_session: AsyncSession,
+) -> None:
+    provider = UpstreamProviderRow(
+        provider_type="generic",
+        base_url="https://custom-upstream.example/v1",
+        api_key="test-key",
+        provider_fee=1.0,
+    )
+    integration_session.add(provider)
+    await integration_session.commit()
+    await integration_session.refresh(provider)
+    assert provider.id is not None
+    await reinitialize_upstreams()
+
+    headers = _admin_headers()
+    create_payload = _model_payload(
+        provider.id,
+        cache_read=2.8e-9,
+        cache_write=3.5e-9,
+    )
+
+    with patch("routstr.payment.models.sats_usd_price", return_value=1e-6):
+        create_response = await integration_client.post(
+            f"/admin/api/upstream-providers/{provider.id}/models",
+            headers=headers,
+            json=create_payload,
+        )
+
+    assert create_response.status_code == 200
+    create_body = create_response.json()
+    assert create_body["pricing"]["input_cache_read"] == pytest.approx(2.8e-9)
+    assert create_body["pricing"]["input_cache_write"] == pytest.approx(3.5e-9)
+
+    row = await integration_session.get(ModelRow, ("custom-cache-model", provider.id))
+    assert row is not None
+    stored_pricing = json.loads(row.pricing)
+    assert stored_pricing["input_cache_read"] == pytest.approx(2.8e-9)
+    assert stored_pricing["input_cache_write"] == pytest.approx(3.5e-9)
+
+    update_payload = _model_payload(
+        provider.id,
+        cache_read=1.25e-9,
+        cache_write=4.5e-9,
+    )
+
+    with patch("routstr.payment.models.sats_usd_price", return_value=1e-6):
+        update_response = await integration_client.post(
+            f"/admin/api/upstream-providers/{provider.id}/models",
+            headers=headers,
+            json=update_payload,
+        )
+
+    assert update_response.status_code == 200
+    update_body = update_response.json()
+    assert update_body["pricing"]["input_cache_read"] == pytest.approx(1.25e-9)
+    assert update_body["pricing"]["input_cache_write"] == pytest.approx(4.5e-9)
+
+    await integration_session.refresh(row)
+    updated_pricing = json.loads(row.pricing)
+    assert updated_pricing["input_cache_read"] == pytest.approx(1.25e-9)
+    assert updated_pricing["input_cache_write"] == pytest.approx(4.5e-9)
+
+    model = get_model_instance("custom-cache-model")
+    assert model is not None
+    assert model.sats_pricing is not None
+    assert model.sats_pricing.input_cache_read == pytest.approx(0.00125)
+    assert model.sats_pricing.input_cache_write == pytest.approx(0.0045)
+
+    with patch("routstr.payment.cost_calculation.sats_usd_price", return_value=1e-6):
+        cost = await calculate_cost(
+            {
+                "model": "custom-cache-model",
+                "usage": {
+                    "prompt_tokens": 1000,
+                    "completion_tokens": 100,
+                    "prompt_tokens_details": {"cached_tokens": 800},
+                },
+            },
+            max_cost=1_000_000,
+        )
+
+    assert isinstance(cost, CostData)
+    assert cost.input_tokens == 200
+    assert cost.cache_read_input_tokens == 800
+    assert cost.cache_read_msats == 1000
+    assert cost.output_msats == 28000
+    assert cost.input_msats == 29000
+    assert cost.total_msats == 57000
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_upstream_response_cost_uses_model_cache_pricing(
+    integration_session: AsyncSession,
+    patched_db_engine: None,
+) -> None:
+    """A model's configured cache price must discount upstream cached-token usage."""
+    provider = UpstreamProviderRow(
+        provider_type="generic",
+        base_url="https://cache-priced-upstream.example/v1",
+        api_key="test-key",
+        provider_fee=1.0,
+    )
+    integration_session.add(provider)
+    await integration_session.commit()
+    await integration_session.refresh(provider)
+    assert provider.id is not None
+
+    row = ModelRow(
+        id="cache-priced-model",
+        name="Cache Priced Model",
+        description="model seeded with explicit cache pricing",
+        created=0,
+        context_length=128000,
+        architecture=json.dumps(
+            {
+                "modality": "text",
+                "input_modalities": ["text"],
+                "output_modalities": ["text"],
+                "tokenizer": "unknown",
+                "instruct_type": None,
+            }
+        ),
+        pricing=json.dumps(
+            {
+                "prompt": 1.4e-7,
+                "completion": 2.8e-7,
+                "input_cache_read": 1.25e-9,
+                "input_cache_write": 4.5e-9,
+            }
+        ),
+        upstream_provider_id=provider.id,
+        enabled=True,
+        forwarded_model_id="cache-priced-model",
+    )
+    integration_session.add(row)
+    await integration_session.commit()
+
+    with patch("routstr.payment.models.sats_usd_price", return_value=1e-6):
+        await reinitialize_upstreams()
+
+    with patch("routstr.payment.cost_calculation.sats_usd_price", return_value=1e-6):
+        cost = await calculate_cost(
+            {
+                "model": "cache-priced-model",
+                "usage": {
+                    "prompt_tokens": 1000,
+                    "completion_tokens": 100,
+                    "prompt_tokens_details": {"cached_tokens": 800},
+                },
+            },
+            max_cost=1_000_000,
+        )
+
+    assert isinstance(cost, CostData)
+    # prompt 1.4e-7 USD/token -> 0.14 sats/token -> 140 msats/token
+    # cache 1.25e-9 USD/token -> 0.00125 sats/token -> 1.25 msats/token
+    # completion 2.8e-7 USD/token -> 0.28 sats/token -> 280 msats/token
+    assert cost.input_tokens == 200
+    assert cost.cache_read_input_tokens == 800
+    assert cost.cache_read_msats == 1000
+    assert cost.input_msats == 29000
+    assert cost.output_msats == 28000
+    assert cost.total_msats == 57000

tests/integration/test_free_response_stale_reservation.py

@@ -0,0 +1,146 @@
+"""Regression tests for charging after stale reservation cleanup."""
+
+import time
+import uuid
+from unittest.mock import patch
+
+import pytest
+from sqlmodel.ext.asyncio.session import AsyncSession
+
+from routstr.core.db import ApiKey
+from routstr.payment.cost_calculation import CostData
+
+
+def _make_key(balance: int, reserved: int) -> ApiKey:
+    return ApiKey(
+        hashed_key=f"test_{uuid.uuid4().hex}",
+        balance=balance,
+        reserved_balance=reserved,
+        total_spent=0,
+        total_requests=1,
+    )
+
+
+def _cost_data(total_msats: int) -> CostData:
+    return CostData(
+        base_msats=0,
+        input_msats=total_msats // 2,
+        output_msats=total_msats - total_msats // 2,
+        total_msats=total_msats,
+        total_usd=0.0,
+        input_tokens=100,
+        output_tokens=100,
+    )
+
+
+@pytest.mark.asyncio
+async def test_overrun_charges_after_reservation_swept(
+    integration_session: AsyncSession,
+) -> None:
+    """Overrun finalize must charge even when the reservation was already released."""
+    from routstr.auth import adjust_payment_for_tokens
+
+    deducted_max_cost = 990  # discounted reservation
+    actual_token_cost = 1000  # actual cost overruns the reservation
+
+    # Sweeper has zeroed reserved_balance but left balance untouched.
+    key = _make_key(balance=1000, reserved=0)
+    integration_session.add(key)
+    await integration_session.commit()
+
+    response_data = {
+        "model": "test-model",
+        "usage": {"prompt_tokens": 100, "completion_tokens": 100},
+    }
+
+    with patch(
+        "routstr.auth.calculate_cost",
+        return_value=_cost_data(actual_token_cost),
+    ):
+        await adjust_payment_for_tokens(
+            key, response_data, integration_session, deducted_max_cost
+        )
+
+    await integration_session.refresh(key)
+
+    assert key.total_spent == actual_token_cost, (
+        f"Request was not billed (total_spent={key.total_spent}) — free response bug"
+    )
+    assert key.balance == 1000 - actual_token_cost, (
+        f"Balance not charged: {key.balance}"
+    )
+    assert key.balance >= 0
+    assert key.reserved_balance == 0
+
+
+@pytest.mark.asyncio
+async def test_free_response_path_closed_end_to_end(
+    integration_session: AsyncSession,
+    patched_db_engine: None,
+) -> None:
+    """A reservation released by the real sweeper must not yield a free response."""
+    from routstr.auth import adjust_payment_for_tokens, pay_for_request
+    from routstr.core.db import create_session, release_stale_reservations
+
+    deducted_max_cost = 990
+    actual_token_cost = 1000
+    key_hash = f"test_sweep_{uuid.uuid4().hex}"
+
+    async with create_session() as session:
+        session.add(
+            ApiKey(
+                hashed_key=key_hash,
+                balance=1000,
+                reserved_balance=0,
+                total_spent=0,
+                total_requests=0,
+            )
+        )
+        await session.commit()
+
+    # Reserve the request, then backdate reserved_at so the sweeper treats it as
+    # stale (simulates a stream that outlived stale_reservation_timeout_seconds).
+    async with create_session() as session:
+        key = await session.get(ApiKey, key_hash)
+        assert key is not None
+        await pay_for_request(key, deducted_max_cost, session)
+        await session.refresh(key)
+        assert key.reserved_balance == deducted_max_cost
+        key.reserved_at = int(time.time()) - 10_000
+        session.add(key)
+        await session.commit()
+
+    # Sweeper releases the stale reservation without charging.
+    async with create_session() as session:
+        released = await release_stale_reservations(session, max_age_seconds=300)
+        assert released == 1
+
+    async with create_session() as session:
+        key = await session.get(ApiKey, key_hash)
+        assert key is not None
+        assert key.reserved_balance == 0, "Precondition: sweeper zeroed the reservation"
+
+        response_data = {
+            "model": "test-model",
+            "usage": {"prompt_tokens": 100, "completion_tokens": 100},
+        }
+        with patch(
+            "routstr.auth.calculate_cost",
+            return_value=_cost_data(actual_token_cost),
+        ):
+            await adjust_payment_for_tokens(
+                key, response_data, session, deducted_max_cost
+            )
+
+    async with create_session() as session:
+        final = await session.get(ApiKey, key_hash)
+        assert final is not None
+
+    assert final.total_spent == actual_token_cost, (
+        f"Free response: total_spent={final.total_spent}, expected {actual_token_cost}"
+    )
+    assert final.balance == 1000 - actual_token_cost, (
+        f"Balance not charged after sweep: {final.balance}"
+    )
+    assert final.balance >= 0
+    assert final.reserved_balance == 0

tests/integration/test_provider_self_lookup.py

@@ -0,0 +1,128 @@
+"""A live upstream provider resolves its OWN database row by stable identity
+(its primary key), not by its mutable/secret ``api_key``.
+
+Today ``from_db_row`` drops ``provider_row.id`` and the two self-referential
+paths — PPQ.AI's insufficient-balance self-disable and the base
+``refresh_models_cache`` — re-find their own row with
+``WHERE base_url == self.base_url AND api_key == self.api_key``. That uses a
+rotatable secret as a self-handle: the moment the row's key changes underneath a
+live object (a rotation racing an in-flight request), the object can no longer
+find itself. These tests pin the invariant that a provider looks itself up by
+identity, so the lookup survives a key change (and, later, key encryption).
+"""
+
+from unittest.mock import AsyncMock, patch
+
+import pytest
+
+from routstr.core.db import UpstreamProviderRow
+from routstr.upstream.ppqai import PPQAIUpstreamProvider
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_provider_object_carries_its_persistent_identity(
+    integration_session: object,
+    patched_db_engine: None,
+) -> None:
+    """``from_db_row`` gives the in-memory object its row's identity (``db_id``)."""
+    row = UpstreamProviderRow(
+        provider_type="ppqai",
+        base_url="https://api.ppq.ai",
+        api_key="sk-original",
+        enabled=True,
+        provider_fee=1.0,
+    )
+    integration_session.add(row)  # type: ignore[attr-defined]
+    await integration_session.commit()  # type: ignore[attr-defined]
+    await integration_session.refresh(row)  # type: ignore[attr-defined]
+
+    provider = PPQAIUpstreamProvider.from_db_row(row)
+    assert provider is not None
+
+    assert provider.db_id == row.id
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_self_disable_targets_own_row_after_key_rotation(
+    integration_session: object,
+    patched_db_engine: None,
+) -> None:
+    """PPQ.AI self-disable must disable *its* row even after the key rotated.
+
+    RED (current): the object holds the pre-rotation key, so the
+    ``(base_url, api_key)`` lookup misses the row → the provider is never
+    disabled. GREEN: lookup by ``id`` finds it and disables it.
+    """
+    row = UpstreamProviderRow(
+        provider_type="ppqai",
+        base_url="https://api.ppq.ai",
+        api_key="sk-original",
+        enabled=True,
+        provider_fee=1.0,
+    )
+    integration_session.add(row)  # type: ignore[attr-defined]
+    await integration_session.commit()  # type: ignore[attr-defined]
+    await integration_session.refresh(row)  # type: ignore[attr-defined]
+
+    provider = PPQAIUpstreamProvider.from_db_row(row)  # captures sk-original
+    assert provider is not None
+
+    # Key is rotated in the DB while `provider` is still live.
+    row.api_key = "sk-rotated"
+    integration_session.add(row)  # type: ignore[attr-defined]
+    await integration_session.commit()  # type: ignore[attr-defined]
+
+    with patch("routstr.proxy.reinitialize_upstreams", new=AsyncMock()):
+        await provider.on_upstream_error_redirect(402, "Insufficient balance")
+
+    await integration_session.refresh(row)  # type: ignore[attr-defined]
+    assert row.enabled is False
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_refresh_models_cache_finds_own_row_after_key_rotation(
+    integration_session: object,
+    patched_db_engine: None,
+) -> None:
+    """``refresh_models_cache`` must resolve its own row after a key rotation.
+
+    ``refresh_models_cache`` swallows every exception (it only logs), so the
+    observable proof it found its row is that it reaches ``list_models`` — which
+    is called with the row's ``id`` only *after* the row is resolved. RED
+    (current): the stale-key ``(base_url, api_key)`` lookup returns nothing, the
+    method raises ``404`` internally and returns before ``list_models`` is ever
+    called. GREEN: lookup by ``id`` finds the row and ``list_models`` runs for
+    that ``id``.
+    """
+    row = UpstreamProviderRow(
+        provider_type="ppqai",
+        base_url="https://api.ppq.ai",
+        api_key="sk-original",
+        enabled=True,
+        provider_fee=1.0,
+    )
+    integration_session.add(row)  # type: ignore[attr-defined]
+    await integration_session.commit()  # type: ignore[attr-defined]
+    await integration_session.refresh(row)  # type: ignore[attr-defined]
+    row_id = row.id
+
+    provider = PPQAIUpstreamProvider.from_db_row(row)  # captures sk-original
+    assert provider is not None
+
+    row.api_key = "sk-rotated"
+    integration_session.add(row)  # type: ignore[attr-defined]
+    await integration_session.commit()  # type: ignore[attr-defined]
+
+    list_models_mock = AsyncMock(return_value=[])
+    with (
+        patch.object(provider, "fetch_models", new=AsyncMock(return_value=[])),
+        patch("routstr.upstream.base.list_models", new=list_models_mock),
+    ):
+        await provider.refresh_models_cache()
+
+    list_models_mock.assert_awaited_once()
+    assert list_models_mock.await_args is not None
+    assert list_models_mock.await_args.kwargs["upstream_id"] == row_id

tests/integration/test_prune_dead_api_keys.py

@@ -0,0 +1,283 @@
+"""
+Tests for prune_dead_api_keys — the janitor that removes provably-dead 0/0/0
+API keys (funded keys fully refunded/expired without ever being used, plus bare
+orphans), while protecting keys that are still meaningful.
+"""
+
+import time
+import uuid
+from collections.abc import Awaitable, Callable
+from typing import Any, cast
+
+import pytest
+from sqlalchemy.sql.dml import Update
+from sqlmodel import col, update
+
+from routstr.core.db import (
+    ApiKey,
+    CashuTransaction,
+    LightningInvoice,
+    create_session,
+    prune_dead_api_keys,
+)
+
+OLD = 100  # min_age_seconds used by the tests
+NOW = int(time.time())
+LONG_AGO = NOW - 10_000  # well past the grace period
+
+
+async def _exists(key_hash: str) -> bool:
+    async with create_session() as session:
+        return (await session.get(ApiKey, key_hash)) is not None
+
+
+def _dead_key(created_at: int | None) -> ApiKey:
+    return ApiKey(
+        hashed_key=f"dead_{uuid.uuid4().hex}",
+        balance=0,
+        reserved_balance=0,
+        total_spent=0,
+        total_requests=0,
+        created_at=created_at,
+    )
+
+
+@pytest.mark.asyncio
+async def test_prunes_old_refunded_zero_key(patched_db_engine: None) -> None:
+    """A funded-then-refunded key (0/0/0, NULL parent, old) is pruned."""
+    key = _dead_key(LONG_AGO)
+    async with create_session() as session:
+        session.add(key)
+        await session.commit()
+
+    async with create_session() as session:
+        pruned = await prune_dead_api_keys(session, OLD)
+
+    assert pruned == 1
+    assert not await _exists(key.hashed_key)
+
+
+@pytest.mark.asyncio
+async def test_grace_period_protects_fresh_key(patched_db_engine: None) -> None:
+    """A dead-looking but recently created key is protected by the grace period."""
+    key = _dead_key(int(time.time()))
+    async with create_session() as session:
+        session.add(key)
+        await session.commit()
+
+    async with create_session() as session:
+        pruned = await prune_dead_api_keys(session, OLD)
+
+    assert pruned == 0
+    assert await _exists(key.hashed_key)
+
+
+@pytest.mark.asyncio
+async def test_used_key_never_pruned(patched_db_engine: None) -> None:
+    """Keys with any spend/requests or live balance are never pruned."""
+    spent = _dead_key(LONG_AGO)
+    spent.total_spent = 1
+    requested = _dead_key(LONG_AGO)
+    requested.total_requests = 1
+    funded = _dead_key(LONG_AGO)
+    funded.balance = 1000
+    reserved = _dead_key(LONG_AGO)
+    reserved.reserved_balance = 500
+
+    async with create_session() as session:
+        for k in (spent, requested, funded, reserved):
+            session.add(k)
+        await session.commit()
+
+    async with create_session() as session:
+        pruned = await prune_dead_api_keys(session, OLD)
+
+    assert pruned == 0
+    for k in (spent, requested, funded, reserved):
+        assert await _exists(k.hashed_key)
+
+
+@pytest.mark.asyncio
+async def test_parent_and_child_keys_are_not_pruned(
+    patched_db_engine: None,
+) -> None:
+    """Pruning must not orphan child keys or delete valid children."""
+    parent = _dead_key(LONG_AGO)
+    child = ApiKey(
+        hashed_key=f"child_{uuid.uuid4().hex}",
+        balance=0,
+        reserved_balance=0,
+        total_spent=0,
+        total_requests=0,
+        created_at=LONG_AGO,
+        parent_key_hash=parent.hashed_key,
+    )
+    async with create_session() as session:
+        session.add(parent)
+        session.add(child)
+        await session.commit()
+
+    async with create_session() as session:
+        pruned = await prune_dead_api_keys(session, OLD)
+
+    assert pruned == 0
+    assert await _exists(parent.hashed_key)
+    assert await _exists(child.hashed_key)
+
+
+@pytest.mark.asyncio
+async def test_pending_invoice_protects_key(patched_db_engine: None) -> None:
+    """A key referenced by a pending topup invoice is never pruned mid-topup."""
+    key = _dead_key(LONG_AGO)
+    invoice = LightningInvoice(
+        id=f"inv_{uuid.uuid4().hex}",
+        bolt11=f"lnbc_{uuid.uuid4().hex}",
+        amount_sats=10,
+        description="topup",
+        payment_hash=uuid.uuid4().hex,
+        status="pending",
+        api_key_hash=key.hashed_key,
+        purpose="topup",
+        expires_at=NOW + 10_000,
+    )
+    async with create_session() as session:
+        session.add(key)
+        session.add(invoice)
+        await session.commit()
+
+    async with create_session() as session:
+        pruned = await prune_dead_api_keys(session, OLD)
+
+    assert pruned == 0
+    assert await _exists(key.hashed_key)
+
+
+@pytest.mark.asyncio
+async def test_paid_invoice_does_not_protect_key(patched_db_engine: None) -> None:
+    """A settled (non-pending) invoice does not keep a dead key alive."""
+    key = _dead_key(LONG_AGO)
+    invoice = LightningInvoice(
+        id=f"inv_{uuid.uuid4().hex}",
+        bolt11=f"lnbc_{uuid.uuid4().hex}",
+        amount_sats=10,
+        description="topup",
+        payment_hash=uuid.uuid4().hex,
+        status="paid",
+        api_key_hash=key.hashed_key,
+        purpose="topup",
+        expires_at=NOW - 1,
+    )
+    async with create_session() as session:
+        session.add(key)
+        session.add(invoice)
+        await session.commit()
+
+    async with create_session() as session:
+        pruned = await prune_dead_api_keys(session, OLD)
+
+    assert pruned == 1
+    assert not await _exists(key.hashed_key)
+
+
+@pytest.mark.asyncio
+async def test_key_that_becomes_meaningful_during_prune_survives(
+    patched_db_engine: None, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """Revalidate before unlink/delete so a late top-up cannot be pruned."""
+    key = _dead_key(LONG_AGO)
+    txn = CashuTransaction(
+        id=uuid.uuid4().hex,
+        token="cashuABC",
+        amount=21,
+        unit="sat",
+        type="in",
+        source="apikey",
+        api_key_hashed_key=key.hashed_key,
+    )
+    async with create_session() as session:
+        session.add(key)
+        session.add(txn)
+        await session.commit()
+
+    async with create_session() as session:
+        original_exec = cast(Callable[..., Awaitable[Any]], session.exec)
+        topped_up = False
+
+        async def exec_with_late_topup(
+            statement: Any, *args: Any, **kwargs: Any
+        ) -> Any:
+            nonlocal topped_up
+            if (
+                not topped_up
+                and isinstance(statement, Update)
+                and getattr(statement.table, "name", None) == "cashu_transactions"
+            ):
+                topped_up = True
+                async with create_session() as topup_session:
+                    await topup_session.exec(  # type: ignore[call-overload]
+                        update(ApiKey)
+                        .where(col(ApiKey.hashed_key) == key.hashed_key)
+                        .values(balance=42)
+                    )
+                    await topup_session.commit()
+            return await original_exec(statement, *args, **kwargs)
+
+        monkeypatch.setattr(session, "exec", exec_with_late_topup)
+        pruned = await prune_dead_api_keys(session, OLD)
+
+    assert pruned == 0
+    assert await _exists(key.hashed_key)
+    async with create_session() as session:
+        surviving = await session.get(CashuTransaction, txn.id)
+        assert surviving is not None
+        assert surviving.api_key_hashed_key == key.hashed_key
+
+
+@pytest.mark.asyncio
+async def test_transaction_audit_trail_preserved(patched_db_engine: None) -> None:
+    """Pruning a refunded key keeps its cashu_transactions, unlinked from the key."""
+    key = _dead_key(LONG_AGO)
+    txn = CashuTransaction(
+        id=uuid.uuid4().hex,
+        token="cashuABC",
+        amount=21,
+        unit="sat",
+        type="in",
+        source="apikey",
+        api_key_hashed_key=key.hashed_key,
+    )
+    async with create_session() as session:
+        session.add(key)
+        session.add(txn)
+        await session.commit()
+
+    async with create_session() as session:
+        pruned = await prune_dead_api_keys(session, OLD)
+
+    assert pruned == 1
+    assert not await _exists(key.hashed_key)
+
+    async with create_session() as session:
+        surviving = await session.get(CashuTransaction, txn.id)
+        assert surviving is not None, "Financial audit row must survive key deletion"
+        assert surviving.api_key_hashed_key is None, "Link must be nulled, not dangling"
+        assert surviving.amount == 21
+
+
+@pytest.mark.asyncio
+async def test_periodic_prune_disabled_returns_immediately(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """Non-positive intervals disable the janitor."""
+    from unittest.mock import AsyncMock
+
+    from routstr import auth
+    from routstr.core.settings import settings
+
+    monkeypatch.setattr(settings, "dead_key_prune_interval_seconds", 0)
+    sleep_mock = AsyncMock()
+    monkeypatch.setattr(auth.asyncio, "sleep", sleep_mock)
+
+    await auth.periodic_dead_key_prune()
+
+    sleep_mock.assert_not_called()

tests/integration/test_swap_fee_retry.py

@@ -0,0 +1,194 @@
+"""
+Integration tests for reactive swap fee retries via the wallet topup endpoint.
+
+Foreign-mint tokens are swapped to the primary mint using the foreign mint's
+melt quote, whose fee_reserve is a non-binding estimate (NUT-05): the mint may
+demand more when re-quoting or at melt execution. These tests cover the
+endpoint behaviour in those cases:
+
+1. The mint demands one sat more at melt time than every quote reported
+   (the mint.cubabitcoin.org incident): the swap retries with a smaller
+   invoice and the topup succeeds, crediting the recomputed amount.
+2. The real melt quote reports a higher fee_reserve than the estimate: the
+   swap re-quotes from the observed fee and the topup succeeds.
+3. The mint escalates its fee demands on every attempt: the retry budget is
+   exhausted and the endpoint returns 400 with a clear error (never 500),
+   without ever executing a melt.
+"""
+
+from collections.abc import Callable
+from unittest.mock import AsyncMock, Mock, patch
+
+import pytest
+from httpx import AsyncClient, Response
+
+from routstr.core.settings import settings
+
+# Captured at collection time, before the integration_app fixture replaces it
+# with the testmint stub that bypasses swapping (see conftest.py).
+from routstr.wallet import recieve_token as _real_recieve_token
+
+PRIMARY_MINT = "http://primary:3338"
+
+
+def _make_swap_mocks(
+    token_amount: int,
+    fee_reserves: list[int],
+    input_fees: int = 0,
+    mint_url: str = "http://foreign-mint:3338",
+) -> tuple[Mock, Mock, Mock]:
+    """Return (token, token_wallet, primary_wallet) mocks that act like a mint.
+
+    Mint quotes pass the requested amount through their ``request`` field and
+    melt quotes echo that amount back, so the mocks stay consistent for
+    whatever amounts the implementation requests. ``fee_reserves`` supplies the
+    fee_reserve of each successive melt quote (the first serves the estimation
+    pass); requesting more quotes than provided fails the test.
+    """
+    mock_token = Mock()
+    mock_token.mint = mint_url
+    mock_token.unit = "sat"
+    mock_token.amount = token_amount
+    mock_token.keysets = ["keyset1"]
+    mock_token.proofs = [Mock(amount=token_amount)]
+
+    mock_token_wallet = Mock()
+    mock_token_wallet.load_mint = AsyncMock()
+    mock_token_wallet.load_proofs = AsyncMock()
+    mock_token_wallet.get_fees_for_proofs = Mock(return_value=input_fees)
+
+    mock_primary_wallet = Mock()
+    mock_primary_wallet.load_mint = AsyncMock()
+    mock_primary_wallet.load_proofs = AsyncMock()
+    mock_primary_wallet.available_balance = Mock(amount=0)
+    mock_primary_wallet.mint = AsyncMock(return_value=Mock())
+
+    fees = iter(fee_reserves)
+
+    def _next_fee() -> int:
+        try:
+            return next(fees)
+        except StopIteration:
+            raise AssertionError(
+                "more melt quotes requested than fee_reserves provided"
+            ) from None
+
+    mock_primary_wallet.request_mint = AsyncMock(
+        side_effect=lambda amount: Mock(quote=f"mint_quote_{amount}", request=amount)
+    )
+    mock_token_wallet.melt_quote = AsyncMock(
+        side_effect=lambda invoice: Mock(
+            quote=f"melt_quote_{invoice}", amount=invoice, fee_reserve=_next_fee()
+        )
+    )
+    mock_token_wallet.melt = AsyncMock(return_value=Mock())
+
+    return mock_token, mock_token_wallet, mock_primary_wallet
+
+
+def _wallet_router(primary_wallet: Mock, token_wallet: Mock) -> Callable[..., Mock]:
+    """Route get_wallet calls to the primary or foreign wallet mock by URL."""
+
+    def fake_get_wallet(mint_url: str, unit: str = "sat", load: bool = True) -> Mock:
+        return primary_wallet if mint_url == PRIMARY_MINT else token_wallet
+
+    return fake_get_wallet
+
+
+async def _post_topup(
+    client: AsyncClient,
+    mock_token: Mock,
+    token_wallet: Mock,
+    primary_wallet: Mock,
+) -> Response:
+    """POST /v1/wallet/topup with the swap layer mocked at the mint boundary.
+
+    The conftest's testmint stub for recieve_token is swapped back for the
+    real implementation so the request exercises the actual swap path.
+    """
+    with patch("routstr.wallet.recieve_token", _real_recieve_token):
+        with patch(
+            "routstr.wallet.deserialize_token_from_string", return_value=mock_token
+        ):
+            with patch(
+                "routstr.wallet.get_wallet",
+                side_effect=_wallet_router(primary_wallet, token_wallet),
+            ):
+                with patch.object(settings, "primary_mint", PRIMARY_MINT):
+                    with patch.object(settings, "primary_mint_unit", "sat"):
+                        with patch.object(settings, "cashu_mints", [PRIMARY_MINT]):
+                            return await client.post(
+                                "/v1/wallet/topup",
+                                params={"cashu_token": "cashuAtest_foreign_token"},
+                            )
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_topup_retries_when_melt_demands_more_than_quoted(
+    authenticated_client: AsyncClient,
+) -> None:
+    """A 179-sat token where every quote reports fee_reserve=1 but the mint
+    rejects the first melt demanding 180. The retry shrinks the invoice to 177
+    and the topup credits 177 sats (177_000 msats)."""
+    mock_token, token_wallet, primary_wallet = _make_swap_mocks(
+        179, fee_reserves=[1, 1, 1], mint_url="http://mint.cubabitcoin.org"
+    )
+    token_wallet.melt.side_effect = [
+        Exception(
+            "Mint Error: not enough inputs provided for melt. "
+            "Provided: 179, needed: 180 (Code: 11000)"
+        ),
+        Mock(),
+    ]
+
+    response = await _post_topup(
+        authenticated_client, mock_token, token_wallet, primary_wallet
+    )
+
+    assert response.status_code == 200
+    assert response.json()["msats"] == 177_000
+    assert token_wallet.melt.call_count == 2
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_topup_retries_when_quote_fee_exceeds_estimate(
+    authenticated_client: AsyncClient,
+) -> None:
+    """A 1000-sat token estimated at fee 20, but the real quote demands 23.
+    The retry recomputes 1000 - 23 = 977, which fits, and the topup credits
+    977 sats (977_000 msats) with a single melt."""
+    mock_token, token_wallet, primary_wallet = _make_swap_mocks(
+        1000, fee_reserves=[20, 23, 23]
+    )
+
+    response = await _post_topup(
+        authenticated_client, mock_token, token_wallet, primary_wallet
+    )
+
+    assert response.status_code == 200
+    assert response.json()["msats"] == 977_000
+    assert token_wallet.melt.call_count == 1
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_topup_returns_422_when_retries_exhausted(
+    authenticated_client: AsyncClient,
+) -> None:
+    """A mint that escalates fee_reserve on every re-quote (1 → 10 → 25 → 50)
+    exhausts the retry budget: clean 422 mint_error/too-small taxonomy, melt
+    never executed."""
+    mock_token, token_wallet, primary_wallet = _make_swap_mocks(
+        1000, fee_reserves=[1, 10, 25, 50]
+    )
+
+    response = await _post_topup(
+        authenticated_client, mock_token, token_wallet, primary_wallet
+    )
+
+    assert response.status_code == 422
+    assert "too small to cover swap fees" in response.json()["detail"]
+    assert token_wallet.melt_quote.call_count == 4  # estimation + 3 attempts
+    token_wallet.melt.assert_not_called()

tests/integration/test_wallet_authentication.py

@@ -12,10 +12,7 @@ from sqlmodel import select
 
 from routstr.core.db import ApiKey
 
-from .utils import (
-    CashuTokenGenerator,
-    ResponseValidator,
-)
+from .utils import ResponseValidator
 
 
 @pytest.mark.integration
@@ -79,29 +76,31 @@ async def test_api_key_generation_invalid_token(
     # Capture initial state
     await db_snapshot.capture()
 
-    # Test various invalid tokens
-    invalid_tokens = [
-        CashuTokenGenerator.generate_invalid_token(),  # Malformed token
-        "not-a-cashu-token",  # Wrong format
-        "cashuA",  # Empty token
-        "cashuA" + "x" * 1000,  # Invalid base64
+    # Non-Cashu bearer values are invalid API keys (401). Malformed values that
+    # look like Cashu tokens use the shared Cashu taxonomy (400 invalid_token).
+    invalid_tokens: list[tuple[str, int, str | None]] = [
+        ("not-a-cashu-token", 401, None),
+        ("sk-not-a-real-api-key", 401, None),
+        ("cashuA", 400, "invalid_cashu_token"),
+        ("cashuA" + "x" * 1000, 400, "invalid_cashu_token"),
     ]
 
-    for invalid_token in invalid_tokens:
+    for invalid_token, expected_status, expected_code in invalid_tokens:
         integration_client.headers["Authorization"] = f"Bearer {invalid_token}"
         response = await integration_client.get("/v1/wallet/info")
 
-        # Should fail with 401
-        assert response.status_code == 401, (
+        assert response.status_code == expected_status, (
             f"Token {invalid_token[:20]}... should be invalid"
         )
 
         # Validate error response
         validator = ResponseValidator()
         error_validation = validator.validate_error_response(
-            response, expected_status=401, expected_error_key="detail"
+            response, expected_status=expected_status, expected_error_key="detail"
         )
         assert error_validation["valid"]
+        if expected_code is not None:
+            assert response.json()["detail"]["error"]["code"] == expected_code
 
     # Verify no database changes
     diff = await db_snapshot.diff()

tests/integration/test_wallet_refund.py

@@ -14,6 +14,7 @@ from httpx import AsyncClient
 from sqlmodel import select
 
 from routstr.core.db import ApiKey, CashuTransaction
+from routstr.wallet import MintConnectionError
 
 
 @pytest.mark.integration
@@ -503,26 +504,19 @@ async def test_mint_unavailability_handling(
 
     # The global mock in conftest.py is already in place,
     # so we need to temporarily modify it
-    from unittest.mock import patch
+    raw_error = "Mint unavailable: Connection refused"
 
-    # Make the send_token method raise an exception
+    # Make the send_token method raise a typed mint connection exception.
     with patch(
         "routstr.balance.send_token",
-        side_effect=Exception("Mint unavailable: Connection refused"),
+        side_effect=MintConnectionError(raw_error),
     ):
-        # The exception should propagate as a 503 error (Service Unavailable)
-        # But we need to handle it properly
-        try:
-            response = await authenticated_client.post("/v1/wallet/refund")
-            # If we get here, check the status code
-            assert response.status_code == 503
-            assert "Mint service unavailable" in response.json()["detail"]
-        except Exception as e:
-            # If the exception propagates, that's also a failure scenario
-            assert "Mint unavailable" in str(e)
+        response = await authenticated_client.post("/v1/wallet/refund")
+        assert response.status_code == 503
+        assert response.json()["detail"] == "Mint service unavailable"
+        assert raw_error not in response.text
 
         # Balance should remain unchanged (transaction should roll back)
-        # Note: Current implementation might not handle this perfectly
         wallet_response = await authenticated_client.get("/v1/wallet/")
         assert wallet_response.status_code == 200
         assert wallet_response.json()["balance"] == 10_000_000

tests/unit/test_admin_withdraw.py

@@ -0,0 +1,51 @@
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, Mock
+
+import pytest
+
+from routstr.core import admin
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("requested_mint", [None, "https://secondary.example"])
+async def test_withdraw_uses_effective_mint_and_records_outgoing_transaction(
+    monkeypatch: pytest.MonkeyPatch, requested_mint: str | None
+) -> None:
+    primary_mint = "https://primary.example"
+    effective_mint = requested_mint or primary_mint
+    wallet = object()
+    proofs = [SimpleNamespace(amount=40), SimpleNamespace(amount=60)]
+    token = "cashuBoutgoing"
+
+    get_wallet = AsyncMock(return_value=wallet)
+    get_proofs = Mock(return_value=proofs)
+    filter_proofs = AsyncMock(return_value=proofs)
+    send_token = AsyncMock(return_value=token)
+    store_transaction = AsyncMock(return_value=True)
+
+    monkeypatch.setattr(admin, "get_wallet", get_wallet)
+    monkeypatch.setattr(admin, "get_proofs_per_mint_and_unit", get_proofs)
+    monkeypatch.setattr(admin, "slow_filter_spend_proofs", filter_proofs)
+    monkeypatch.setattr(admin, "send_token", send_token)
+    monkeypatch.setattr(admin, "store_cashu_transaction", store_transaction)
+    monkeypatch.setattr(admin.settings, "primary_mint", primary_mint)
+
+    result = await admin.withdraw(
+        Mock(),
+        admin.WithdrawRequest(amount=75, mint_url=requested_mint, unit="sat"),
+    )
+
+    assert result == {"token": token}
+    get_wallet.assert_awaited_once_with(effective_mint, "sat")
+    get_proofs.assert_called_once_with(wallet, effective_mint, "sat", not_reserved=True)
+    filter_proofs.assert_awaited_once_with(proofs, wallet)
+    send_token.assert_awaited_once_with(75, "sat", effective_mint)
+    store_transaction.assert_awaited_once_with(
+        token=token,
+        amount=75,
+        unit="sat",
+        mint_url=effective_mint,
+        typ="out",
+        collected=False,
+        source="admin",
+    )

tests/unit/test_algorithm.py

@@ -137,8 +137,8 @@ def test_create_model_mappings_includes_db_override_for_missing_cached_model(
 
     model_instances, provider_map, unique_models = create_model_mappings(
         upstreams=[provider],
-        overrides_by_id={"azure/gpt-4o": (override_row, 1.01)},
-        disabled_model_ids=set(),
+        overrides_by_key={("azure/gpt-4o", 7): (override_row, 1.01)},
+        disabled_model_keys=set(),
     )
 
     assert "azure/gpt-4o" in model_instances
@@ -182,11 +182,73 @@ def test_create_model_mappings_dedupes_with_provider_identity_not_provider_type(
 
     _, provider_map, _ = create_model_mappings(
         upstreams=[provider_a, provider_b],
-        overrides_by_id={"azure/gpt-4o": (override_row, 1.01)},
-        disabled_model_ids=set(),
+        overrides_by_key={("azure/gpt-4o", 2): (override_row, 1.01)},
+        disabled_model_keys=set(),
     )
 
     providers_for_alias = provider_map["azure/gpt-4o"]
     assert provider_a in providers_for_alias
     assert provider_b in providers_for_alias
     assert len(providers_for_alias) == 2
+
+
+def test_create_model_mappings_applies_override_only_to_matching_provider(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """Same-id overrides must not add provider-specific aliases to other providers."""
+    provider_a_model = create_test_model("same-id", prompt_price=0.01)
+    provider_a = create_test_provider(
+        "provider-a",
+        "https://provider-a.example/v1",
+        db_id=1,
+        models=[provider_a_model],
+    )
+    provider_b_model = create_test_model("same-id", prompt_price=0.02)
+    provider_b = create_test_provider(
+        "provider-b",
+        "https://provider-b.example/v1",
+        db_id=2,
+        models=[provider_b_model],
+    )
+
+    override_model = create_test_model("same-id", prompt_price=0.001)
+    override_model.alias_ids = ["provider-b-only"]
+    override_row = SimpleNamespace(id="same-id", upstream_provider_id=2, enabled=True)
+
+    def fake_row_to_model(*args, **kwargs) -> Model:  # type: ignore[no-untyped-def]
+        return override_model
+
+    monkeypatch.setattr("routstr.payment.models._row_to_model", fake_row_to_model)
+
+    _, provider_map, _ = create_model_mappings(
+        upstreams=[provider_a, provider_b],
+        overrides_by_key={("same-id", 2): (override_row, 1.01)},
+        disabled_model_keys=set(),
+    )
+
+    assert provider_map["provider-b-only"] == [provider_b]
+    assert set(provider_map["same-id"]) == {provider_a, provider_b}
+
+
+def test_create_model_mappings_disables_only_matching_provider() -> None:
+    """Disabled overrides are scoped to the provider row, not the shared model id."""
+    provider_a = create_test_provider(
+        "provider-a",
+        "https://provider-a.example/v1",
+        db_id=1,
+        models=[create_test_model("same-id")],
+    )
+    provider_b = create_test_provider(
+        "provider-b",
+        "https://provider-b.example/v1",
+        db_id=2,
+        models=[create_test_model("same-id")],
+    )
+
+    _, provider_map, _ = create_model_mappings(
+        upstreams=[provider_a, provider_b],
+        overrides_by_key={},
+        disabled_model_keys={("same-id", 2)},
+    )
+
+    assert provider_map["same-id"] == [provider_a]

tests/unit/test_auth_cashu.py

@@ -1,8 +1,9 @@
 import hashlib
 from types import SimpleNamespace
-from typing import AsyncGenerator
+from typing import AsyncGenerator, cast
 from unittest.mock import AsyncMock, patch
 
+import httpx
 import pytest
 from fastapi import HTTPException
 from sqlalchemy.ext.asyncio import AsyncEngine, create_async_engine
@@ -12,6 +13,19 @@ from sqlmodel.ext.asyncio.session import AsyncSession
 
 from routstr.auth import validate_bearer_key
 from routstr.core.db import ApiKey
+from routstr.wallet import MintConnectionError
+
+
+def _value_error_wrapping_transport() -> ValueError:
+    """A ValueError re-raised ``from`` a real httpx transport error, mirroring
+    ``wallet.py`` wrapping a connection failure. The sanitized classifier must
+    still see the mint-unreachable signal through the ``__cause__`` chain."""
+    try:
+        raise httpx.ConnectError("All connection attempts failed")
+    except httpx.ConnectError as exc:
+        err = ValueError("Failed to estimate fees: connection failed")
+        err.__cause__ = exc
+        return err
 
 
 def _make_engine() -> AsyncEngine:
@@ -57,3 +71,223 @@ async def test_failed_first_cashu_redemption_rolls_back_empty_api_key(
             await validate_bearer_key(token, session)
 
     assert await session.get(ApiKey, hashed_key) is None
+
+
+@pytest.mark.parametrize(
+    ("error", "expected_status", "expected_type", "expected_message", "expected_code"),
+    [
+        (
+            ValueError("Mint Error: Token already spent. (Code: 11001)"),
+            400,
+            "token_already_spent",
+            "Cashu token already spent",
+            "cashu_token_already_spent",
+        ),
+        (
+            # Raw httpx transport error propagated unwrapped from cashu.
+            httpx.ConnectError("All connection attempts failed"),
+            503,
+            "mint_unreachable",
+            "Cashu mint is unreachable",
+            "cashu_mint_unreachable",
+        ),
+        (
+            # Typed error raised by wallet.py at a wrap site.
+            MintConnectionError("connect to http://mint:3338 refused"),
+            503,
+            "mint_unreachable",
+            "Cashu mint is unreachable",
+            "cashu_mint_unreachable",
+        ),
+        (
+            # ValueError wrapping the httpx error in its __cause__ chain.
+            _value_error_wrapping_transport(),
+            503,
+            "mint_unreachable",
+            "Cashu mint is unreachable",
+            "cashu_mint_unreachable",
+        ),
+        (
+            # asyncio.TimeoutError is builtin TimeoutError on 3.11+.
+            TimeoutError("Timed out connecting to Cashu mint http://mint:3338"),
+            503,
+            "mint_unreachable",
+            "Cashu mint is unreachable",
+            "cashu_mint_unreachable",
+        ),
+        (
+            ValueError(
+                "Token amount (5 sat) is insufficient to cover melt fees. "
+                "Needed: 7 sat (amount: 5 + fee: 1 + input_fees: 1)"
+            ),
+            422,
+            "mint_error",
+            "Token value is too small to cover swap fees",
+            "cashu_token_swap_fees_exceed_amount",
+        ),
+        (
+            ValueError(
+                "Failed to estimate fees: Fees (7 sat) exceed token amount (5 sat)"
+            ),
+            422,
+            "mint_error",
+            "Token value is too small to cover swap fees",
+            "cashu_token_swap_fees_exceed_amount",
+        ),
+        (
+            ValueError(
+                "Failed to melt token from foreign mint http://foreign:3338: boom"
+            ),
+            422,
+            "mint_error",
+            "Failed to swap token from foreign mint",
+            "cashu_foreign_mint_swap_failed",
+        ),
+        (
+            ValueError("could not decode token"),
+            400,
+            "invalid_token",
+            "Invalid Cashu token",
+            "invalid_cashu_token",
+        ),
+        (
+            ValueError("some unexpected wallet condition"),
+            400,
+            "cashu_error",
+            "Failed to redeem Cashu token",
+            "cashu_token_redemption_failed",
+        ),
+        (
+            ValueError("Redeemed token amount must be positive, got 0 msats"),
+            400,
+            "cashu_error",
+            "Failed to redeem Cashu token: token yielded no value",
+            "cashu_token_zero_value",
+        ),
+    ],
+)
+@pytest.mark.asyncio
+async def test_redemption_failure_returns_sanitized_error(
+    session: AsyncSession,
+    error: Exception,
+    expected_status: int,
+    expected_type: str,
+    expected_message: str,
+    expected_code: str,
+) -> None:
+    """Redemption failures reuse the shared X-Cashu taxonomy (carried in
+    ``type``), expose stable sanitized messages and granular machine-readable
+    ``code`` values, and leave no orphan ApiKey row."""
+    token = "cashuAredemption_fails_with_specific_error"
+    hashed_key = hashlib.sha256(token.encode()).hexdigest()
+    token_obj = SimpleNamespace(mint="http://mint:3338", unit="sat")
+
+    from routstr.core.settings import settings
+
+    with (
+        patch.object(settings, "cashu_mints", ["http://mint:3338"]),
+        patch("routstr.auth.deserialize_token_from_string", return_value=token_obj),
+        patch(
+            "routstr.auth.credit_balance",
+            new=AsyncMock(side_effect=error),
+        ),
+    ):
+        with pytest.raises(HTTPException) as exc_info:
+            await validate_bearer_key(token, session)
+
+    assert exc_info.value.status_code == expected_status
+    detail = cast(dict[str, dict[str, object]], exc_info.value.detail)
+    error_detail = detail["error"]
+    assert error_detail["type"] == expected_type
+    assert error_detail["code"] == expected_code
+    assert error_detail["message"] == expected_message
+    assert str(error) not in cast(str, error_detail["message"])
+    assert await session.get(ApiKey, hashed_key) is None
+
+
+@pytest.mark.asyncio
+async def test_unexpected_redemption_error_returns_internal_error(
+    session: AsyncSession,
+) -> None:
+    """Unexpected (non-wallet) failures surface as generic 500s without
+    leaking internal details, instead of masquerading as token errors."""
+    token = "cashuAredemption_fails_with_internal_error"
+    hashed_key = hashlib.sha256(token.encode()).hexdigest()
+    token_obj = SimpleNamespace(mint="http://mint:3338", unit="sat")
+
+    from routstr.core.settings import settings
+
+    with (
+        patch.object(settings, "cashu_mints", ["http://mint:3338"]),
+        patch("routstr.auth.deserialize_token_from_string", return_value=token_obj),
+        patch(
+            "routstr.auth.credit_balance",
+            new=AsyncMock(side_effect=RuntimeError("db exploded at /var/lib/secret")),
+        ),
+    ):
+        with pytest.raises(HTTPException) as exc_info:
+            await validate_bearer_key(token, session)
+
+    assert exc_info.value.status_code == 500
+    detail = cast(dict[str, dict[str, str]], exc_info.value.detail)
+    error_detail = detail["error"]
+    assert error_detail["code"] == "internal_error"
+    assert "/var/lib/secret" not in error_detail["message"]
+    assert await session.get(ApiKey, hashed_key) is None
+
+
+@pytest.mark.asyncio
+async def test_internal_error_with_invalid_keyword_does_not_masquerade(
+    session: AsyncSession,
+) -> None:
+    """A non-wallet fault whose text merely contains "invalid" (but not
+    "token") must fall through to a generic 500, not a 401 token error.
+
+    Guards the anchored `"invalid"/"decode"` + `"token"` gate against stdlib/
+    driver strings like "Invalid isoformat string" leaking as token errors."""
+    token = "cashuAinternal_fault_mentions_invalid"
+    hashed_key = hashlib.sha256(token.encode()).hexdigest()
+    token_obj = SimpleNamespace(mint="http://mint:3338", unit="sat")
+
+    from routstr.core.settings import settings
+
+    with (
+        patch.object(settings, "cashu_mints", ["http://mint:3338"]),
+        patch("routstr.auth.deserialize_token_from_string", return_value=token_obj),
+        patch(
+            "routstr.auth.credit_balance",
+            new=AsyncMock(
+                side_effect=RuntimeError("Invalid isoformat string: '2020-13-99'")
+            ),
+        ),
+    ):
+        with pytest.raises(HTTPException) as exc_info:
+            await validate_bearer_key(token, session)
+
+    assert exc_info.value.status_code == 500
+    detail = cast(dict[str, dict[str, str]], exc_info.value.detail)
+    assert detail["error"]["code"] == "internal_error"
+    assert await session.get(ApiKey, hashed_key) is None
+
+
+@pytest.mark.asyncio
+async def test_malformed_cashu_token_returns_400_invalid_token(
+    session: AsyncSession,
+) -> None:
+    """A malformed 'cashu...' token that fails to decode maps to 400
+    invalid_cashu_token (shared taxonomy), not the generic 401 invalid_api_key."""
+    token = "cashuAthis_is_not_a_valid_token"
+
+    with patch(
+        "routstr.auth.deserialize_token_from_string",
+        side_effect=ValueError("unable to decode token: bad base64"),
+    ):
+        with pytest.raises(HTTPException) as exc_info:
+            await validate_bearer_key(token, session)
+
+    assert exc_info.value.status_code == 400
+    detail = cast(dict[str, dict[str, str]], exc_info.value.detail)
+    assert detail["error"]["type"] == "invalid_token"
+    assert detail["error"]["code"] == "invalid_cashu_token"
+    # Raw decoder text must not leak to the client.
+    assert "base64" not in detail["error"]["message"]

tests/unit/test_auto_topup.py

@@ -0,0 +1,143 @@
+import json
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+from routstr.core.db import CashuTransaction
+from routstr.upstream.auto_topup import _check_and_topup
+
+
+def _row() -> MagicMock:
+    row = MagicMock()
+    row.id = "provider-1"
+    row.base_url = "https://provider.test"
+    row.api_key = "secret"
+    row.provider_settings = json.dumps(
+        {
+            "auto_topup": True,
+            "topup_threshold": 100,
+            "topup_amount_limit": 50,
+            "topup_mint_url": "https://mint.test",
+        }
+    )
+    return row
+
+
+class _Session:
+    def __init__(self, transaction: CashuTransaction) -> None:
+        self.transaction = transaction
+        self.commit = AsyncMock()
+
+    async def __aenter__(self) -> "_Session":
+        return self
+
+    async def __aexit__(self, *args: object) -> None:
+        return None
+
+    async def exec(self, query: object) -> MagicMock:
+        result = MagicMock()
+        result.first.return_value = self.transaction
+        return result
+
+    def add(self, transaction: CashuTransaction) -> None:
+        self.transaction = transaction
+
+
+@pytest.mark.asyncio
+async def test_auto_topup_persists_before_sending_and_marks_success_collected() -> None:
+    provider = MagicMock()
+    provider.get_balance = AsyncMock(return_value=0)
+    provider.topup = AsyncMock(return_value={"balance": 50})
+    transaction = CashuTransaction(
+        token="cashu-token", amount=50, unit="sat", source="auto_topup"
+    )
+    session = _Session(transaction)
+
+    with (
+        patch(
+            "routstr.upstream.auto_topup.RoutstrUpstreamProvider.from_db_row",
+            return_value=provider,
+        ),
+        patch(
+            "routstr.upstream.auto_topup.send_token",
+            AsyncMock(return_value="cashu-token"),
+        ),
+        patch(
+            "routstr.upstream.auto_topup.store_cashu_transaction",
+            AsyncMock(return_value=True),
+        ) as store,
+        patch("routstr.upstream.auto_topup.create_session", return_value=session),
+    ):
+        await _check_and_topup(_row())
+
+    store.assert_awaited_once_with(
+        token="cashu-token",
+        amount=50,
+        unit="sat",
+        mint_url="https://mint.test",
+        typ="out",
+        collected=False,
+        source="auto_topup",
+    )
+    provider.topup.assert_awaited_once_with("cashu-token")
+    assert transaction.collected is True
+    session.commit.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("outcome", [{"error": "rejected"}, RuntimeError("network")])
+async def test_auto_topup_failure_leaves_persisted_token_uncollected(
+    outcome: object,
+) -> None:
+    provider = MagicMock()
+    provider.get_balance = AsyncMock(return_value=0)
+    provider.topup = AsyncMock(
+        side_effect=outcome if isinstance(outcome, Exception) else None,
+        return_value=outcome,
+    )
+
+    with (
+        patch(
+            "routstr.upstream.auto_topup.RoutstrUpstreamProvider.from_db_row",
+            return_value=provider,
+        ),
+        patch(
+            "routstr.upstream.auto_topup.send_token",
+            AsyncMock(return_value="cashu-token"),
+        ),
+        patch(
+            "routstr.upstream.auto_topup.store_cashu_transaction",
+            AsyncMock(return_value=True),
+        ),
+        patch("routstr.upstream.auto_topup.create_session") as create_session,
+    ):
+        if isinstance(outcome, Exception):
+            with pytest.raises(RuntimeError):
+                await _check_and_topup(_row())
+        else:
+            await _check_and_topup(_row())
+
+    create_session.assert_not_called()
+
+
+@pytest.mark.asyncio
+async def test_auto_topup_does_not_send_untracked_token() -> None:
+    provider = MagicMock()
+    provider.get_balance = AsyncMock(return_value=0)
+    provider.topup = AsyncMock()
+    with (
+        patch(
+            "routstr.upstream.auto_topup.RoutstrUpstreamProvider.from_db_row",
+            return_value=provider,
+        ),
+        patch(
+            "routstr.upstream.auto_topup.send_token",
+            AsyncMock(return_value="cashu-token"),
+        ),
+        patch(
+            "routstr.upstream.auto_topup.store_cashu_transaction",
+            AsyncMock(return_value=False),
+        ),
+    ):
+        await _check_and_topup(_row())
+    provider.topup.assert_not_awaited()

tests/unit/test_balance.py

@@ -1,12 +1,13 @@
 import json
 from unittest.mock import AsyncMock, MagicMock, patch
 
+import httpx
 import pytest
 from fastapi.responses import JSONResponse
 
-from routstr.balance import refund_wallet_endpoint
+from routstr.balance import refund_wallet_endpoint, topup_wallet_endpoint
 from routstr.core.db import ApiKey, CashuTransaction
-from routstr.wallet import credit_balance
+from routstr.wallet import MintConnectionError, credit_balance
 
 
 def _make_cashu_tx(
@@ -103,6 +104,64 @@ async def test_refund_x_cashu_not_found_raises_404() -> None:
     assert exc_info.value.status_code == 404
 
 
+@pytest.mark.asyncio
+async def test_refund_x_cashu_pending_raises_425() -> None:
+    """in row exists with a request_id but out row not yet created → 425.
+
+    This is the race condition where /v1/wallet/refund is polled while the
+    upstream request is still in flight. The endpoint must signal "retry"
+    rather than a permanent 404.
+    """
+    from fastapi import HTTPException
+
+    x_cashu_token = "cashuApending_token"
+    in_tx = _make_cashu_tx(
+        token=x_cashu_token, amount=0, unit="msat", type="in", request_id="req-pending"
+    )
+
+    session = MagicMock()
+    session.exec = AsyncMock(side_effect=[_exec_result(in_tx), _exec_result(None)])
+    session.add = MagicMock()
+    session.commit = AsyncMock()
+
+    with pytest.raises(HTTPException) as exc_info:
+        await refund_wallet_endpoint(
+            authorization="Bearer sk-somekey",
+            x_cashu=x_cashu_token,
+            session=session,
+        )
+
+    assert exc_info.value.status_code == 425
+    assert exc_info.value.headers == {"Retry-After": "2"}
+
+
+@pytest.mark.asyncio
+async def test_refund_x_cashu_in_tx_without_request_id_raises_404() -> None:
+    """in row exists but has no request_id (cannot link to a refund) → 404.
+
+    This is a genuine "no refund will ever exist" case, distinct from the
+    pending 425 path.
+    """
+    from fastapi import HTTPException
+
+    x_cashu_token = "cashuAnoreqid_token"
+    in_tx = _make_cashu_tx(
+        token=x_cashu_token, amount=0, unit="msat", type="in", request_id=None
+    )
+
+    session = MagicMock()
+    session.exec = AsyncMock(side_effect=[_exec_result(in_tx)])
+
+    with pytest.raises(HTTPException) as exc_info:
+        await refund_wallet_endpoint(
+            authorization="Bearer sk-somekey",
+            x_cashu=x_cashu_token,
+            session=session,
+        )
+
+    assert exc_info.value.status_code == 404
+
+
 @pytest.mark.asyncio
 async def test_refund_x_cashu_swept_raises_410() -> None:
     from fastapi import HTTPException
@@ -339,7 +398,10 @@ async def test_apikey_refund_restores_balance_on_mint_failure() -> None:
 
     with (
         patch("routstr.balance.get_billing_key", AsyncMock(return_value=key)),
-        patch("routstr.balance.send_token", AsyncMock(side_effect=Exception("mint down"))),
+        patch(
+            "routstr.balance.send_token",
+            AsyncMock(side_effect=MintConnectionError("raw mint outage detail")),
+        ),
         patch("routstr.balance.store_cashu_transaction", AsyncMock()),
         patch("routstr.balance._refund_cache_get", AsyncMock(return_value=None)),
         patch("routstr.balance._refund_cache_set", AsyncMock()),
@@ -353,10 +415,46 @@ async def test_apikey_refund_restores_balance_on_mint_failure() -> None:
             )
 
     assert exc_info.value.status_code == 503
+    assert exc_info.value.detail == "Mint service unavailable"
+    assert "raw mint outage detail" not in exc_info.value.detail
     # Verify two exec calls: debit + restore
     assert session.exec.await_count == 2
 
 
+@pytest.mark.asyncio
+async def test_apikey_refund_generic_failure_is_sanitized_500() -> None:
+    """Unexpected send-side failures restore balance without leaking exception text."""
+    from fastapi import HTTPException
+
+    key = _make_api_key(balance=5000, refund_currency="sat")
+    raw_error = "database secret token raw-mint-response"
+
+    session = MagicMock()
+    session.get = AsyncMock(return_value=key)
+    session.exec = AsyncMock(side_effect=[_update_result(1), _update_result(1)])
+    session.commit = AsyncMock()
+
+    with (
+        patch("routstr.balance.get_billing_key", AsyncMock(return_value=key)),
+        patch("routstr.balance.send_token", AsyncMock(side_effect=RuntimeError(raw_error))),
+        patch("routstr.balance.store_cashu_transaction", AsyncMock()),
+        patch("routstr.balance._refund_cache_get", AsyncMock(return_value=None)),
+        patch("routstr.balance._refund_cache_set", AsyncMock()),
+        patch("routstr.balance.logger"),
+    ):
+        with pytest.raises(HTTPException) as exc_info:
+            await refund_wallet_endpoint(
+                authorization="Bearer sk-testhash",
+                x_cashu=None,
+                session=session,
+            )
+
+    assert exc_info.value.status_code == 500
+    assert exc_info.value.detail == "Refund failed"
+    assert raw_error not in exc_info.value.detail
+    assert session.exec.await_count == 2
+
+
 # ---------------------------------------------------------------------------
 # no-create guarantee: fresh Cashu/unknown sk- tokens must not create API keys
 # ---------------------------------------------------------------------------
@@ -401,3 +499,190 @@ async def test_refund_unknown_sk_bearer_returns_401() -> None:
 
     assert exc_info.value.status_code == 401
     session.get.assert_awaited_once()
+
+
+# --- Topup redemption error taxonomy (POST /v1/wallet/topup) ------------------
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "error",
+    [
+        httpx.ConnectError("All connection attempts failed"),
+        MintConnectionError("connect to mint refused"),
+        TimeoutError("timed out connecting to mint"),
+    ],
+)
+async def test_topup_mint_unreachable_returns_503(error: Exception) -> None:
+    """A down mint must surface 503 (retryable), not 400 or 500 — the token is
+    fine, so the client should retry once the mint recovers."""
+    from fastapi import HTTPException
+
+    key = _make_api_key(balance=1000)
+    session = MagicMock()
+
+    with (
+        patch("routstr.balance.get_billing_key", AsyncMock(return_value=key)),
+        patch("routstr.balance.credit_balance", AsyncMock(side_effect=error)),
+    ):
+        with pytest.raises(HTTPException) as exc_info:
+            await topup_wallet_endpoint(
+                cashu_token="cashuAtoken", key=key, session=session
+            )
+
+    assert exc_info.value.status_code == 503
+    assert exc_info.value.detail == "Cashu mint is unreachable"
+
+
+@pytest.mark.asyncio
+async def test_topup_already_spent_still_returns_400() -> None:
+    """Regression: the mint-unreachable short-circuit must not swallow the
+    existing ValueError substring buckets."""
+    from fastapi import HTTPException
+
+    key = _make_api_key(balance=1000)
+    session = MagicMock()
+
+    with (
+        patch("routstr.balance.get_billing_key", AsyncMock(return_value=key)),
+        patch(
+            "routstr.balance.credit_balance",
+            AsyncMock(side_effect=ValueError("Token already spent")),
+        ),
+    ):
+        with pytest.raises(HTTPException) as exc_info:
+            await topup_wallet_endpoint(
+                cashu_token="cashuAtoken", key=key, session=session
+            )
+
+    assert exc_info.value.status_code == 400
+    assert exc_info.value.detail == "Cashu token already spent"
+
+
+@pytest.mark.asyncio
+async def test_topup_zero_value_returns_400_zero_value_message() -> None:
+    """A dust/zero redemption maps to the documented zero-value message, not the
+    generic redemption-failed one."""
+    from fastapi import HTTPException
+
+    key = _make_api_key(balance=1000)
+    session = MagicMock()
+
+    with (
+        patch("routstr.balance.get_billing_key", AsyncMock(return_value=key)),
+        patch(
+            "routstr.balance.credit_balance",
+            AsyncMock(
+                side_effect=ValueError("Redeemed token amount must be positive, got 0 msats")
+            ),
+        ),
+    ):
+        with pytest.raises(HTTPException) as exc_info:
+            await topup_wallet_endpoint(
+                cashu_token="cashuAtoken", key=key, session=session
+            )
+
+    assert exc_info.value.status_code == 400
+    assert exc_info.value.detail == "Failed to redeem Cashu token: token yielded no value"
+
+
+@pytest.mark.asyncio
+async def test_topup_token_consumed_returns_500() -> None:
+    """A post-redemption crediting failure (token spent) is a non-retryable 500,
+    not a 4xx that invites a retry."""
+    from fastapi import HTTPException
+
+    from routstr.wallet import TokenConsumedError
+
+    key = _make_api_key(balance=1000)
+    session = MagicMock()
+
+    with (
+        patch("routstr.balance.get_billing_key", AsyncMock(return_value=key)),
+        patch(
+            "routstr.balance.credit_balance",
+            AsyncMock(side_effect=TokenConsumedError("credit failed")),
+        ),
+    ):
+        with pytest.raises(HTTPException) as exc_info:
+            await topup_wallet_endpoint(
+                cashu_token="cashuAtoken", key=key, session=session
+            )
+
+    assert exc_info.value.status_code == 500
+    assert exc_info.value.detail == (
+        "Token was redeemed but could not be credited; do not retry"
+    )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("error", "expected_status", "expected_detail"),
+    [
+        (
+            ValueError(
+                "Failed to estimate fees: Fees (7 sat) exceed token amount (5 sat)"
+            ),
+            422,
+            "Token value is too small to cover swap fees",
+        ),
+        (
+            ValueError(
+                "Token amount (5 sat) is insufficient to cover melt fees."
+            ),
+            422,
+            "Token value is too small to cover swap fees",
+        ),
+        (
+            ValueError("Failed to melt token from foreign mint http://m: boom"),
+            422,
+            "Failed to swap token from foreign mint",
+        ),
+    ],
+)
+async def test_topup_fee_and_swap_failures_return_422(
+    error: Exception, expected_status: int, expected_detail: str
+) -> None:
+    """Fee/swap failures map to 422 (shared taxonomy), matching the bearer and
+    X-Cashu paths — previously top-up flattened these to 400."""
+    from fastapi import HTTPException
+
+    key = _make_api_key(balance=1000)
+    session = MagicMock()
+
+    with (
+        patch("routstr.balance.get_billing_key", AsyncMock(return_value=key)),
+        patch("routstr.balance.credit_balance", AsyncMock(side_effect=error)),
+    ):
+        with pytest.raises(HTTPException) as exc_info:
+            await topup_wallet_endpoint(
+                cashu_token="cashuAtoken", key=key, session=session
+            )
+
+    assert exc_info.value.status_code == expected_status
+    assert exc_info.value.detail == expected_detail
+
+
+@pytest.mark.asyncio
+async def test_topup_unexpected_non_valueerror_returns_500() -> None:
+    """A non-ValueError, non-transport fault is an internal error (500), not a
+    sanitized 400 — the merged except must preserve this."""
+    from fastapi import HTTPException
+
+    key = _make_api_key(balance=1000)
+    session = MagicMock()
+
+    with (
+        patch("routstr.balance.get_billing_key", AsyncMock(return_value=key)),
+        patch(
+            "routstr.balance.credit_balance",
+            AsyncMock(side_effect=RuntimeError("db exploded")),
+        ),
+    ):
+        with pytest.raises(HTTPException) as exc_info:
+            await topup_wallet_endpoint(
+                cashu_token="cashuAtoken", key=key, session=session
+            )
+
+    assert exc_info.value.status_code == 500
+    assert exc_info.value.detail == "Internal server error"

tests/unit/test_cache_breakpoints.py

@@ -0,0 +1,189 @@
+"""Tests for Anthropic cache-breakpoint injection on forwarded requests.
+
+Anthropic prompt caching is explicit; a client that doesn't recognise a routstr
+URL as Anthropic-backed never sends ``cache_control`` markers, so caching never
+engages over routstr. ``prepare_request_body`` must stamp the standard
+breakpoints for Anthropic-family models while always deferring to client-set
+markers and never touching automatic-cache providers.
+"""
+
+import json
+import os
+
+import pytest
+
+os.environ.setdefault("UPSTREAM_BASE_URL", "http://test")
+os.environ.setdefault("UPSTREAM_API_KEY", "test")
+os.environ.setdefault("LIGHTNING_ADDRESS", "test@stm.to")
+
+from routstr.upstream import GenericUpstreamProvider
+from routstr.upstream.cache_breakpoints import (
+    body_has_cache_control,
+    inject_anthropic_cache_breakpoints,
+    is_explicit_cache_model,
+)
+
+
+def _chat_body() -> dict:
+    return {
+        "model": "anthropic/claude-sonnet-4.5",
+        "stream": True,
+        "messages": [
+            {"role": "system", "content": "You are concise."},
+            {"role": "user", "content": "Hello"},
+        ],
+        "tools": [
+            {"type": "function", "function": {"name": "a"}},
+            {"type": "function", "function": {"name": "b"}},
+        ],
+    }
+
+
+# ---------------------------------------------------------------------------
+# Model detection
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.parametrize(
+    "model_id,expected",
+    [
+        ("anthropic/claude-sonnet-4.5", True),
+        ("claude-haiku-4-5-20251001", True),
+        # Alibaba explicit-cache models share Anthropic's wire format
+        ("qwen/qwen3-max", True),
+        ("qwen/qwen3-coder-plus", True),
+        ("deepseek/deepseek-v3.2", True),
+        # Automatic-cache providers need no markers
+        ("openai/gpt-4o", False),
+        ("google/gemini-2.5-flash", False),
+        ("deepseek/deepseek-chat", False),
+        ("qwen/qwen3.5-plus-02-15", False),  # snapshot, no explicit caching
+        (None, False),
+    ],
+)
+def test_is_explicit_cache_model(model_id: str | None, expected: bool) -> None:
+    assert is_explicit_cache_model(model_id) is expected
+
+
+def test_is_explicit_cache_model_uses_fallbacks() -> None:
+    # routstr id is opaque but a forwarded/canonical alias reveals the family.
+    assert is_explicit_cache_model("model-xyz", None, "anthropic/claude-opus-4.1")
+
+
+# ---------------------------------------------------------------------------
+# Breakpoint placement
+# ---------------------------------------------------------------------------
+
+
+def test_injects_three_breakpoints() -> None:
+    data = _chat_body()
+    assert inject_anthropic_cache_breakpoints(data) is True
+
+    # system prompt promoted to array form with a marker
+    system = data["messages"][0]["content"]
+    assert system == [
+        {
+            "type": "text",
+            "text": "You are concise.",
+            "cache_control": {"type": "ephemeral"},
+        }
+    ]
+    # last tool marked
+    assert data["tools"][-1]["cache_control"] == {"type": "ephemeral"}
+    assert "cache_control" not in data["tools"][0]
+    # last user message marked
+    user = data["messages"][1]["content"]
+    assert user[-1]["cache_control"] == {"type": "ephemeral"}
+
+
+def test_defers_to_client_supplied_cache_control() -> None:
+    data = _chat_body()
+    data["messages"][1]["content"] = [
+        {"type": "text", "text": "Hello", "cache_control": {"type": "ephemeral"}}
+    ]
+    assert body_has_cache_control(data) is True
+    # No additional stamping when the client already controls caching.
+    assert inject_anthropic_cache_breakpoints(data) is False
+    assert "cache_control" not in data["tools"][-1]
+
+
+def test_marks_last_text_part_of_array_content() -> None:
+    data = _chat_body()
+    data["messages"][1]["content"] = [
+        {"type": "text", "text": "first"},
+        {"type": "image_url", "image_url": {"url": "x"}},
+        {"type": "text", "text": "last"},
+    ]
+    inject_anthropic_cache_breakpoints(data)
+    parts = data["messages"][1]["content"]
+    assert parts[2]["cache_control"] == {"type": "ephemeral"}
+    assert "cache_control" not in parts[0]
+
+
+def test_noop_without_messages() -> None:
+    assert inject_anthropic_cache_breakpoints({"prompt": "x"}) is False
+
+
+# ---------------------------------------------------------------------------
+# prepare_request_body integration
+# ---------------------------------------------------------------------------
+
+
+def _model(model_id: str):  # type: ignore[no-untyped-def]
+    from routstr.payment.models import Architecture, Model, Pricing
+
+    return Model(
+        id=model_id,
+        name=model_id,
+        created=0,
+        description="",
+        context_length=200000,
+        architecture=Architecture(
+            modality="text->text",
+            input_modalities=["text"],
+            output_modalities=["text"],
+            tokenizer="Claude",
+            instruct_type=None,
+        ),
+        pricing=Pricing(prompt=0.0, completion=0.0),
+    )
+
+
+def _openrouter_provider() -> "GenericUpstreamProvider":
+    # OpenRouter endpoint via the generic provider — recognised by base URL.
+    return GenericUpstreamProvider(base_url="https://openrouter.ai/api/v1")
+
+
+@pytest.mark.parametrize(
+    "model_id", ["anthropic/claude-sonnet-4.5", "qwen/qwen3-max", "deepseek/deepseek-v3.2"]
+)
+def test_prepare_request_body_injects_for_explicit_models(model_id: str) -> None:
+    provider = _openrouter_provider()
+    body = json.dumps(_chat_body()).encode()
+    out = provider.prepare_request_body(body, _model(model_id))
+    assert out is not None
+    data = json.loads(out)
+    assert body_has_cache_control(data) is True
+    assert data["tools"][-1]["cache_control"] == {"type": "ephemeral"}
+
+
+def test_prepare_request_body_skips_for_automatic_provider_model() -> None:
+    provider = _openrouter_provider()
+    body = json.dumps(_chat_body()).encode()
+    out = provider.prepare_request_body(body, _model("openai/gpt-4o"))
+    assert out is not None
+    data = json.loads(out)
+    assert body_has_cache_control(data) is False
+
+
+def test_prepare_request_body_skips_when_upstream_rejects_markers() -> None:
+    # Claude id but a non-OpenRouter/Anthropic upstream → must NOT inject,
+    # since the markers could be rejected by an upstream that doesn't accept them.
+    from routstr.upstream import GenericUpstreamProvider
+
+    provider = GenericUpstreamProvider(base_url="https://some-gateway.example/v1")
+    body = json.dumps(_chat_body()).encode()
+    out = provider.prepare_request_body(body, _model("anthropic/claude-sonnet-4.5"))
+    assert out is not None
+    data = json.loads(out)
+    assert body_has_cache_control(data) is False

tests/unit/test_cache_pricing.py

@@ -0,0 +1,347 @@
+"""Tests for cache-aware pricing of cached input tokens.
+
+Specifies two things:
+
+1. ``backfill_cache_pricing`` — when the OpenRouter model feed omits cache
+   rates (it does for most DeepSeek models and e.g. openai/gpt-4o), they are
+   filled from litellm's bundled cost map instead of silently billing cache
+   reads at the full input rate. Existing OpenRouter values are never
+   overwritten, and provider fees apply to backfilled rates like any other.
+2. ``calculate_cost`` — cached tokens are billed at the cache rates from the
+   model's sats_pricing; the full input rate remains only as the documented
+   last resort when no cache rate could be resolved anywhere.
+"""
+
+import os
+from unittest.mock import Mock, patch
+
+import litellm
+import pytest
+
+os.environ.setdefault("UPSTREAM_BASE_URL", "http://test")
+os.environ.setdefault("UPSTREAM_API_KEY", "test")
+os.environ.setdefault("LIGHTNING_ADDRESS", "test@stm.to")
+
+from routstr.core.settings import settings
+from routstr.payment.cost_calculation import CostData, calculate_cost
+from routstr.payment.models import (
+    Architecture,
+    Model,
+    Pricing,
+    backfill_cache_pricing,
+)
+from routstr.upstream import GenericUpstreamProvider
+
+
+def _make_model(model_id: str, pricing: Pricing) -> Model:
+    return Model(
+        id=model_id,
+        name=model_id,
+        created=0,
+        description="",
+        context_length=64000,
+        architecture=Architecture(
+            modality="text->text",
+            input_modalities=["text"],
+            output_modalities=["text"],
+            tokenizer="Other",
+            instruct_type=None,
+        ),
+        pricing=pricing,
+    )
+
+
+# ============================================================================
+# backfill_cache_pricing — litellm as fallback source for missing cache rates
+# ============================================================================
+
+
+def test_backfill_deepseek_cache_read_from_litellm() -> None:
+    """deepseek/deepseek-chat has no input_cache_read on OpenRouter; litellm
+    knows the real rate (10x cheaper than input)."""
+    pricing = Pricing(prompt=2.8e-07, completion=4.2e-07)
+
+    result = backfill_cache_pricing("deepseek/deepseek-chat", pricing)
+
+    expected = litellm.model_cost["deepseek/deepseek-chat"][
+        "cache_read_input_token_cost"
+    ]
+    assert result.input_cache_read == expected
+    assert result.input_cache_read < pricing.prompt  # sanity: it's a discount
+
+
+def test_backfill_strips_vendor_prefix_for_litellm_lookup() -> None:
+    """OpenRouter ids are vendor-prefixed (openai/gpt-4o); litellm keys most
+    non-DeepSeek models without the prefix (gpt-4o)."""
+    pricing = Pricing(prompt=2.5e-06, completion=1e-05)
+
+    result = backfill_cache_pricing("openai/gpt-4o", pricing)
+
+    expected = litellm.model_cost["gpt-4o"]["cache_read_input_token_cost"]
+    assert result.input_cache_read == expected
+
+
+def test_backfill_case_insensitive_lookup() -> None:
+    """A generic upstream may report a mixed-case id
+    (deepseek-ai/DeepSeek-V4-Flash); litellm keys are lowercase. The
+    case-insensitive fallback still resolves the cache rate."""
+    pricing = Pricing(prompt=1.4e-07, completion=2.8e-07)
+
+    result = backfill_cache_pricing("deepseek-ai/DeepSeek-V4-Flash", pricing)
+
+    expected = litellm.model_cost["deepseek-v4-flash"][
+        "cache_read_input_token_cost"
+    ]
+    assert result.input_cache_read == expected
+    assert result.input_cache_read < pricing.prompt  # sanity: it's a discount
+
+
+def test_backfill_fills_cache_write_rate() -> None:
+    """Anthropic cache writes cost more than input (1.25x); billing them at
+    the input rate undercharges. litellm carries the write rate."""
+    pricing = Pricing(prompt=3e-06, completion=1.5e-05)
+
+    result = backfill_cache_pricing("anthropic/claude-sonnet-4-5", pricing)
+
+    expected = litellm.model_cost["claude-sonnet-4-5"][
+        "cache_creation_input_token_cost"
+    ]
+    assert result.input_cache_write == expected
+    assert result.input_cache_write > pricing.prompt  # sanity: write premium
+
+
+def test_backfill_never_overwrites_openrouter_rates() -> None:
+    """When OpenRouter provides a cache rate, it is authoritative."""
+    pricing = Pricing(
+        prompt=2.1e-07, completion=7.9e-07, input_cache_read=1.3e-07
+    )
+
+    result = backfill_cache_pricing("deepseek/deepseek-chat", pricing)
+
+    assert result.input_cache_read == 1.3e-07
+
+
+def test_backfill_unknown_model_unchanged() -> None:
+    """Models litellm doesn't know stay untouched (last-resort fallback to
+    the input rate happens later, at billing time)."""
+    pricing = Pricing(prompt=1e-06, completion=2e-06)
+
+    result = backfill_cache_pricing("artificial-dumbness/dumb-1", pricing)
+
+    assert result.input_cache_read == 0.0
+    assert result.input_cache_write == 0.0
+
+
+def test_provider_fee_applies_to_backfilled_cache_rates() -> None:
+    """Backfill happens before the provider fee, so cache rates carry the
+    same markup as every other price component."""
+    provider = GenericUpstreamProvider(
+        base_url="http://upstream.example", provider_fee=2.0
+    )
+    model = _make_model(
+        "deepseek/deepseek-chat", Pricing(prompt=2.8e-07, completion=4.2e-07)
+    )
+
+    adjusted = provider._apply_provider_fee_to_model(model)
+
+    litellm_read = litellm.model_cost["deepseek/deepseek-chat"][
+        "cache_read_input_token_cost"
+    ]
+    assert adjusted.pricing.input_cache_read == pytest.approx(litellm_read * 2.0)
+    assert adjusted.pricing.prompt == pytest.approx(2.8e-07 * 2.0)
+
+
+def test_row_to_model_backfills_cache_rate() -> None:
+    """The DB-override path (admin-configured providers, e.g. a generic
+    upstream) stores pricing without cache rates. ``_row_to_model`` must
+    backfill them from litellm just like ``_apply_provider_fee_to_model``,
+    otherwise cache reads bill at the full input rate."""
+    import json
+
+    from routstr.core.db import ModelRow
+    from routstr.payment.models import _row_to_model
+
+    row = ModelRow(
+        id="deepseek-v4-flash",
+        name="deepseek-v4-flash",
+        created=0,
+        description="",
+        context_length=1000000,
+        architecture=json.dumps(
+            {
+                "modality": "text",
+                "input_modalities": ["text"],
+                "output_modalities": ["text"],
+                "tokenizer": "unknown",
+                "instruct_type": None,
+            }
+        ),
+        # Stored pricing omits input_cache_read (generic provider never sets it).
+        pricing=json.dumps({"prompt": 1.4e-07, "completion": 2.8e-07}),
+        enabled=True,
+        upstream_provider_id=1,
+    )
+
+    with patch(
+        "routstr.payment.models.sats_usd_price", return_value=5.0e-5
+    ):
+        model = _row_to_model(row, apply_provider_fee=True, provider_fee=1.0)
+
+    litellm_read = litellm.model_cost["deepseek-v4-flash"][
+        "cache_read_input_token_cost"
+    ]
+    assert model.pricing.input_cache_read == pytest.approx(litellm_read)
+    assert model.pricing.input_cache_read < model.pricing.prompt  # a discount
+    assert model.sats_pricing is not None
+    assert model.sats_pricing.input_cache_read > 0
+
+
+def test_row_to_model_backfills_via_forwarded_model_id() -> None:
+    """An alias row (id != forwarded_model_id) must backfill cache rates from
+    the *forwarded* model name — the real upstream model litellm prices —
+    not the alias id, which litellm doesn't know."""
+    import json
+
+    from routstr.core.db import ModelRow
+    from routstr.payment.models import _row_to_model
+
+    row = ModelRow(
+        id="local-alias",  # litellm has no such key
+        name="local-alias",
+        created=0,
+        description="",
+        context_length=1000000,
+        architecture=json.dumps(
+            {
+                "modality": "text",
+                "input_modalities": ["text"],
+                "output_modalities": ["text"],
+                "tokenizer": "unknown",
+                "instruct_type": None,
+            }
+        ),
+        pricing=json.dumps({"prompt": 1.4e-07, "completion": 2.8e-07}),
+        enabled=True,
+        upstream_provider_id=1,
+        forwarded_model_id="deepseek-v4-flash",
+    )
+
+    with patch(
+        "routstr.payment.models.sats_usd_price", return_value=5.0e-5
+    ):
+        model = _row_to_model(row, apply_provider_fee=True, provider_fee=1.0)
+
+    litellm_read = litellm.model_cost["deepseek-v4-flash"][
+        "cache_read_input_token_cost"
+    ]
+    assert model.pricing.input_cache_read == pytest.approx(litellm_read)
+    assert model.pricing.input_cache_read < model.pricing.prompt
+
+
+# ============================================================================
+# calculate_cost — cached tokens billed at cache rates
+# ============================================================================
+
+
+@pytest.fixture(autouse=True)
+def patch_sats_usd_price() -> None:  # type: ignore[misc]
+    with patch("routstr.payment.cost_calculation.sats_usd_price", return_value=5.0e-5):
+        yield
+
+
+@pytest.fixture
+def model_pricing(monkeypatch: pytest.MonkeyPatch) -> Mock:
+    """Model-based pricing: 1 msat per input token, 2 per output token,
+    0.1 per cache-read token, 1.25 per cache-write token."""
+    monkeypatch.setattr(settings, "fixed_pricing", False)
+    model = Mock()
+    model.sats_pricing = Pricing(
+        prompt=0.001,
+        completion=0.002,
+        input_cache_read=0.0001,
+        input_cache_write=0.00125,
+    )
+    return model
+
+
+@pytest.mark.asyncio
+async def test_deepseek_cache_hits_billed_at_cache_rate(model_pricing: Mock) -> None:
+    """The reported overcharge scenario: a 10k-token prompt with 90% cache
+    hits costs 2900 msats at honest rates, not the 11000 msats that billing
+    every prompt token at the full input rate would charge."""
+    response = {
+        "model": "deepseek-chat",
+        "usage": {
+            "prompt_tokens": 10000,
+            "completion_tokens": 500,
+            "prompt_cache_hit_tokens": 9000,
+            "prompt_cache_miss_tokens": 1000,
+        },
+    }
+
+    with patch("routstr.proxy.get_model_instance", return_value=model_pricing):
+        result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    # 1000 input @ 1 msat + 9000 cache reads @ 0.1 msat + 500 output @ 2 msat.
+    # input_msats folds the cache-read cost in (1000 + 900) so a dashboard
+    # rendering I/O/T sees input + output == total; the cache portion stays
+    # visible in cache_read_msats.
+    assert result.cache_read_msats == 900
+    assert result.output_msats == 1000
+    assert result.input_msats == 1900
+    assert result.input_msats + result.output_msats == result.total_msats
+    assert result.total_msats == 2900
+
+
+@pytest.mark.asyncio
+async def test_anthropic_cache_write_billed_at_write_rate(model_pricing: Mock) -> None:
+    """Cache writes carry their premium rate (1.25x input here), instead of
+    being silently billed at the plain input rate."""
+    response = {
+        "model": "claude-sonnet-4-5",
+        "usage": {
+            "input_tokens": 300,
+            "output_tokens": 100,
+            "cache_read_input_tokens": 500,
+            "cache_creation_input_tokens": 2000,
+        },
+    }
+
+    with patch("routstr.proxy.get_model_instance", return_value=model_pricing):
+        result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    # 300 @ 1 + 500 @ 0.1 + 2000 @ 1.25 + 100 @ 2
+    assert result.cache_read_msats == 50
+    assert result.cache_creation_msats == 2500
+    assert result.total_msats == 3050
+
+
+@pytest.mark.asyncio
+async def test_missing_cache_rate_falls_back_to_input_rate(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """Documented last resort: when no cache rate could be resolved anywhere
+    (OpenRouter and litellm both silent), cache reads bill at the input rate —
+    never cheaper, never free."""
+    monkeypatch.setattr(settings, "fixed_pricing", False)
+    model = Mock()
+    model.sats_pricing = Pricing(prompt=0.001, completion=0.002)
+
+    response = {
+        "model": "dumb-1",
+        "usage": {
+            "prompt_tokens": 10000,
+            "completion_tokens": 500,
+            "prompt_cache_hit_tokens": 9000,
+            "prompt_cache_miss_tokens": 1000,
+        },
+    }
+
+    with patch("routstr.proxy.get_model_instance", return_value=model):
+        result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    # 1000 @ 1 + 9000 @ 1 (fallback) + 500 @ 2
+    assert result.total_msats == 11000

tests/unit/test_cost_calculation_caching.py

@@ -1,10 +1,11 @@
 """Tests for cache token handling in cost calculation.
 
-Covers OpenAI vs Anthropic caching formats, edge cases, and billing accuracy.
+Covers OpenAI, Anthropic and DeepSeek caching formats, dialect precedence,
+edge cases, and billing accuracy.
 """
 
 import os
-from unittest.mock import AsyncMock, patch
+from unittest.mock import patch
 
 import pytest
 
@@ -16,12 +17,6 @@ from routstr.core.settings import settings
 from routstr.payment.cost_calculation import CostData, MaxCostData, calculate_cost
 
 
-@pytest.fixture
-def mock_session() -> AsyncMock:
-    """Mock AsyncSession for cost calculation tests."""
-    return AsyncMock()
-
-
 @pytest.fixture(autouse=True)
 def mock_fixed_pricing(monkeypatch: pytest.MonkeyPatch) -> None:
     """Mock settings and price to use fixed pricing."""
@@ -41,7 +36,7 @@ def patch_sats_usd_price() -> None:  # type: ignore[misc]
 # Test 1: OpenAI Cache Format
 # ============================================================================
 @pytest.mark.asyncio
-async def test_openai_cache_subtraction(mock_session: AsyncMock) -> None:
+async def test_openai_cache_subtraction() -> None:
     """OpenAI includes cached_tokens in prompt_tokens, subtract them."""
     response = {
         "model": "gpt-4",
@@ -53,7 +48,7 @@ async def test_openai_cache_subtraction(mock_session: AsyncMock) -> None:
             }
         }
     }
-    result = await calculate_cost(response, max_cost=100000, session=mock_session)
+    result = await calculate_cost(response, max_cost=100000)
 
     assert isinstance(result, CostData)
     assert result.input_tokens == 1000  # 2000 - 1000
@@ -65,7 +60,7 @@ async def test_openai_cache_subtraction(mock_session: AsyncMock) -> None:
 # Test 2: Anthropic Cache Format
 # ============================================================================
 @pytest.mark.asyncio
-async def test_anthropic_cache_additive(mock_session: AsyncMock, mock_fixed_pricing: None) -> None:
+async def test_anthropic_cache_additive(mock_fixed_pricing: None) -> None:
     """Anthropic cache tokens are separate (additive) from input_tokens."""
     response = {
         "model": "claude-3-5-sonnet",
@@ -76,7 +71,7 @@ async def test_anthropic_cache_additive(mock_session: AsyncMock, mock_fixed_pric
             "cache_read_input_tokens": 0,
         }
     }
-    result = await calculate_cost(response, max_cost=100000, session=mock_session)
+    result = await calculate_cost(response, max_cost=100000)
 
     assert isinstance(result, CostData)
     assert result.input_tokens == 500
@@ -89,7 +84,7 @@ async def test_anthropic_cache_additive(mock_session: AsyncMock, mock_fixed_pric
 # Test 3: Invalid Cache (Edge Case)
 # ============================================================================
 @pytest.mark.asyncio
-async def test_cache_read_exceeds_prompt_tokens(mock_session: AsyncMock, mock_fixed_pricing: None) -> None:
+async def test_cache_read_exceeds_prompt_tokens(mock_fixed_pricing: None) -> None:
     """Handle buggy upstream reporting cached > prompt_tokens."""
     response = {
         "model": "gpt-4",
@@ -101,7 +96,7 @@ async def test_cache_read_exceeds_prompt_tokens(mock_session: AsyncMock, mock_fi
             }
         }
     }
-    result = await calculate_cost(response, max_cost=100000, session=mock_session)
+    result = await calculate_cost(response, max_cost=100000)
 
     # Should not go negative
     assert isinstance(result, CostData)
@@ -114,7 +109,7 @@ async def test_cache_read_exceeds_prompt_tokens(mock_session: AsyncMock, mock_fi
 # Test 4: Malformed Token Values
 # ============================================================================
 @pytest.mark.asyncio
-async def test_malformed_cache_tokens_coerce_to_zero(mock_session: AsyncMock, mock_fixed_pricing: None) -> None:
+async def test_malformed_cache_tokens_coerce_to_zero(mock_fixed_pricing: None) -> None:
     """Handle non-numeric cache token values."""
     response = {
         "model": "gpt-4",
@@ -127,7 +122,7 @@ async def test_malformed_cache_tokens_coerce_to_zero(mock_session: AsyncMock, mo
             }
         }
     }
-    result = await calculate_cost(response, max_cost=100000, session=mock_session)
+    result = await calculate_cost(response, max_cost=100000)
 
     # Both should coerce to 0
     assert isinstance(result, CostData)
@@ -139,7 +134,7 @@ async def test_malformed_cache_tokens_coerce_to_zero(mock_session: AsyncMock, mo
 # Test 5: Anthropic Cache Not Subtracted
 # ============================================================================
 @pytest.mark.asyncio
-async def test_anthropic_cache_not_subtracted(mock_session: AsyncMock, mock_fixed_pricing: None) -> None:
+async def test_anthropic_cache_not_subtracted(mock_fixed_pricing: None) -> None:
     """Anthropic cache fields should NOT be subtracted from input_tokens."""
     response = {
         "model": "claude-3-5-sonnet",
@@ -149,7 +144,7 @@ async def test_anthropic_cache_not_subtracted(mock_session: AsyncMock, mock_fixe
             "cache_read_input_tokens": 200,  # ← Additive, don't subtract
         }
     }
-    result = await calculate_cost(response, max_cost=100000, session=mock_session)
+    result = await calculate_cost(response, max_cost=100000)
 
     # Anthropic: input_tokens stays as-is
     assert isinstance(result, CostData)
@@ -161,7 +156,7 @@ async def test_anthropic_cache_not_subtracted(mock_session: AsyncMock, mock_fixe
 # Test 6: Only Cache Read, No Regular Input
 # ============================================================================
 @pytest.mark.asyncio
-async def test_only_cache_read_tokens(mock_session: AsyncMock, mock_fixed_pricing: None) -> None:
+async def test_only_cache_read_tokens(mock_fixed_pricing: None) -> None:
     """Handle response with only cache read tokens."""
     response = {
         "model": "gpt-4",
@@ -173,7 +168,7 @@ async def test_only_cache_read_tokens(mock_session: AsyncMock, mock_fixed_pricin
             }
         }
     }
-    result = await calculate_cost(response, max_cost=100000, session=mock_session)
+    result = await calculate_cost(response, max_cost=100000)
 
     assert isinstance(result, CostData)
     assert result.input_tokens == 0  # max(0, 0 - 1000)
@@ -185,7 +180,7 @@ async def test_only_cache_read_tokens(mock_session: AsyncMock, mock_fixed_pricin
 # Test 7: Only Cache Creation
 # ============================================================================
 @pytest.mark.asyncio
-async def test_only_cache_creation_tokens(mock_session: AsyncMock, mock_fixed_pricing: None) -> None:
+async def test_only_cache_creation_tokens(mock_fixed_pricing: None) -> None:
     """Handle response with only cache creation tokens (Anthropic)."""
     response = {
         "model": "claude-3-5-sonnet",
@@ -196,7 +191,7 @@ async def test_only_cache_creation_tokens(mock_session: AsyncMock, mock_fixed_pr
             "cache_read_input_tokens": 0,
         }
     }
-    result = await calculate_cost(response, max_cost=100000, session=mock_session)
+    result = await calculate_cost(response, max_cost=100000)
 
     assert isinstance(result, CostData)
     assert result.input_tokens == 500
@@ -209,7 +204,7 @@ async def test_only_cache_creation_tokens(mock_session: AsyncMock, mock_fixed_pr
 # Test 8: Both Cache Read and Creation
 # ============================================================================
 @pytest.mark.asyncio
-async def test_both_cache_read_and_creation(mock_session: AsyncMock, mock_fixed_pricing: None) -> None:
+async def test_both_cache_read_and_creation(mock_fixed_pricing: None) -> None:
     """Handle response with both cache read and creation."""
     response = {
         "model": "claude-3-5-sonnet",
@@ -220,7 +215,7 @@ async def test_both_cache_read_and_creation(mock_session: AsyncMock, mock_fixed_
             "cache_read_input_tokens": 500,
         }
     }
-    result = await calculate_cost(response, max_cost=100000, session=mock_session)
+    result = await calculate_cost(response, max_cost=100000)
 
     assert isinstance(result, CostData)
     assert result.input_tokens == 300
@@ -233,7 +228,7 @@ async def test_both_cache_read_and_creation(mock_session: AsyncMock, mock_fixed_
 # Test 9: Token Field Fallback
 # ============================================================================
 @pytest.mark.asyncio
-async def test_token_field_fallback_order(mock_session: AsyncMock, mock_fixed_pricing: None) -> None:
+async def test_token_field_fallback_order(mock_fixed_pricing: None) -> None:
     """Verify fallback order for token extraction."""
     # When prompt_tokens is not present, fall back to input_tokens
     response = {
@@ -243,7 +238,7 @@ async def test_token_field_fallback_order(mock_session: AsyncMock, mock_fixed_pr
             "completion_tokens": 50,
         }
     }
-    result = await calculate_cost(response, max_cost=100000, session=mock_session)
+    result = await calculate_cost(response, max_cost=100000)
 
     assert isinstance(result, CostData)
     assert result.input_tokens == 250
@@ -254,20 +249,21 @@ async def test_token_field_fallback_order(mock_session: AsyncMock, mock_fixed_pr
 # Test 10: Float Token Values
 # ============================================================================
 @pytest.mark.asyncio
-async def test_float_token_values_coerced_to_int(mock_session: AsyncMock, mock_fixed_pricing: None) -> None:
+async def test_float_token_values_coerced_to_int(mock_fixed_pricing: None) -> None:
     """Handle float token values by converting to int."""
     response = {
         "model": "gpt-4",
         "usage": {
             "prompt_tokens": 100.7,  # Float
             "completion_tokens": 50.3,  # Float
-            "cache_read_input_tokens": 25.9,  # Float
+            "prompt_tokens_details": {"cached_tokens": 25.9},  # Float
         }
     }
-    result = await calculate_cost(response, max_cost=100000, session=mock_session)
+    result = await calculate_cost(response, max_cost=100000)
 
     assert isinstance(result, CostData)
-    assert result.input_tokens == 100  # Floored
+    # cached_tokens are part of prompt_tokens (OpenAI dialect) → subtracted: 100 - 25
+    assert result.input_tokens == 75  # Floored
     assert result.output_tokens == 50  # Floored
     assert result.cache_read_input_tokens == 25  # Floored
 
@@ -276,7 +272,7 @@ async def test_float_token_values_coerced_to_int(mock_session: AsyncMock, mock_f
 # Test 11: Boolean Cache Tokens
 # ============================================================================
 @pytest.mark.asyncio
-async def test_boolean_cache_tokens_coerced_to_zero(mock_session: AsyncMock, mock_fixed_pricing: None) -> None:
+async def test_boolean_cache_tokens_coerced_to_zero(mock_fixed_pricing: None) -> None:
     """Handle boolean cache token values by coercing to zero."""
     response = {
         "model": "gpt-4",
@@ -286,7 +282,7 @@ async def test_boolean_cache_tokens_coerced_to_zero(mock_session: AsyncMock, moc
             "cache_read_input_tokens": True,  # Boolean
         }
     }
-    result = await calculate_cost(response, max_cost=100000, session=mock_session)
+    result = await calculate_cost(response, max_cost=100000)
 
     assert isinstance(result, CostData)
     assert result.cache_read_input_tokens == 0  # Boolean coerced to 0
@@ -297,7 +293,7 @@ async def test_boolean_cache_tokens_coerced_to_zero(mock_session: AsyncMock, moc
 # Test 12: Zero Cache Tokens
 # ============================================================================
 @pytest.mark.asyncio
-async def test_zero_cache_tokens(mock_session: AsyncMock, mock_fixed_pricing: None) -> None:
+async def test_zero_cache_tokens(mock_fixed_pricing: None) -> None:
     """Handle explicit zero cache tokens."""
     response = {
         "model": "gpt-4",
@@ -309,21 +305,344 @@ async def test_zero_cache_tokens(mock_session: AsyncMock, mock_fixed_pricing: No
             }
         }
     }
-    result = await calculate_cost(response, max_cost=100000, session=mock_session)
+    result = await calculate_cost(response, max_cost=100000)
 
     assert isinstance(result, CostData)
     assert result.cache_read_input_tokens == 0
     assert result.input_tokens == 100
 
 
+# ============================================================================
+# DeepSeek Cache Format
+# DeepSeek emits neither OpenAI's prompt_tokens_details nor Anthropic's
+# cache_read_input_tokens — only prompt_cache_hit_tokens and
+# prompt_cache_miss_tokens, with the documented guarantee
+# prompt_tokens = hit + miss. Hits are ~10x cheaper upstream, so billing
+# them as regular input is a large overcharge.
+# ============================================================================
+@pytest.mark.asyncio
+async def test_deepseek_cache_hit_tokens_extracted() -> None:
+    """DeepSeek cache hits are extracted and removed from regular input.
+
+    Payload shape verbatim from the DeepSeek API reference (usage object).
+    """
+    response = {
+        "model": "deepseek-chat",
+        "usage": {
+            "prompt_tokens": 10000,  # = hit + miss
+            "completion_tokens": 500,
+            "total_tokens": 10500,
+            "prompt_cache_hit_tokens": 9000,
+            "prompt_cache_miss_tokens": 1000,
+        },
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.input_tokens == 1000  # only the cache misses
+    assert result.cache_read_input_tokens == 9000
+    assert result.output_tokens == 500
+
+
+@pytest.mark.asyncio
+async def test_deepseek_all_tokens_cached() -> None:
+    """A fully cached DeepSeek prompt bills zero regular input tokens."""
+    response = {
+        "model": "deepseek-chat",
+        "usage": {
+            "prompt_tokens": 5000,
+            "completion_tokens": 100,
+            "prompt_cache_hit_tokens": 5000,
+            "prompt_cache_miss_tokens": 0,
+        },
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.input_tokens == 0
+    assert result.cache_read_input_tokens == 5000
+
+
+@pytest.mark.asyncio
+async def test_dialect_precedence_never_double_subtracts() -> None:
+    """If a vendor emits both OpenAI-style and DeepSeek-style cache fields for
+    the same cached tokens, they are counted once, not subtracted twice."""
+    response = {
+        "model": "deepseek-chat",
+        "usage": {
+            "prompt_tokens": 10000,
+            "completion_tokens": 500,
+            "prompt_tokens_details": {"cached_tokens": 9000},
+            "prompt_cache_hit_tokens": 9000,
+            "prompt_cache_miss_tokens": 1000,
+        },
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.input_tokens == 1000  # 10000 - 9000, applied exactly once
+    assert result.cache_read_input_tokens == 9000
+
+
+@pytest.mark.asyncio
+async def test_deepseek_malformed_hit_tokens_coerce_to_zero() -> None:
+    """Malformed DeepSeek cache fields degrade to billing all input at full
+    rate instead of crashing or going negative."""
+    response = {
+        "model": "deepseek-chat",
+        "usage": {
+            "prompt_tokens": 1000,
+            "completion_tokens": 50,
+            "prompt_cache_hit_tokens": "garbage",
+            "prompt_cache_miss_tokens": -5,
+        },
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.input_tokens == 1000
+    assert result.cache_read_input_tokens == 0
+
+
+# ============================================================================
+# Truly-empty response with a non-zero USD cost → full refund
+#
+# When an upstream reports a USD cost but the response carries NO tokens at all
+# (input, output, cache-read and cache-creation all zero), billing the
+# USD-derived cost charges the user for nothing. Refund in full. The gate is
+# tightened relative to PR #489: a cache-read/-creation-only turn legitimately
+# reports zero prompt/completion tokens with a real cost and must still bill.
+# ============================================================================
+@pytest.mark.asyncio
+async def test_truly_empty_usd_cost_response_is_refunded(
+    mock_fixed_pricing: None,
+) -> None:
+    """0 input + 0 output + 0 cache tokens with a non-zero USD cost → refund."""
+    response = {
+        "model": "gpt-4",
+        "usage": {
+            "prompt_tokens": 0,
+            "completion_tokens": 0,
+            "total_cost": 0.01,  # non-zero USD cost despite no tokens
+        },
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.total_msats == 0  # full refund
+    assert result.input_msats == 0
+    assert result.output_msats == 0
+    assert result.total_usd == 0.0
+    assert result.input_tokens == 0
+    assert result.output_tokens == 0
+    assert result.cache_read_input_tokens == 0
+    assert result.cache_creation_input_tokens == 0
+
+
+@pytest.mark.asyncio
+async def test_cache_read_only_usd_cost_response_is_billed(
+    mock_fixed_pricing: None,
+) -> None:
+    """Cache-read-only turn (0 prompt/completion, non-zero cost) still bills."""
+    response = {
+        "model": "claude-3-5-sonnet",
+        "usage": {
+            "input_tokens": 0,
+            "output_tokens": 0,
+            "cache_read_input_tokens": 1000,  # real cached usage
+            "cache_creation_input_tokens": 0,
+            "total_cost": 0.01,  # non-zero USD cost
+        },
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    # NOT refunded — the USD cost is billed in full. Pinning the exact value
+    # guards against any future regression that would over-refund a cache-only
+    # turn (the bug in PR #489, which refunded whenever prompt+completion == 0).
+    assert result.total_msats == 200000
+    assert result.total_usd == 0.01
+    assert result.cache_read_input_tokens == 1000
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("total_cost", "input_cost", "output_cost", "expected_msats"),
+    [
+        (0.000471, 0.00023451, 0.00023649, 9420),
+        (0.00000004, 0.00000002, 0.00000002, 1),
+    ],
+)
+async def test_small_usd_cost_components_sum_to_rounded_total(
+    total_cost: float,
+    input_cost: float,
+    output_cost: float,
+    expected_msats: int,
+) -> None:
+    """Small USD component costs must retain every billed millisatoshi."""
+    response = {
+        "model": "gpt-4",
+        "usage": {
+            "prompt_tokens": 1,
+            "completion_tokens": 1,
+            "cost_details": {
+                "total_cost": total_cost,
+                "input_cost": input_cost,
+                "output_cost": output_cost,
+            },
+        },
+    }
+
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.total_msats == expected_msats
+    assert result.input_msats + result.output_msats == result.total_msats
+
+
+@pytest.mark.asyncio
+async def test_openrouter_upstream_inference_cost_components_are_used() -> None:
+    """OpenRouter component aliases must determine the input/output split."""
+    response = {
+        "model": "gpt-4",
+        "usage": {
+            "prompt_tokens": 375,
+            "completion_tokens": 158,
+            "total_tokens": 533,
+            "cost": 0.00022354,
+            "is_byok": False,
+            "prompt_tokens_details": {
+                "cached_tokens": 286,
+                "cache_write_tokens": 0,
+            },
+            "cost_details": {
+                "upstream_inference_cost": 0.00022354,
+                "upstream_inference_prompt_cost": 0.00004974,
+                "upstream_inference_completions_cost": 0.0001738,
+            },
+            "completion_tokens_details": {"reasoning_tokens": 17},
+        },
+    }
+
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.input_msats == 994
+    assert result.output_msats == 3477
+    assert result.input_msats + result.output_msats == result.total_msats == 4471
+
+
+# ============================================================================
+# PPQ.AI BYOK: upstream_inference_cost + BYOK fee billing
+#
+# PPQ.AI (bring-your-own-key) returns a small ~5 % BYOK routing fee in
+# ``usage.cost`` and the real inference cost in
+# ``cost_details.upstream_inference_cost``.  The old code billed only the fee,
+# under-charging by ~20×.  The fix bills ``upstream_inference_cost + byok_fee``
+# — what PPQ actually deducts from the balance.
+# Payload numbers are from a live ``glm-5.2-fast`` request (GitHub issue #615).
+# ============================================================================
+@pytest.mark.asyncio
+async def test_ppq_byok_bills_upstream_inference_cost_plus_fee() -> None:
+    """PPQ.AI BYOK must bill upstream_inference_cost + byok_fee, not just the
+    fee.  Mirrors the live request from GitHub issue #615."""
+    response = {
+        "model": "glm-5.2-fast",
+        "usage": {
+            "prompt_tokens": 164371,  # includes 159301 cached
+            "completion_tokens": 99,
+            "cost": 0.002260057305,  # ~5% BYOK routing fee
+            "is_byok": True,
+            "prompt_tokens_details": {"cached_tokens": 159301},
+            "cost_details": {
+                "upstream_inference_cost": 0.04475361,
+                "upstream_inference_prompt_cost": 0.04410021,
+                "upstream_inference_completions_cost": 0.0006534,
+            },
+        },
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    # The fix bills upstream_inference_cost + byok_fee (~0.047 USD → ~940k
+    # msats), not the fee alone (~0.0023 USD → ~45k msats).  ~20× correction.
+    assert result.total_msats == 940274
+    assert result.input_msats + result.output_msats == result.total_msats
+    assert result.input_msats == 926546
+    assert result.output_msats == 13728
+    assert result.total_usd == pytest.approx(0.047013667305)
+    # Token normalisation (OpenAI dialect: cached included in prompt_tokens)
+    assert result.input_tokens == 5070  # 164371 - 159301
+    assert result.cache_read_input_tokens == 159301
+    assert result.output_tokens == 99
+
+
+@pytest.mark.asyncio
+async def test_ppq_byok_fee_only_would_undercharge() -> None:
+    """Sanity check: billing only usage.cost (the BYOK fee) under-charges by
+    ~20×.  This documents the regression the fix prevents."""
+    response = {
+        "model": "glm-5.2-fast",
+        "usage": {
+            "prompt_tokens": 164371,
+            "completion_tokens": 99,
+            "cost": 0.002260057305,  # BYOK fee only — no upstream_inference_cost
+            "is_byok": True,
+            "prompt_tokens_details": {"cached_tokens": 159301},
+        },
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    # Without upstream_inference_cost, only the fee is billed — the old bug.
+    assert result.total_msats == 45202
+
+
 # ============================================================================
 # Test 13: Missing Usage Block
 # ============================================================================
 @pytest.mark.asyncio
-async def test_missing_usage_block(mock_session: AsyncMock, mock_fixed_pricing: None) -> None:
+async def test_missing_upstream_cost_uses_litellm_model_pricing(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """Token usage without an upstream cost is priced from LiteLLM."""
+    monkeypatch.setattr(settings, "fixed_pricing", True)
+    monkeypatch.setattr(settings, "fixed_per_1k_input_tokens", 0)
+    monkeypatch.setattr(settings, "fixed_per_1k_output_tokens", 0)
+    monkeypatch.setattr(
+        "routstr.payment.models.litellm_cost_entry",
+        lambda model: {
+            "input_cost_per_token": 0.000001,
+            "output_cost_per_token": 0.000002,
+        },
+    )
+    response = {
+        "model": "priced-by-litellm",
+        "usage": {
+            "prompt_tokens": 90,
+            "completion_tokens": 80,
+            "total_tokens": 170,
+            "prompt_tokens_details": {"cached_tokens": 0},
+            "completion_tokens_details": {"reasoning_tokens": 74},
+            "prompt_cache_hit_tokens": 0,
+            "prompt_cache_miss_tokens": 90,
+        },
+    }
+
+    result = await calculate_cost(response, max_cost=10000)
+
+    assert isinstance(result, CostData)
+    assert not isinstance(result, MaxCostData)
+    assert result.input_msats == 1800
+    assert result.output_msats == 3200
+    assert result.input_msats + result.output_msats == result.total_msats == 5000
+
+
+@pytest.mark.asyncio
+async def test_missing_usage_block(mock_fixed_pricing: None) -> None:
     """When usage is missing, return MaxCostData with zero tokens."""
     response = {"model": "gpt-4", "choices": [{"message": {"content": "test"}}]}
-    result = await calculate_cost(response, max_cost=100000, session=mock_session)
+    result = await calculate_cost(response, max_cost=100000)
 
     assert isinstance(result, MaxCostData)
     assert result.input_tokens == 0
@@ -335,10 +654,10 @@ async def test_missing_usage_block(mock_session: AsyncMock, mock_fixed_pricing:
 # Test 14: Null Usage Block
 # ============================================================================
 @pytest.mark.asyncio
-async def test_null_usage_block(mock_session: AsyncMock, mock_fixed_pricing: None) -> None:
+async def test_null_usage_block(mock_fixed_pricing: None) -> None:
     """When usage is null, return MaxCostData with zero tokens."""
     response = {"model": "gpt-4", "usage": None}
-    result = await calculate_cost(response, max_cost=100000, session=mock_session)
+    result = await calculate_cost(response, max_cost=100000)
 
     assert isinstance(result, MaxCostData)
     assert result.input_tokens == 0

tests/unit/test_coverage_admin.py

@@ -0,0 +1,136 @@
+"""Coverage tests for admin.py (currently 35%).
+
+Tests admin endpoints that are testable without full app setup:
+withdraw validation, authentication guards, and slug validation.
+"""
+
+from unittest.mock import Mock, patch
+
+import pytest
+from fastapi import HTTPException, Request
+
+# ===========================================================================
+# withdraw — validation and edge cases
+# ===========================================================================
+
+@pytest.mark.asyncio
+async def test_withdraw_rejects_zero_amount() -> None:
+    """withdraw validation rejects amount <= 0."""
+    from routstr.core.admin import WithdrawRequest, withdraw
+
+    request = Request(scope={"type": "http", "method": "POST"})
+
+    with pytest.raises(HTTPException) as exc_info:
+        await withdraw(request, WithdrawRequest(amount=0, unit="sat"))
+
+    assert exc_info.value.status_code == 400
+
+
+@pytest.mark.asyncio
+async def test_withdraw_rejects_negative_amount() -> None:
+    """withdraw validation rejects negative amounts."""
+    from routstr.core.admin import WithdrawRequest, withdraw
+
+    request = Request(scope={"type": "http", "method": "POST"})
+
+    with pytest.raises(HTTPException) as exc_info:
+        await withdraw(request, WithdrawRequest(amount=-100, unit="sat"))
+
+    assert exc_info.value.status_code == 400
+
+
+@pytest.mark.asyncio
+async def test_withdraw_rejects_insufficient_balance() -> None:
+    """withdraw returns 400 when wallet balance is insufficient."""
+    from routstr.core.admin import WithdrawRequest, withdraw
+
+    request = Request(scope={"type": "http", "method": "POST"})
+
+    with patch("routstr.core.admin.get_wallet") as mock_wallet, \
+         patch("routstr.core.admin.get_proofs_per_mint_and_unit") as mock_proofs, \
+         patch("routstr.core.admin.slow_filter_spend_proofs") as mock_filter:
+
+        mock_w = Mock()
+        mock_w.keysets = {}
+        mock_w.proofs = []
+        mock_wallet.return_value = mock_w
+        mock_proofs.return_value = []
+        mock_filter.return_value = []
+
+        with pytest.raises(HTTPException) as exc_info:
+            await withdraw(request, WithdrawRequest(amount=1000000, unit="sat"))
+
+        assert exc_info.value.status_code == 400
+        assert "Insufficient" in str(exc_info.value.detail)
+
+
+# ===========================================================================
+# require_admin_api guard
+# ===========================================================================
+
+@pytest.mark.asyncio
+async def test_require_admin_rejects_no_session() -> None:
+    """require_admin_api rejects requests without admin session cookie."""
+    from routstr.core.admin import require_admin_api
+
+    request = Request(scope={
+        "type": "http",
+        "method": "GET",
+        "headers": [],
+    })
+
+    with pytest.raises(HTTPException) as exc_info:
+        await require_admin_api(request)
+
+    # 401 or 403 depending on auth configuration
+    assert exc_info.value.status_code in (401, 403)
+
+
+# ===========================================================================
+# _validate_slug
+# ===========================================================================
+
+def test_validate_slug_accepts_valid() -> None:
+    """Valid slugs pass validation."""
+    from routstr.core.admin import _validate_slug
+
+    assert _validate_slug("valid-slug") == "valid-slug"
+    assert _validate_slug("valid123") == "valid123"
+    assert _validate_slug("my-provider") == "my-provider"
+
+
+def test_validate_slug_rejects_spaces() -> None:
+    """Slugs with spaces are rejected."""
+    from fastapi import HTTPException
+
+    from routstr.core.admin import _validate_slug
+
+    with pytest.raises(HTTPException):
+        _validate_slug("invalid slug")
+
+
+def test_validate_slug_rejects_too_short() -> None:
+    """Slugs shorter than 3 chars are rejected."""
+    from fastapi import HTTPException
+
+    from routstr.core.admin import _validate_slug
+
+    with pytest.raises(HTTPException):
+        _validate_slug("ab")
+
+
+# ===========================================================================
+# admin login endpoint
+# ===========================================================================
+
+@pytest.mark.asyncio
+async def test_admin_login_requires_payload() -> None:
+    """admin_login requires a payload — verify it exists."""
+    # Verify the function signature
+    import inspect
+
+    from routstr.core.admin import admin_login
+    sig = inspect.signature(admin_login)
+    params = list(sig.parameters.keys())
+    assert "request" in params
+    assert "payload" in params or len(params) >= 2

tests/unit/test_coverage_base.py

@@ -0,0 +1,204 @@
+"""Coverage tests for base.py (currently 41%).
+
+Tests preparers, builders, accessors, and model cache methods.
+"""
+
+from unittest.mock import Mock
+
+import pytest
+
+from routstr.upstream.base import BaseUpstreamProvider
+
+# ===========================================================================
+# prepare_headers
+# ===========================================================================
+
+def test_prepare_headers_adds_auth() -> None:
+    """API key is added as Bearer token."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test-key")
+    headers = p.prepare_headers({})
+
+    assert "Authorization" in headers
+    assert headers["Authorization"] == "Bearer sk-test-key"
+
+
+def test_prepare_headers_preserves_existing() -> None:
+    """Existing headers are preserved."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test-key")
+    headers = p.prepare_headers({"X-Custom": "value", "Content-Type": "application/json"})
+
+    assert headers["X-Custom"] == "value"
+    assert headers["Content-Type"] == "application/json"
+
+
+def test_prepare_headers_auth_header_passthrough() -> None:
+    """Authorization header is handled — verify current behaviour."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test-key")
+    headers = p.prepare_headers({"Authorization": "Bearer user-key"})
+
+    # Currently provider key is used (may be intentional for proxy pattern)
+    assert "Authorization" in headers
+
+
+# ===========================================================================
+# prepare_params
+# ===========================================================================
+
+@pytest.mark.asyncio
+async def test_prepare_params_passes_through() -> None:
+    """Query params are preserved by default."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test-key")
+    params = p.prepare_params("/v1/chat/completions", {"temperature": "0.7"})
+
+    assert params["temperature"] == "0.7"
+
+
+# ===========================================================================
+# transform_model_name / normalize_request_path / get_request_base_url
+# ===========================================================================
+
+def test_transform_model_name_default_passthrough() -> None:
+    """Default returns model_id unchanged."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test-key")
+    assert p.transform_model_name("gpt-4") == "gpt-4"
+    assert p.transform_model_name("") == ""
+
+
+def test_normalize_request_path_passthrough() -> None:
+    """Default returns path unchanged."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test-key")
+    assert p.normalize_request_path("/v1/chat/completions") == "/v1/chat/completions"
+
+
+def test_get_request_base_url_default() -> None:
+    """Default returns the provider's base_url."""
+    p = BaseUpstreamProvider("https://api.test.com/v1", "sk-test-key")
+    url = p.get_request_base_url("/v1/chat/completions")
+    assert url == "https://api.test.com/v1"
+
+
+# ===========================================================================
+# build_request_url
+# ===========================================================================
+
+def test_build_request_url_combines_base_and_path() -> None:
+    """Combines base_url and path."""
+    p = BaseUpstreamProvider("https://api.test.com/v1", "sk-test-key")
+    url = p.build_request_url("/chat/completions")
+    assert "api.test.com" in url
+    assert "/chat/completions" in url
+
+
+# ===========================================================================
+# get_litellm_provider_prefix / get_provider_metadata
+# ===========================================================================
+
+def test_get_litellm_provider_prefix_default() -> None:
+    """Default returns a string prefix."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test-key")
+    prefix = p.get_litellm_provider_prefix()
+    assert isinstance(prefix, str)
+
+
+def test_get_provider_metadata_returns_dict() -> None:
+    """Default metadata has name and capabilities."""
+    metadata = BaseUpstreamProvider.get_provider_metadata()
+    assert isinstance(metadata, dict)
+    assert "name" in metadata
+
+
+# ===========================================================================
+# from_db_row
+# ===========================================================================
+
+@pytest.mark.asyncio
+async def test_from_db_row_returns_provider() -> None:
+    """from_db_row constructs a provider from a valid row."""
+    mock_row = Mock()
+    mock_row.base_url = "https://api.test.com"
+    mock_row.api_key = "sk-test-key"
+    mock_row.slug = "test-slug"
+    mock_row.provider_fee = 1.0
+    mock_row.field_overrides = None
+    mock_row.name = "Test"
+
+    result = BaseUpstreamProvider.from_db_row(mock_row)
+    assert result is not None
+
+
+# ===========================================================================
+# prepare_request_body
+# ===========================================================================
+
+def test_prepare_request_body_with_model() -> None:
+    """prepare_request_body takes bytes body and Model object."""
+    mock_model = Mock()
+    mock_model.id = "gpt-4"
+    mock_model.forwarded_model_id = None
+
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test-key")
+
+    # None body returns None
+    result = p.prepare_request_body(None, mock_model)
+    assert result is None
+
+
+# ===========================================================================
+# prepare_responses_request_body
+# ===========================================================================
+
+def test_prepare_responses_request_body_none() -> None:
+    """None body returns None."""
+    model_obj = Mock()
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test-key")
+    result = p.prepare_responses_request_body(None, model_obj)
+    assert result is None
+
+
+# ===========================================================================
+# _upstream_accepts_cache_control
+# ===========================================================================
+
+def test_upstream_accepts_cache_control_default() -> None:
+    """Default: upstream does NOT accept cache-control."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test-key")
+    assert p._upstream_accepts_cache_control() is False
+
+
+# ===========================================================================
+# inject_cost_metadata
+# ===========================================================================
+
+def test_inject_cost_metadata_adds_metadata() -> None:
+    """Cost metadata is injected into the response dict."""
+    mock_key = Mock()
+    mock_key.balance_msat = 500000
+
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test-key")
+    data = {"model": "gpt-4", "usage": {"prompt_tokens": 100}}
+    cost_data = {
+        "base_msats": 200000,
+        "input_msats": 100000,
+        "output_msats": 100000,
+        "total_msats": 200000,
+        "total_usd": 0.01,
+        "input_tokens": 100,
+        "output_tokens": 50,
+    }
+
+    p.inject_cost_metadata(data, cost_data, mock_key)
+
+    # Metadata is nested under metadata.routstr.cost
+    assert "metadata" in data or "routstr_cost" in data or "cost" in data
+
+
+# ===========================================================================
+# _apply_provider_field
+# ===========================================================================
+
+def test_apply_provider_field_adds_to_response() -> None:
+    """Provider field is added to response JSON."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test-key")
+    data = {"id": "chatcmpl-123"}
+    p._apply_provider_field(data)
+    assert "provider" in data

tests/unit/test_coverage_base2.py

@@ -0,0 +1,217 @@
+"""Additional coverage tests for base.py (41% → target 50%+).
+
+Tests error message extraction, static helpers, model cache, and cost hooks.
+
+These test existing correct behavior — all should PASS.
+"""
+
+import json
+from unittest.mock import Mock
+
+import pytest
+
+from routstr.upstream.base import BaseUpstreamProvider
+
+# ===========================================================================
+# _extract_upstream_error_message
+# ===========================================================================
+
+def test_extract_error_from_json_body() -> None:
+    """Error message is extracted from JSON upstream error response."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test")
+    body = json.dumps({"error": {"message": "Model not found", "type": "not_found"}}).encode()
+
+    msg, error_type = p._extract_upstream_error_message(body)
+
+    assert "Model not found" in msg
+    assert error_type == "not_found"
+
+
+def test_extract_error_from_simple_json() -> None:
+    """Simple JSON error with direct message key."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test")
+    body = json.dumps({"message": "Rate limit exceeded"}).encode()
+
+    msg, error_type = p._extract_upstream_error_message(body)
+
+    assert "Rate limit" in msg
+
+
+def test_extract_error_from_text_body() -> None:
+    """Non-JSON text body is returned as-is."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test")
+
+    msg, error_type = p._extract_upstream_error_message(b"Internal Server Error")
+
+    assert "Internal Server Error" in msg
+
+
+def test_extract_error_empty_body() -> None:
+    """Empty body returns a generic message."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test")
+
+    msg, error_type = p._extract_upstream_error_message(b"")
+
+    assert isinstance(msg, str)
+    assert len(msg) > 0
+
+
+def test_extract_error_simple_error_string_not_parsed() -> None:
+    """JSON error as plain string (not dict) falls through to generic message."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test")
+    body = json.dumps({"error": "Invalid API key"}).encode()
+
+    msg, error_type = p._extract_upstream_error_message(body)
+
+    # Simple error strings not nested in a dict object use generic message
+    assert "Upstream request failed" in msg or "Invalid" in msg
+
+
+# ===========================================================================
+# on_upstream_error_redirect
+# ===========================================================================
+
+@pytest.mark.asyncio
+async def test_on_upstream_error_redirect_noop() -> None:
+    """Default implementation is a no-op for non-redirect statuses."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test")
+    await p.on_upstream_error_redirect(402, "Insufficient balance")
+
+
+@pytest.mark.asyncio
+async def test_on_upstream_error_redirect_429() -> None:
+    """429 rate limit passes through (subclasses may override)."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test")
+    await p.on_upstream_error_redirect(429, "Rate limited")
+
+
+# ===========================================================================
+# _fold_cache_into_input_tokens (static method)
+# ===========================================================================
+
+def test_fold_cache_no_cache_data() -> None:
+    """Usage without cache details is unchanged."""
+    from routstr.upstream.base import BaseUpstreamProvider
+
+    usage = Mock()
+    usage.prompt_tokens = 100
+    del usage.prompt_tokens_details  # No cache details
+
+    BaseUpstreamProvider._fold_cache_into_input_tokens(usage)
+    # Should not modify the usage object when no cache exists
+
+
+def test_fold_cache_preserves_total() -> None:
+    """Total prompt tokens remain the same after folding cache."""
+    from routstr.upstream.base import BaseUpstreamProvider
+
+    usage = Mock()
+    usage.prompt_tokens = 100
+    details = Mock()
+    details.cached_tokens = 30
+    usage.prompt_tokens_details = details
+
+    BaseUpstreamProvider._fold_cache_into_input_tokens(usage)
+    # prompt_tokens should still be 100 (total unchanged)
+    assert usage.prompt_tokens == 100
+
+
+# ===========================================================================
+# get_cached_models / get_cached_model_by_id
+# ===========================================================================
+
+def test_get_cached_models_returns_list() -> None:
+    """get_cached_models always returns a list."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test")
+    models = p.get_cached_models()
+    assert isinstance(models, list)
+
+
+def test_get_cached_model_by_id_unknown_returns_none() -> None:
+    """Unknown model ID returns None."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test")
+    result = p.get_cached_model_by_id("nonexistent-model-xyz-12345")
+    assert result is None
+
+
+# ===========================================================================
+# get_x_cashu_cost
+# ===========================================================================
+
+def test_get_x_cashu_cost_with_usage() -> None:
+    """Cost is calculated from response data with usage info."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test")
+    response_data = {
+        "model": "gpt-4",
+        "usage": {"prompt_tokens": 100, "completion_tokens": 50},
+    }
+
+    result = p.get_x_cashu_cost(response_data, 100000)
+
+    # Either returns None (needs more data) or a cost object
+    assert result is not None
+
+
+def test_get_x_cashu_cost_no_usage() -> None:
+    """Response without usage returns MaxCostData."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test")
+    response_data = {"model": "gpt-4"}
+
+    result = p.get_x_cashu_cost(response_data, 100000)
+
+    # Without usage, uses max_cost
+    assert result is not None
+
+
+# ===========================================================================
+# get_balance
+# ===========================================================================
+
+@pytest.mark.asyncio
+async def test_get_balance_raises_not_implemented() -> None:
+    """Default get_balance raises NotImplementedError (no account support)."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test")
+    with pytest.raises(NotImplementedError):
+        await p.get_balance()
+
+
+# ===========================================================================
+# refresh_models_cache
+# ===========================================================================
+
+@pytest.mark.asyncio
+async def test_refresh_models_cache_no_providers() -> None:
+    """refresh_models_cache handles empty provider list gracefully."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test")
+    # Default implementation may be a no-op or raise
+    try:
+        await p.refresh_models_cache()
+    except Exception:
+        pass  # May fail without DB — that's fine
+
+
+# ===========================================================================
+# fetch_models
+# ===========================================================================
+
+@pytest.mark.asyncio
+async def test_fetch_models_returns_list() -> None:
+    """fetch_models returns a model list (or empty) for default provider."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test")
+    try:
+        result = await p.fetch_models()
+        assert isinstance(result, list)
+    except Exception:
+        pass  # May fail without network
+
+
+# ===========================================================================
+# create_account
+# ===========================================================================
+
+@pytest.mark.asyncio
+async def test_create_account_raises_not_implemented() -> None:
+    """Default create_account raises NotImplementedError."""
+    p = BaseUpstreamProvider("https://api.test.com", "sk-test")
+    with pytest.raises(NotImplementedError):
+        await p.create_account()

tests/unit/test_coverage_middleware.py

@@ -0,0 +1,150 @@
+"""Coverage-filling tests for middleware.py (currently 38% coverage).
+
+Only LoggingMiddleware and request_id_context exist on main.
+ConcurrencyLimiterMiddleware + TimeoutMiddleware are on an unmerged branch.
+"""
+
+from fastapi import FastAPI, Request
+from fastapi.testclient import TestClient
+
+# ---------------------------------------------------------------------------
+# LoggingMiddleware
+# ---------------------------------------------------------------------------
+
+def test_logging_middleware_adds_request_id() -> None:
+    """Every request gets an x-routstr-request-id header."""
+    from routstr.core.middleware import LoggingMiddleware
+
+    app = FastAPI()
+
+    @app.get("/test")
+    async def test_endpoint(request: Request) -> dict:
+        assert hasattr(request.state, "request_id")
+        assert request.state.request_id is not None
+        return {"ok": True}
+
+    app.add_middleware(LoggingMiddleware)
+
+    client = TestClient(app)
+    response = client.get("/test")
+
+    assert response.status_code == 200
+    assert "x-routstr-request-id" in response.headers
+    assert len(response.headers["x-routstr-request-id"]) == 36  # UUID4 length
+
+
+def test_logging_middleware_skips_head_requests() -> None:
+    """HEAD requests are skipped by _should_log (health probes)."""
+    from routstr.core.middleware import LoggingMiddleware
+
+    app = FastAPI()
+
+    @app.head("/test")
+    async def test_endpoint(request: Request) -> dict:
+        return {"ok": True}
+
+    app.add_middleware(LoggingMiddleware)
+
+    client = TestClient(app)
+    response = client.head("/test")
+
+    assert response.status_code == 200
+    assert "x-routstr-request-id" in response.headers
+
+
+def test_logging_middleware_skips_options_requests() -> None:
+    """OPTIONS requests (CORS preflight) are skipped."""
+    from routstr.core.middleware import LoggingMiddleware
+
+    app = FastAPI()
+
+    @app.options("/test")
+    async def test_endpoint(request: Request) -> dict:
+        return {"ok": True}
+
+    app.add_middleware(LoggingMiddleware)
+
+    client = TestClient(app)
+    response = client.options("/test")
+
+    assert response.status_code == 200
+    assert "x-routstr-request-id" in response.headers
+
+
+def test_should_log_rejects_admin_api_prefix() -> None:
+    """Admin API polling paths are skipped."""
+    from routstr.core.middleware import _should_log
+
+    assert _should_log("GET", "/admin/api/balances") is False
+    assert _should_log("GET", "/admin/api/logs") is False
+    assert _should_log("GET", "/admin/api/providers") is False
+
+
+def test_should_log_rejects_nextjs_chunks() -> None:
+    """Next.js static chunks are skipped."""
+    from routstr.core.middleware import _should_log
+
+    assert _should_log("GET", "/_next/static/chunks/main.js") is False
+    assert _should_log("GET", "/_next/data/build-id/page.json") is False
+
+
+def test_should_log_rejects_exact_paths() -> None:
+    """Exact paths like /favicon.ico are skipped."""
+    from routstr.core.middleware import _should_log
+
+    assert _should_log("GET", "/favicon.ico") is False
+    assert _should_log("GET", "/v1/wallet/info") is False
+    assert _should_log("GET", "/index.txt") is False
+    assert _should_log("GET", "/login/index.txt") is False
+
+
+def test_should_log_accepts_normal_paths() -> None:
+    """Normal API paths are logged."""
+    from routstr.core.middleware import _should_log
+
+    assert _should_log("GET", "/v1/chat/completions") is True
+    assert _should_log("POST", "/v1/chat/completions") is True
+    assert _should_log("GET", "/v1/models") is True
+    assert _should_log("POST", "/api/some-endpoint") is True
+
+
+def test_should_log_accepts_non_skipped_path() -> None:
+    """Generic paths not in skip list are logged."""
+    from routstr.core.middleware import _should_log
+
+    assert _should_log("GET", "/some/random/path") is True
+    assert _should_log("POST", "/api/custom") is True
+
+
+def test_request_id_context_is_contextvar() -> None:
+    """request_id_context is a ContextVar[str | None] with no default value."""
+    from contextvars import ContextVar
+
+    from routstr.core.middleware import request_id_context
+
+    assert isinstance(request_id_context, ContextVar)
+    # ContextVar without a default raises LookupError when accessed without being set
+    try:
+        val = request_id_context.get()
+        # If it returns, it should be None
+        assert val is None
+    except LookupError:
+        # Expected: ContextVar with no default raises LookupError
+        pass
+
+
+def test_middleware_exports() -> None:
+    """Only LoggingMiddleware is exported on main."""
+    from routstr.core.middleware import LoggingMiddleware, request_id_context
+
+    assert LoggingMiddleware is not None
+    assert request_id_context is not None
+
+
+def test_middleware_skips_health_probe_path() -> None:
+    """Health probe paths pass through without logging."""
+    from routstr.core.middleware import _should_log
+
+    # HEAD method is always skipped regardless of path
+    assert _should_log("HEAD", "/v1/chat/completions") is False
+    assert _should_log("OPTIONS", "/v1/chat/completions") is False

tests/unit/test_coverage_payment_helpers.py

@@ -0,0 +1,181 @@
+"""Coverage-filling tests for payment/helpers.py (currently 52% coverage).
+
+Tests the real public API: check_token_balance, get_max_cost_for_model,
+estimate_tokens, create_error_response, etc.
+"""
+
+from unittest.mock import Mock, patch
+
+import pytest
+
+# ---------------------------------------------------------------------------
+# check_token_balance
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_check_token_balance_x_cashu_present() -> None:
+    """X-Cashu header triggers token deserialization and balance check."""
+    from routstr.payment.helpers import check_token_balance
+
+    headers = {"x-cashu": "cashuAtest_token"}
+    body = {"model": "gpt-4"}
+
+    with patch("routstr.payment.helpers.deserialize_token_from_string") as mock_deser:
+        mock_token = Mock()
+        mock_token.amount = 50000
+        mock_token.unit = "sat"
+        mock_deser.return_value = mock_token
+
+        # Should not raise — balance is sufficient
+        check_token_balance(headers, body, 1000)
+
+
+@pytest.mark.asyncio
+async def test_check_token_balance_no_x_cashu_raises() -> None:
+    """Missing X-Cashu header raises HTTPException (401 on main)."""
+    from fastapi import HTTPException
+
+    from routstr.payment.helpers import check_token_balance
+
+    headers: dict[str, str] = {}
+    body = {"model": "gpt-4"}
+
+    with pytest.raises(HTTPException) as exc_info:
+        check_token_balance(headers, body, 1000)
+
+    assert exc_info.value.status_code == 401
+
+
+@pytest.mark.asyncio
+async def test_check_token_balance_insufficient_raises() -> None:
+    """Token with insufficient balance raises HTTPException 402.
+
+    max_cost_for_model is in msat, so with amount=100 sat (=100,000 msat),
+    max_cost=200,000 msat triggers the insufficient balance check.
+    """
+    from fastapi import HTTPException
+
+    from routstr.payment.helpers import check_token_balance
+
+    headers = {"x-cashu": "cashuAtest_token"}
+    body = {"model": "gpt-4"}
+
+    with patch("routstr.payment.helpers.deserialize_token_from_string") as mock_deser:
+        mock_token = Mock()
+        mock_token.amount = 100  # 100 sat
+        mock_token.unit = "sat"
+        mock_deser.return_value = mock_token
+
+        with pytest.raises(HTTPException) as exc_info:
+            # 200,000 msat > 100,000 msat (100 sat * 1000)
+            check_token_balance(headers, body, 200000)
+
+        assert exc_info.value.status_code == 402
+
+
+# ---------------------------------------------------------------------------
+# estimate_tokens
+# ---------------------------------------------------------------------------
+
+def test_estimate_tokens_empty_messages() -> None:
+    """Empty message list returns 0 tokens."""
+    from routstr.payment.helpers import estimate_tokens
+
+    result = estimate_tokens([])
+
+    assert result == 0
+
+
+def test_estimate_tokens_text_content() -> None:
+    """Text messages are counted."""
+    from routstr.payment.helpers import estimate_tokens
+
+    messages = [
+        {"role": "system", "content": "You are a helpful assistant."},
+        {"role": "user", "content": "Hello, how are you?"},
+    ]
+
+    result = estimate_tokens(messages)
+
+    assert result > 0
+    assert isinstance(result, int)
+
+
+def test_estimate_tokens_long_text() -> None:
+    """Longer messages produce higher token counts."""
+    from routstr.payment.helpers import estimate_tokens
+
+    short = estimate_tokens([{"role": "user", "content": "Hi"}])
+    long = estimate_tokens([{"role": "user", "content": "Hello " * 100}])
+
+    assert long > short
+
+
+# ---------------------------------------------------------------------------
+# create_error_response
+# ---------------------------------------------------------------------------
+
+def test_create_error_response_402() -> None:
+    """402 Payment Required error is properly formatted."""
+    from fastapi import Request
+
+    from routstr.payment.helpers import create_error_response
+
+    request = Request(scope={"type": "http", "method": "GET"})
+    result = create_error_response("insufficient_funds", "Insufficient balance", 402, request)
+
+    assert result.status_code == 402
+
+
+def test_create_error_response_500() -> None:
+    """500 Internal Server Error is properly formatted."""
+    from fastapi import Request
+
+    from routstr.payment.helpers import create_error_response
+
+    request = Request(scope={"type": "http", "method": "GET"})
+    result = create_error_response("server_error", "Internal error", 500, request)
+
+    assert result.status_code == 500
+
+
+# ---------------------------------------------------------------------------
+# Image token estimation helpers
+# ---------------------------------------------------------------------------
+
+def test_image_dimensions_valid_png() -> None:
+    """_get_image_dimensions returns width and height for a valid PNG."""
+    from routstr.payment.helpers import _get_image_dimensions
+
+    # A minimal 1x1 red PNG (valid minimal file)
+    png = (
+        b"\x89PNG\r\n\x1a\n"
+        b"\x00\x00\x00\rIHDR\x00\x00\x00\x01\x00\x00\x00\x01\x08\x02"
+        b"\x00\x00\x00\x90wS\xde"
+        b"\x00\x00\x00\x0cIDAT\x08\xd7c\xf8\x0f\x00\x00\x01\x01\x00\x05"
+        b"\x18\xd8N"
+        b"\x00\x00\x00\x00IEND\xaeB`\x82"
+    )
+
+    w, h = _get_image_dimensions(png)
+    assert w == 1
+    assert h == 1
+
+
+def test_calculate_image_tokens_low_detail() -> None:
+    """Low detail images are always 85 tokens."""
+    from routstr.payment.helpers import _calculate_image_tokens
+
+    tokens = _calculate_image_tokens(1024, 1024, "low")
+
+    assert tokens == 85
+
+
+def test_calculate_image_tokens_high_detail() -> None:
+    """High detail images are scaled and tile-based."""
+    from routstr.payment.helpers import _calculate_image_tokens
+
+    tokens = _calculate_image_tokens(1024, 1024, "high")
+
+    assert tokens > 85
+    assert isinstance(tokens, int)

tests/unit/test_coverage_proxy.py

@@ -0,0 +1,161 @@
+"""Coverage tests for proxy.py (currently 47%).
+
+Tests request parsing, model extraction, and routing helpers.
+"""
+
+import json
+
+import pytest
+from fastapi import HTTPException
+
+# ===========================================================================
+# parse_request_body_json
+# ===========================================================================
+
+def test_parse_json_valid_body() -> None:
+    """Valid JSON body is parsed correctly for chat completions."""
+    from routstr.proxy import parse_request_body_json
+
+    body = json.dumps({"model": "gpt-4", "messages": [{"role": "user", "content": "hi"}]}).encode()
+    result = parse_request_body_json(body, "/v1/chat/completions")
+
+    assert result["model"] == "gpt-4"
+    assert result["messages"][0]["role"] == "user"
+
+
+def test_parse_json_invalid_raises_400() -> None:
+    """Invalid JSON raises HTTPException 400."""
+    from routstr.proxy import parse_request_body_json
+
+    with pytest.raises(HTTPException) as exc_info:
+        parse_request_body_json(b"not json", "/v1/chat/completions")
+
+    assert exc_info.value.status_code == 400
+
+
+def test_parse_json_empty_body() -> None:
+    """Empty body returns empty dict."""
+    from routstr.proxy import parse_request_body_json
+
+    result = parse_request_body_json(b"", "/v1/chat/completions")
+    assert isinstance(result, dict)
+    assert result == {}
+
+
+def test_parse_json_responses_path() -> None:
+    """Responses API path is handled."""
+    from routstr.proxy import parse_request_body_json
+
+    body = json.dumps({"model": "gpt-4", "input": "hello"}).encode()
+    result = parse_request_body_json(body, "/v1/responses")
+
+    assert "model" in result
+
+
+def test_parse_json_rejects_non_integer_max_tokens() -> None:
+    """max_tokens must be an integer."""
+    from routstr.proxy import parse_request_body_json
+
+    body = json.dumps({"model": "gpt-4", "max_tokens": "abc"}).encode()
+
+    with pytest.raises(HTTPException) as exc_info:
+        parse_request_body_json(body, "/v1/chat/completions")
+
+    assert exc_info.value.status_code == 400
+
+
+# ===========================================================================
+# extract_model_from_responses_request
+# ===========================================================================
+
+def test_extract_model_from_responses() -> None:
+    """Model name is extracted from Responses API request."""
+    from routstr.proxy import extract_model_from_responses_request
+
+    body = {"model": "gpt-4o", "input": "test"}
+    model = extract_model_from_responses_request(body)
+    assert model == "gpt-4o"
+
+
+def test_extract_model_returns_unknown_for_missing() -> None:
+    """Missing model field returns 'unknown'."""
+    from routstr.proxy import extract_model_from_responses_request
+
+    body = {"input": "test"}
+    model = extract_model_from_responses_request(body)
+    assert model == "unknown"
+
+
+def test_extract_model_empty_body_returns_unknown() -> None:
+    """Empty body returns 'unknown'."""
+    from routstr.proxy import extract_model_from_responses_request
+
+    model = extract_model_from_responses_request({})
+    assert model == "unknown"
+
+
+def test_extract_model_from_input_nested() -> None:
+    """Model nested in input dict is found."""
+    from routstr.proxy import extract_model_from_responses_request
+
+    body = {"input": {"model": "claude-sonnet", "text": "hi"}}
+    model = extract_model_from_responses_request(body)
+    # The function checks input_data.get("model") for nested
+    assert model in ("claude-sonnet", "unknown")
+
+
+# ===========================================================================
+# get_model_instance / get_provider_for_model / get_unique_models
+# ===========================================================================
+
+def test_get_model_instance_unknown_returns_none() -> None:
+    """Unknown model ID returns None."""
+    from routstr.proxy import get_model_instance
+
+    result = get_model_instance("nonexistent-model-xyz-12345")
+    assert result is None
+
+
+def test_get_provider_for_model_unknown_returns_none() -> None:
+    """Unknown model returns None."""
+    from routstr.proxy import get_provider_for_model
+
+    result = get_provider_for_model("nonexistent-model-xyz-12345")
+    assert result is None
+
+
+def test_get_unique_models_returns_list() -> None:
+    """get_unique_models always returns a list."""
+    from routstr.proxy import get_unique_models
+
+    result = get_unique_models()
+    assert isinstance(result, list)
+
+
+def test_get_upstreams_returns_list() -> None:
+    """get_upstreams returns a list of providers."""
+    from routstr.proxy import get_upstreams
+
+    result = get_upstreams()
+    assert isinstance(result, list)
+
+
+# ===========================================================================
+# parse_request_body_json — nested objects
+# ===========================================================================
+
+def test_parse_body_preserves_nested_objects() -> None:
+    """Nested JSON objects are preserved during parsing."""
+    from routstr.proxy import parse_request_body_json
+
+    body = json.dumps({
+        "model": "claude-3",
+        "messages": [{"role": "system", "content": "You are helpful."}],
+        "temperature": 0.7,
+        "max_tokens": 1024,
+    }).encode()
+
+    result = parse_request_body_json(body, "/v1/chat/completions")
+    assert result["temperature"] == 0.7
+    assert result["max_tokens"] == 1024
+    assert len(result["messages"]) == 1

tests/unit/test_ehbp_finalize_payment.py

@@ -0,0 +1,185 @@
+from __future__ import annotations
+
+from typing import AsyncGenerator
+
+import pytest
+from sqlalchemy.ext.asyncio import AsyncEngine, create_async_engine
+from sqlalchemy.pool import StaticPool
+from sqlmodel import SQLModel, select
+from sqlmodel.ext.asyncio.session import AsyncSession
+
+from routstr.core.db import ApiKey
+from routstr.upstream.ehbp import (
+    finalize_ehbp_actual_cost_payment,
+    finalize_ehbp_max_cost_payment,
+)
+
+
+def _make_engine() -> AsyncEngine:
+    return create_async_engine(
+        "sqlite+aiosqlite://",
+        poolclass=StaticPool,
+        connect_args={"check_same_thread": False},
+    )
+
+
+@pytest.fixture
+async def session(monkeypatch: pytest.MonkeyPatch) -> AsyncGenerator[AsyncSession, None]:
+    monkeypatch.setattr("routstr.upstream.ehbp.ROUTSTR_FEE_PERCENT", 0)
+    engine = _make_engine()
+    async with engine.begin() as conn:
+        await conn.run_sync(SQLModel.metadata.create_all)
+    db_session = AsyncSession(engine, expire_on_commit=False)
+    try:
+        yield db_session
+    finally:
+        await db_session.close()
+        await engine.dispose()
+
+
+async def _api_key(session: AsyncSession, hashed_key: str) -> ApiKey | None:
+    return (
+        await session.exec(select(ApiKey).where(ApiKey.hashed_key == hashed_key))
+    ).one_or_none()
+
+
+@pytest.mark.asyncio
+async def test_finalize_actual_cost_payment_updates_balance_and_releases_reserve(
+    session: AsyncSession,
+) -> None:
+    key = ApiKey(
+        hashed_key="ehbp-actual",
+        balance=10_000,
+        reserved_balance=3_000,
+        reserved_at=123,
+    )
+    session.add(key)
+    await session.commit()
+
+    await finalize_ehbp_actual_cost_payment(
+        key,
+        session,
+        reserved_cost_for_model=3_000,
+        model_id="tinfoil/model",
+        cost_info={
+            "total_msats": 1_200,
+            "input_tokens": 10,
+            "output_tokens": 20,
+            "input_msats": 500,
+            "output_msats": 700,
+        },
+    )
+
+    updated = await _api_key(session, "ehbp-actual")
+    assert updated is not None
+    assert updated.balance == 8_800
+    assert updated.reserved_balance == 0
+    assert updated.reserved_at is None
+    assert updated.total_spent == 1_200
+
+
+@pytest.mark.asyncio
+async def test_finalize_max_cost_payment_updates_parent_and_child_spend(
+    session: AsyncSession,
+) -> None:
+    parent = ApiKey(
+        hashed_key="ehbp-parent",
+        balance=10_000,
+        reserved_balance=3_000,
+        reserved_at=123,
+    )
+    child = ApiKey(
+        hashed_key="ehbp-child",
+        balance=0,
+        reserved_balance=3_000,
+        reserved_at=123,
+        parent_key_hash="ehbp-parent",
+    )
+    session.add(parent)
+    session.add(child)
+    await session.commit()
+
+    await finalize_ehbp_max_cost_payment(
+        child,
+        session,
+        max_cost_for_model=3_000,
+        model_id="tinfoil/model",
+    )
+
+    updated_parent = await _api_key(session, "ehbp-parent")
+    updated_child = await _api_key(session, "ehbp-child")
+    assert updated_parent is not None
+    assert updated_child is not None
+    assert updated_parent.balance == 7_000
+    assert updated_parent.reserved_balance == 0
+    assert updated_parent.reserved_at is None
+    assert updated_parent.total_spent == 3_000
+    assert updated_child.balance == 0
+    assert updated_child.reserved_balance == 0
+    assert updated_child.reserved_at is None
+    assert updated_child.total_spent == 3_000
+
+
+@pytest.mark.asyncio
+async def test_finalize_actual_cost_payment_rolls_back_when_parent_update_matches_no_rows(
+    session: AsyncSession,
+) -> None:
+    key = ApiKey(
+        hashed_key="ehbp-missing-parent",
+        balance=10_000,
+        reserved_balance=3_000,
+        reserved_at=123,
+    )
+    session.add(key)
+    await session.commit()
+    await session.delete(key)
+    await session.commit()
+
+    await finalize_ehbp_actual_cost_payment(
+        key,
+        session,
+        reserved_cost_for_model=3_000,
+        model_id="tinfoil/model",
+        cost_info={"total_msats": 1_200},
+    )
+
+    assert await _api_key(session, "ehbp-missing-parent") is None
+
+
+@pytest.mark.asyncio
+async def test_finalize_max_cost_payment_rolls_back_parent_when_child_update_matches_no_rows(
+    session: AsyncSession,
+) -> None:
+    parent = ApiKey(
+        hashed_key="ehbp-rollback-parent",
+        balance=10_000,
+        reserved_balance=3_000,
+        reserved_at=123,
+    )
+    child = ApiKey(
+        hashed_key="ehbp-missing-child",
+        balance=0,
+        reserved_balance=3_000,
+        reserved_at=123,
+        parent_key_hash="ehbp-rollback-parent",
+    )
+    session.add(parent)
+    session.add(child)
+    await session.commit()
+    await session.delete(child)
+    await session.commit()
+
+    await finalize_ehbp_max_cost_payment(
+        child,
+        session,
+        max_cost_for_model=3_000,
+        model_id="tinfoil/model",
+    )
+
+    updated_parent = await _api_key(session, "ehbp-rollback-parent")
+    assert updated_parent is not None
+    assert updated_parent.balance == 10_000
+    assert updated_parent.reserved_balance == 3_000
+    assert updated_parent.reserved_at == 123
+    assert updated_parent.total_spent == 0
+    assert await _api_key(session, "ehbp-missing-child") is None

tests/unit/test_fee_payout_crash_safety.py

@@ -0,0 +1,160 @@
+import asyncio
+from collections.abc import AsyncIterator
+from contextlib import asynccontextmanager
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, Mock, patch
+
+import pytest
+from sqlalchemy.ext.asyncio import create_async_engine
+from sqlmodel import SQLModel
+from sqlmodel.ext.asyncio.session import AsyncSession
+
+from routstr import wallet
+from routstr.core import db
+
+
+@asynccontextmanager
+async def _session_context(session: Mock) -> AsyncIterator[Mock]:
+    yield session
+
+
+@pytest.mark.asyncio
+async def test_fee_payout_checkpoint_is_atomic_and_durable() -> None:
+    engine = create_async_engine("sqlite+aiosqlite://")
+    async with engine.begin() as connection:
+        await connection.run_sync(SQLModel.metadata.create_all)
+
+    async with AsyncSession(engine) as session:
+        session.add(db.RoutstrFee(id=1, accumulated_msats=5_000))
+        await session.commit()
+
+        assert await db.reset_routstr_fee(session, 5_000) is True
+        assert await db.reset_routstr_fee(session, 5_000) is False
+
+        fee = await db.get_routstr_fee(session)
+        await session.refresh(fee)
+        assert fee.accumulated_msats == 0
+        assert fee.payout_in_progress_msats == 5_000
+        assert fee.total_paid_msats == 0
+
+        assert await db.complete_routstr_fee_payout(session, 5_000) is True
+        await session.refresh(fee)
+        assert fee.payout_in_progress_msats == 0
+        assert fee.total_paid_msats == 5_000
+        assert fee.last_paid_at is not None
+
+    await engine.dispose()
+
+
+@pytest.mark.asyncio
+async def test_fee_payout_checkpoints_before_sending() -> None:
+    session = Mock()
+    fee = SimpleNamespace(
+        accumulated_msats=5_000,
+        payout_in_progress_msats=0,
+        payout_started_at=None,
+    )
+    payout_wallet = Mock()
+    events: list[str] = []
+
+    async def checkpoint(*_args: object) -> bool:
+        events.append("checkpoint")
+        return True
+
+    async def send(*_args: object, **_kwargs: object) -> int:
+        events.append("send")
+        return 5
+
+    async def complete(*_args: object) -> bool:
+        events.append("complete")
+        return True
+
+    with (
+        patch("routstr.auth.ROUTSTR_FEE_DEFAULT_PAYOUT", 1),
+        patch("routstr.auth.ROUTSTR_FEE_PAYOUT_INTERVAL_SECONDS", 1),
+        patch("routstr.auth.ROUTSTR_LN_ADDRESS", "fees@example.com"),
+        patch(
+            "routstr.wallet.asyncio.sleep",
+            AsyncMock(side_effect=[None, asyncio.CancelledError()]),
+        ),
+        patch("routstr.wallet.db.create_session", return_value=_session_context(session)),
+        patch("routstr.wallet.db.get_routstr_fee", AsyncMock(return_value=fee)),
+        patch("routstr.wallet.db.reset_routstr_fee", side_effect=checkpoint),
+        patch("routstr.wallet.db.complete_routstr_fee_payout", side_effect=complete),
+        patch("routstr.wallet.get_wallet", AsyncMock(return_value=payout_wallet)),
+        patch("routstr.wallet.get_proofs_per_mint_and_unit", return_value=[]),
+        patch("routstr.wallet.raw_send_to_lnurl", side_effect=send),
+    ):
+        with pytest.raises(asyncio.CancelledError):
+            await wallet.periodic_routstr_fee_payout()
+
+    assert events == ["checkpoint", "send", "complete"]
+
+
+@pytest.mark.asyncio
+async def test_fee_payout_does_not_retry_an_unresolved_checkpoint() -> None:
+    session = Mock()
+    fee = SimpleNamespace(
+        accumulated_msats=10_000,
+        payout_in_progress_msats=5_000,
+        payout_started_at=123,
+    )
+
+    with (
+        patch("routstr.auth.ROUTSTR_FEE_PAYOUT_INTERVAL_SECONDS", 1),
+        patch("routstr.auth.ROUTSTR_LN_ADDRESS", "fees@example.com"),
+        patch(
+            "routstr.wallet.asyncio.sleep",
+            AsyncMock(side_effect=[None, asyncio.CancelledError()]),
+        ),
+        patch("routstr.wallet.db.create_session", return_value=_session_context(session)),
+        patch("routstr.wallet.db.get_routstr_fee", AsyncMock(return_value=fee)),
+        patch("routstr.wallet.db.reset_routstr_fee", AsyncMock()) as checkpoint,
+        patch("routstr.wallet.get_wallet", AsyncMock()) as get_wallet,
+        patch("routstr.wallet.raw_send_to_lnurl", AsyncMock()) as send,
+        patch("routstr.wallet.logger.critical") as critical,
+    ):
+        with pytest.raises(asyncio.CancelledError):
+            await wallet.periodic_routstr_fee_payout()
+
+    checkpoint.assert_not_awaited()
+    get_wallet.assert_not_awaited()
+    send.assert_not_awaited()
+    critical.assert_called_once()
+
+
+@pytest.mark.asyncio
+async def test_fee_payout_keeps_checkpoint_when_send_outcome_is_unknown() -> None:
+    session = Mock()
+    fee = SimpleNamespace(
+        accumulated_msats=5_000,
+        payout_in_progress_msats=0,
+        payout_started_at=None,
+    )
+    complete = AsyncMock()
+
+    with (
+        patch("routstr.auth.ROUTSTR_FEE_DEFAULT_PAYOUT", 1),
+        patch("routstr.auth.ROUTSTR_FEE_PAYOUT_INTERVAL_SECONDS", 1),
+        patch("routstr.auth.ROUTSTR_LN_ADDRESS", "fees@example.com"),
+        patch(
+            "routstr.wallet.asyncio.sleep",
+            AsyncMock(side_effect=[None, asyncio.CancelledError()]),
+        ),
+        patch("routstr.wallet.db.create_session", return_value=_session_context(session)),
+        patch("routstr.wallet.db.get_routstr_fee", AsyncMock(return_value=fee)),
+        patch("routstr.wallet.db.reset_routstr_fee", AsyncMock(return_value=True)),
+        patch("routstr.wallet.db.complete_routstr_fee_payout", complete),
+        patch("routstr.wallet.get_wallet", AsyncMock(return_value=Mock())),
+        patch("routstr.wallet.get_proofs_per_mint_and_unit", return_value=[]),
+        patch(
+            "routstr.wallet.raw_send_to_lnurl",
+            AsyncMock(side_effect=TimeoutError("unknown outcome")),
+        ),
+        patch("routstr.wallet.logger.critical") as critical,
+    ):
+        with pytest.raises(asyncio.CancelledError):
+            await wallet.periodic_routstr_fee_payout()
+
+    complete.assert_not_awaited()
+    critical.assert_called_once()

tests/unit/test_fee_payout_migration.py

@@ -0,0 +1,46 @@
+import os
+import sqlite3
+import subprocess
+import sys
+from pathlib import Path
+
+
+def _run_alembic(root: Path, database_url: str, revision: str) -> None:
+    env = os.environ.copy()
+    env["DATABASE_URL"] = database_url
+    subprocess.run(
+        [sys.executable, "-m", "alembic", "upgrade", revision],
+        cwd=root,
+        env=env,
+        check=True,
+        capture_output=True,
+        text=True,
+    )
+
+
+def test_fee_payout_checkpoint_migration_preserves_existing_row(
+    tmp_path: Path,
+) -> None:
+    root = Path(__file__).resolve().parents[2]
+    database_path = tmp_path / "migration.db"
+    database_url = f"sqlite+aiosqlite:///{database_path}"
+    _run_alembic(root, database_url, "c6d7e8f9a0b1")
+
+    with sqlite3.connect(database_path) as connection:
+        result = connection.execute(
+            "UPDATE routstr_fees SET accumulated_msats = 5000, "
+            "total_paid_msats = 1000, last_paid_at = 123 WHERE id = 1"
+        )
+        assert result.rowcount == 1
+        connection.commit()
+
+    _run_alembic(root, database_url, "head")
+
+    with sqlite3.connect(database_path) as connection:
+        row = connection.execute(
+            "SELECT accumulated_msats, total_paid_msats, last_paid_at, "
+            "payout_in_progress_msats, payout_started_at "
+            "FROM routstr_fees WHERE id = 1"
+        ).fetchone()
+
+    assert row == (5000, 1000, 123, 0, None)

tests/unit/test_fetch_all_balances.py

@@ -0,0 +1,73 @@
+from contextlib import asynccontextmanager
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+from routstr.wallet import fetch_all_balances
+
+
+@asynccontextmanager
+async def _fake_session():  # type: ignore[no-untyped-def]
+    yield MagicMock()
+
+
+def _patches(proof_amount: int = 1000):  # type: ignore[no-untyped-def]
+    proof = MagicMock(amount=proof_amount)
+    return [
+        patch("routstr.wallet.get_wallet", AsyncMock(return_value=MagicMock())),
+        patch(
+            "routstr.wallet.get_proofs_per_mint_and_unit",
+            MagicMock(return_value=[proof]),
+        ),
+        patch(
+            "routstr.wallet.slow_filter_spend_proofs",
+            AsyncMock(side_effect=lambda proofs, wallet: proofs),
+        ),
+        patch(
+            "routstr.wallet.db.balances_for_mint_and_unit",
+            AsyncMock(return_value=0),
+        ),
+        patch("routstr.wallet.db.create_session", _fake_session),
+    ]
+
+
+@pytest.mark.asyncio
+async def test_fetch_all_balances_falls_back_to_primary_mint() -> None:
+    """With empty cashu_mints, balances are still fetched for primary_mint."""
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "cashu_mints", []), patch.object(
+        settings, "primary_mint", "http://primary:3338"
+    ):
+        for p in _patches(proof_amount=1000):
+            p.start()
+        try:
+            details, total_wallet, total_user, owner = await fetch_all_balances(
+                units=["sat"]
+            )
+        finally:
+            patch.stopall()
+
+    assert [d["mint_url"] for d in details] == ["http://primary:3338"]
+    assert total_wallet == 1000
+
+
+@pytest.mark.asyncio
+async def test_fetch_all_balances_no_duplicate_primary_mint() -> None:
+    """primary_mint already in cashu_mints is not inspected twice."""
+    from routstr.core.settings import settings
+
+    with patch.object(
+        settings, "cashu_mints", ["http://primary:3338"]
+    ), patch.object(settings, "primary_mint", "http://primary:3338"):
+        for p in _patches(proof_amount=1000):
+            p.start()
+        try:
+            details, total_wallet, _total_user, _owner = await fetch_all_balances(
+                units=["sat"]
+            )
+        finally:
+            patch.stopall()
+
+    assert [d["mint_url"] for d in details] == ["http://primary:3338"]
+    assert total_wallet == 1000

tests/unit/test_lnurl_melt_timeout.py

@@ -0,0 +1,74 @@
+"""raw_send_to_lnurl() must not hang forever on an unresponsive mint.
+
+The Cashu library issues POST /v1/melt/bolt11 with timeout=None, so a hung
+mint would block the melt (and the payout loop) indefinitely. raw_send_to_lnurl
+now wraps wallet.melt() in asyncio.wait_for(MELT_TIMEOUT_SECONDS) and surfaces a
+timeout as LNURLError instead of hanging.
+"""
+
+import asyncio
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+from routstr.payment import lnurl
+from routstr.payment.lnurl import LNURLError, raw_send_to_lnurl
+
+
+@pytest.mark.asyncio
+async def test_raw_send_to_lnurl_times_out_on_hung_melt() -> None:
+    proofs = [MagicMock(amount=1000)]
+
+    wallet = MagicMock()
+    wallet.melt_quote = AsyncMock(return_value=MagicMock(fee_reserve=1, quote="q"))
+    wallet.select_to_send = AsyncMock(return_value=(proofs, None))
+
+    async def _hang(**kwargs: object) -> None:
+        await asyncio.sleep(5)  # far longer than the patched timeout
+
+    wallet.melt = AsyncMock(side_effect=_hang)
+
+    lnurl_data = {
+        "callback_url": "https://ln.tld/cb",
+        "min_sendable": 1_000,
+        "max_sendable": 100_000_000,
+    }
+
+    with patch.object(lnurl, "MELT_TIMEOUT_SECONDS", 0.05), patch(
+        "routstr.payment.lnurl.get_lnurl_data", AsyncMock(return_value=lnurl_data)
+    ), patch(
+        "routstr.payment.lnurl.get_lnurl_invoice",
+        AsyncMock(return_value=("lnbc1...", {})),
+    ):
+        with pytest.raises(LNURLError, match="Melt timed out"):
+            await raw_send_to_lnurl(wallet, proofs, "owner@ln.tld", "sat", amount=1000)
+
+
+@pytest.mark.asyncio
+async def test_raw_send_to_lnurl_succeeds_within_timeout() -> None:
+    """A prompt melt still returns the net amount, unaffected by the guard."""
+    proofs = [MagicMock(amount=1000)]
+
+    wallet = MagicMock()
+    wallet.melt_quote = AsyncMock(return_value=MagicMock(fee_reserve=1, quote="q"))
+    wallet.select_to_send = AsyncMock(return_value=(proofs, None))
+    wallet.melt = AsyncMock(return_value=MagicMock())
+
+    lnurl_data = {
+        "callback_url": "https://ln.tld/cb",
+        "min_sendable": 1_000,
+        "max_sendable": 100_000_000,
+    }
+
+    with patch.object(lnurl, "MELT_TIMEOUT_SECONDS", 5), patch(
+        "routstr.payment.lnurl.get_lnurl_data", AsyncMock(return_value=lnurl_data)
+    ), patch(
+        "routstr.payment.lnurl.get_lnurl_invoice",
+        AsyncMock(return_value=("lnbc1...", {})),
+    ):
+        paid = await raw_send_to_lnurl(
+            wallet, proofs, "owner@ln.tld", "sat", amount=1000
+        )
+
+    assert paid > 0
+    wallet.melt.assert_awaited_once()

tests/unit/test_messages_litellm_dispatch.py

@@ -11,6 +11,7 @@ import os
 from typing import Any, AsyncIterator
 from unittest.mock import AsyncMock, MagicMock, patch
 
+import httpx
 import pytest
 from fastapi.responses import Response, StreamingResponse
 
@@ -21,6 +22,7 @@ from routstr.core.db import ApiKey  # noqa: E402
 from routstr.payment.cost_calculation import CostData  # noqa: E402
 from routstr.payment.models import Architecture, Model, Pricing  # noqa: E402
 from routstr.upstream.base import BaseUpstreamProvider  # noqa: E402
+from routstr.wallet import MintConnectionError, TokenConsumedError  # noqa: E402
 
 # ---------------------------------------------------------------------------
 # Fixtures
@@ -500,7 +502,7 @@ async def test_streaming_emits_sse_and_reconciles_cost_at_end() -> None:
     captured_cost_call: dict[str, Any] = {}
 
     async def fake_adjust(
-        fresh_key: Any, combined_data: Any, sess: Any, max_cost: int
+        fresh_key: Any, combined_data: Any, sess: Any, max_cost: int, usage: Any = None
     ) -> dict:
         captured_cost_call["combined_data"] = combined_data
         captured_cost_call["max_cost"] = max_cost
@@ -589,7 +591,7 @@ async def test_streaming_handles_iterator_yielding_raw_sse_bytes() -> None:
     captured: dict[str, Any] = {}
 
     async def fake_adjust(
-        fresh_key: Any, combined_data: Any, sess: Any, max_cost: int
+        fresh_key: Any, combined_data: Any, sess: Any, max_cost: int, usage: Any = None
     ) -> dict:
         captured["combined_data"] = combined_data
         return fake_cost
@@ -1308,3 +1310,320 @@ async def test_dispatch_uses_url_detected_prefix_for_fireworks_custom_row() -> N
         "fireworks_ai/accounts/fireworks/models/glm-5"
     )
     assert captured_kwargs["api_base"] == "https://api.fireworks.ai/inference/v1"
+
+
+# ---------------------------------------------------------------------------
+# X-Cashu redemption error taxonomy (unreachable mint + string codes)
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "handler_name",
+    ["handle_x_cashu", "handle_x_cashu_responses"],
+)
+@pytest.mark.parametrize(
+    "error",
+    [
+        httpx.ConnectError("All connection attempts failed"),
+        MintConnectionError("Cashu mint is unreachable"),
+        TimeoutError("timed out connecting to mint"),
+    ],
+)
+async def test_x_cashu_mint_unreachable_returns_503(
+    handler_name: str, error: Exception
+) -> None:
+    """Both X-Cashu entrypoints classify a down mint as 503 mint_unreachable,
+    not a generic 400 cashu_error."""
+    provider = _make_provider()
+    model = _make_model()
+    request = _make_request()
+
+    with patch(
+        "routstr.upstream.base.recieve_token", new=AsyncMock(side_effect=error)
+    ):
+        handler = getattr(provider, handler_name)
+        response = await handler(
+            request=request,
+            x_cashu_token="cashuAtoken",
+            path="v1/chat/completions",
+            max_cost_for_model=10_000,
+            model_obj=model,
+        )
+
+    assert response.status_code == 503
+    body = json.loads(bytes(response.body))
+    assert body["error"]["type"] == "mint_unreachable"
+    assert body["error"]["message"] == "Cashu mint is unreachable"
+    assert body["error"]["code"] == "cashu_mint_unreachable"
+    if str(error) != body["error"]["message"]:
+        assert str(error) not in body["error"]["message"]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "handler_name",
+    ["handle_x_cashu", "handle_x_cashu_responses"],
+)
+@pytest.mark.parametrize(
+    (
+        "error",
+        "expected_status",
+        "expected_type",
+        "expected_message",
+        "expected_code",
+    ),
+    [
+        (
+            ValueError("Mint Error: Token already spent"),
+            400,
+            "token_already_spent",
+            "Cashu token already spent",
+            "cashu_token_already_spent",
+        ),
+        (
+            ValueError("invalid token: could not decode"),
+            400,
+            "invalid_token",
+            "Invalid Cashu token",
+            "invalid_cashu_token",
+        ),
+        (
+            # Fee/swap failures now map to a granular 422 on the X-Cashu path,
+            # matching the bearer path (previously flattened to 400).
+            ValueError(
+                "Failed to estimate fees: Fees (7 sat) exceed token amount (5 sat)"
+            ),
+            422,
+            "mint_error",
+            "Token value is too small to cover swap fees",
+            "cashu_token_swap_fees_exceed_amount",
+        ),
+        (
+            ValueError("Failed to melt token from foreign mint http://m: boom"),
+            422,
+            "mint_error",
+            "Failed to swap token from foreign mint",
+            "cashu_foreign_mint_swap_failed",
+        ),
+        (
+            ValueError("some unexpected wallet condition"),
+            400,
+            "cashu_error",
+            "Failed to redeem Cashu token",
+            "cashu_token_redemption_failed",
+        ),
+        (
+            # Non-ValueError faults are internal errors (500), not token errors.
+            RuntimeError("db exploded"),
+            500,
+            "api_error",
+            "Internal error during token redemption",
+            "internal_error",
+        ),
+    ],
+)
+async def test_x_cashu_error_code_is_stable_string(
+    handler_name: str,
+    error: Exception,
+    expected_status: int,
+    expected_type: str,
+    expected_message: str,
+    expected_code: str,
+) -> None:
+    """X-Cashu emits a stable string ``code`` on every branch, matching the
+    bearer path's taxonomy instead of an int HTTP status."""
+    provider = _make_provider()
+    model = _make_model()
+    request = _make_request()
+
+    with patch(
+        "routstr.upstream.base.recieve_token", new=AsyncMock(side_effect=error)
+    ):
+        handler = getattr(provider, handler_name)
+        response = await handler(
+            request=request,
+            x_cashu_token="cashuAtoken",
+            path="v1/chat/completions",
+            max_cost_for_model=10_000,
+            model_obj=model,
+        )
+
+    assert response.status_code == expected_status
+    body = json.loads(bytes(response.body))
+    assert body["error"]["type"] == expected_type
+    assert body["error"]["message"] == expected_message
+    assert body["error"]["code"] == expected_code
+    assert isinstance(body["error"]["code"], str)
+    if str(error) != expected_message:
+        assert str(error) not in body["error"]["message"]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("handler_name", "forward_attr"),
+    [
+        ("handle_x_cashu", "forward_x_cashu_request"),
+        ("handle_x_cashu_responses", "forward_x_cashu_responses_request"),
+    ],
+)
+async def test_x_cashu_transport_error_after_redemption_is_not_retryable(
+    handler_name: str, forward_attr: str
+) -> None:
+    """A transport failure while forwarding (after the token is spent) maps to
+    502 upstream_error, never a retryable cashu_mint_unreachable."""
+    provider = _make_provider()
+    model = _make_model()
+    request = _make_request()
+
+    with (
+        patch(
+            "routstr.upstream.base.recieve_token",
+            new=AsyncMock(return_value=(5_000, "sat", "https://mint")),
+        ),
+        patch("routstr.upstream.base.store_cashu_transaction", new=AsyncMock()),
+        patch.object(
+            provider,
+            forward_attr,
+            new=AsyncMock(side_effect=httpx.ConnectError("upstream down")),
+        ),
+    ):
+        handler = getattr(provider, handler_name)
+        response = await handler(
+            request=request,
+            x_cashu_token="cashuAtoken",
+            path="v1/chat/completions",
+            max_cost_for_model=10_000,
+            model_obj=model,
+        )
+
+    assert response.status_code == 502
+    body = json.loads(bytes(response.body))
+    assert body["error"]["type"] == "upstream_error"
+    assert body["error"]["code"] != "cashu_mint_unreachable"
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "handler_name",
+    ["handle_x_cashu", "handle_x_cashu_responses"],
+)
+async def test_x_cashu_token_consumed_returns_500_and_no_echo(
+    handler_name: str,
+) -> None:
+    """A post-redemption failure (token spent, crediting/minting failed) is a
+    non-retryable 500 token_consumed and must NOT echo the spent token back."""
+    provider = _make_provider()
+    model = _make_model()
+    request = _make_request()
+
+    with patch(
+        "routstr.upstream.base.recieve_token",
+        new=AsyncMock(side_effect=TokenConsumedError("credit failed")),
+    ):
+        handler = getattr(provider, handler_name)
+        response = await handler(
+            request=request,
+            x_cashu_token="cashuAtoken",
+            path="v1/chat/completions",
+            max_cost_for_model=10_000,
+            model_obj=model,
+        )
+
+    assert response.status_code == 500
+    body = json.loads(bytes(response.body))
+    assert body["error"]["type"] == "token_consumed"
+    assert body["error"]["code"] == "cashu_token_consumed"
+    assert "X-Cashu" not in response.headers
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("handler_name", "error", "echoed"),
+    [
+        # Spent token: must NOT be echoed.
+        ("handle_x_cashu", ValueError("Token already spent"), False),
+        ("handle_x_cashu_responses", ValueError("Token already spent"), False),
+        # Unspent but unreachable mint: echo so the client can retry the token.
+        ("handle_x_cashu", MintConnectionError("mint down"), True),
+        ("handle_x_cashu_responses", MintConnectionError("mint down"), True),
+        # Consumed token (post-redemption): must NOT be echoed.
+        ("handle_x_cashu", TokenConsumedError("credit failed"), False),
+        ("handle_x_cashu_responses", TokenConsumedError("credit failed"), False),
+    ],
+)
+async def test_x_cashu_echoes_token_only_when_recoverable(
+    handler_name: str, error: Exception, echoed: bool
+) -> None:
+    provider = _make_provider()
+    model = _make_model()
+    request = _make_request()
+
+    with patch(
+        "routstr.upstream.base.recieve_token", new=AsyncMock(side_effect=error)
+    ):
+        handler = getattr(provider, handler_name)
+        response = await handler(
+            request=request,
+            x_cashu_token="cashuAtoken",
+            path="v1/chat/completions",
+            max_cost_for_model=10_000,
+            model_obj=model,
+        )
+
+    if echoed:
+        assert response.headers.get("X-Cashu") == "cashuAtoken"
+    else:
+        assert "X-Cashu" not in response.headers
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("handler_name", "forward_attr"),
+    [
+        ("handle_x_cashu", "forward_x_cashu_request"),
+        ("handle_x_cashu_responses", "forward_x_cashu_responses_request"),
+    ],
+)
+@pytest.mark.parametrize("amount", [0, -5])
+async def test_x_cashu_zero_value_rejected_not_forwarded(
+    handler_name: str, forward_attr: str, amount: int
+) -> None:
+    """A token that redeems to <= 0 must be rejected as cashu_token_zero_value
+    (400) and NEVER forwarded as a free request — the X-Cashu path lacked the
+    guard that credit_balance has."""
+    provider = _make_provider()
+    model = _make_model()
+    request = _make_request()
+
+    with (
+        patch(
+            "routstr.upstream.base.recieve_token",
+            new=AsyncMock(return_value=(amount, "sat", "https://mint")),
+        ),
+        patch("routstr.upstream.base.store_cashu_transaction", new=AsyncMock()),
+        patch.object(
+            provider,
+            forward_attr,
+            new=AsyncMock(side_effect=AssertionError("must not forward a zero-value token")),
+        ),
+    ):
+        handler = getattr(provider, handler_name)
+        response = await handler(
+            request=request,
+            x_cashu_token="cashuAtoken",
+            path="v1/chat/completions",
+            max_cost_for_model=10_000,
+            model_obj=model,
+        )
+
+    assert response.status_code == 400
+    body = json.loads(bytes(response.body))
+    assert body["error"]["type"] == "cashu_error"
+    assert (
+        body["error"]["message"]
+        == "Failed to redeem Cashu token: token yielded no value"
+    )
+    assert body["error"]["code"] == "cashu_token_zero_value"
+    # Spent-to-zero token must not be echoed back for retry.
+    assert "X-Cashu" not in response.headers

tests/unit/test_periodic_payout.py

@@ -0,0 +1,154 @@
+"""Tests for periodic_payout() resilience fixes.
+
+Covers two regressions from the auto-payout / primary-mint audit
+(docs/auto-payout-primary-mint-failure-report.md):
+
+1. periodic_payout() must include settings.primary_mint even when it is not
+   listed in settings.cashu_mints, matching fetch_all_balances(); otherwise
+   primary-mint funds never auto-payout.
+2. A failure on one mint/unit must not abort payout for the remaining
+   mint/units in the same cycle (the try/except is now per mint/unit).
+"""
+
+from collections.abc import Callable, Coroutine
+from contextlib import asynccontextmanager
+from typing import Any
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+from routstr.wallet import periodic_payout
+
+# Sentinel interval used to break the otherwise-infinite payout loop after
+# exactly one full cycle.
+_INTERVAL = 987
+
+
+class _LoopBreak(Exception):
+    """Raised via the patched sleep to stop periodic_payout after one cycle."""
+
+
+@asynccontextmanager
+async def _fake_session():  # type: ignore[no-untyped-def]
+    yield MagicMock()
+
+
+def _one_cycle_sleep() -> Callable[[float], Coroutine[Any, Any, None]]:
+    """Return an async sleep stub that lets exactly one payout cycle run.
+
+    The top-of-loop sleep uses the sentinel interval; the second time it is
+    seen (start of the second cycle) we raise to break out. The inner
+    ``asyncio.sleep(5)`` pass-through is ignored.
+    """
+    seen = {"interval": 0}
+
+    async def _sleep(seconds: float) -> None:
+        if seconds == _INTERVAL:
+            seen["interval"] += 1
+            if seen["interval"] >= 2:
+                raise _LoopBreak()
+
+    return _sleep
+
+
+@pytest.mark.asyncio
+async def test_periodic_payout_includes_primary_mint_not_in_cashu_mints() -> None:
+    """primary_mint absent from cashu_mints is still paid out."""
+    from routstr.core.settings import settings
+
+    get_wallet = AsyncMock(return_value=MagicMock())
+    raw_send = AsyncMock(return_value=1000)
+
+    with patch.object(settings, "cashu_mints", []), patch.object(
+        settings, "primary_mint", "http://primary:3338"
+    ), patch.object(settings, "receive_ln_address", "owner@ln.tld"), patch.object(
+        settings, "payout_interval_seconds", _INTERVAL
+    ), patch.object(settings, "min_payout_sat", 10), patch(
+        "routstr.wallet.asyncio.sleep", _one_cycle_sleep()
+    ), patch("routstr.wallet.db.create_session", _fake_session), patch(
+        "routstr.wallet.get_wallet", get_wallet
+    ), patch(
+        "routstr.wallet.get_proofs_per_mint_and_unit",
+        MagicMock(return_value=[MagicMock(amount=100_000)]),
+    ), patch(
+        "routstr.wallet.slow_filter_spend_proofs",
+        AsyncMock(side_effect=lambda proofs, wallet: proofs),
+    ), patch(
+        "routstr.wallet.db.balances_for_mint_and_unit", AsyncMock(return_value=0)
+    ), patch("routstr.wallet.raw_send_to_lnurl", raw_send):
+        with pytest.raises(_LoopBreak):
+            await periodic_payout()
+
+    processed = {call.args[0] for call in get_wallet.await_args_list}
+    assert processed == {"http://primary:3338"}
+    assert raw_send.await_count >= 1
+
+
+@pytest.mark.asyncio
+async def test_periodic_payout_isolates_failing_mint() -> None:
+    """A failing mint does not prevent payout for the other mints."""
+    from routstr.core.settings import settings
+
+    async def _get_wallet(mint_url: str, unit: str) -> MagicMock:
+        if mint_url == "http://bad:3338":
+            raise RuntimeError("mint unreachable")
+        return MagicMock()
+
+    get_wallet = AsyncMock(side_effect=_get_wallet)
+    raw_send = AsyncMock(return_value=1000)
+
+    with patch.object(
+        settings, "cashu_mints", ["http://bad:3338", "http://good:3338"]
+    ), patch.object(settings, "primary_mint", "http://good:3338"), patch.object(
+        settings, "receive_ln_address", "owner@ln.tld"
+    ), patch.object(settings, "payout_interval_seconds", _INTERVAL), patch.object(
+        settings, "min_payout_sat", 10
+    ), patch("routstr.wallet.asyncio.sleep", _one_cycle_sleep()), patch(
+        "routstr.wallet.db.create_session", _fake_session
+    ), patch("routstr.wallet.get_wallet", get_wallet), patch(
+        "routstr.wallet.get_proofs_per_mint_and_unit",
+        MagicMock(return_value=[MagicMock(amount=100_000)]),
+    ), patch(
+        "routstr.wallet.slow_filter_spend_proofs",
+        AsyncMock(side_effect=lambda proofs, wallet: proofs),
+    ), patch(
+        "routstr.wallet.db.balances_for_mint_and_unit", AsyncMock(return_value=0)
+    ), patch("routstr.wallet.raw_send_to_lnurl", raw_send):
+        with pytest.raises(_LoopBreak):
+            await periodic_payout()
+
+    # The bad mint raised on get_wallet for both units, yet the good mint was
+    # still reached and paid out for both units — failures are isolated.
+    good_calls = [
+        c for c in get_wallet.await_args_list if c.args[0] == "http://good:3338"
+    ]
+    assert len(good_calls) == 2  # sat + msat
+    assert raw_send.await_count == 2  # good mint paid for both units
+
+
+@pytest.mark.asyncio
+async def test_periodic_payout_handles_session_creation_failure() -> None:
+    """A db.create_session failure is logged and the payout loop continues."""
+    from routstr.core.settings import settings
+
+    create_session = MagicMock(side_effect=RuntimeError("db unavailable"))
+    logger = MagicMock()
+
+    with patch.object(settings, "cashu_mints", ["http://mint:3338"]), patch.object(
+        settings, "primary_mint", "http://mint:3338"
+    ), patch.object(settings, "receive_ln_address", "owner@ln.tld"), patch.object(
+        settings, "payout_interval_seconds", _INTERVAL
+    ), patch(
+        "routstr.wallet.asyncio.sleep", _one_cycle_sleep()
+    ), patch(
+        "routstr.wallet.db.create_session", create_session
+    ), patch("routstr.wallet.logger", logger):
+        with pytest.raises(_LoopBreak):
+            await periodic_payout()
+
+    create_session.assert_called_once()
+    logger.error.assert_called_once()
+    message = logger.error.call_args.args[0]
+    extra = logger.error.call_args.kwargs["extra"]
+    assert message == "Error in periodic payout cycle: RuntimeError"
+    assert extra == {"error": "db unavailable"}

tests/unit/test_periodic_refund_sweep.py

@@ -0,0 +1,126 @@
+from contextlib import asynccontextmanager
+from typing import AsyncGenerator
+from unittest.mock import AsyncMock
+
+import pytest
+from sqlalchemy.ext.asyncio import AsyncEngine, create_async_engine
+from sqlalchemy.pool import StaticPool
+from sqlmodel import SQLModel, select
+from sqlmodel.ext.asyncio.session import AsyncSession
+
+from routstr import wallet
+from routstr.core.db import CashuTransaction
+
+
+def _make_engine() -> AsyncEngine:
+    return create_async_engine(
+        "sqlite+aiosqlite://",
+        poolclass=StaticPool,
+        connect_args={"check_same_thread": False},
+    )
+
+
+@pytest.fixture
+async def refund_db(
+    monkeypatch: pytest.MonkeyPatch,
+) -> AsyncGenerator[AsyncEngine, None]:
+    engine = _make_engine()
+    async with engine.begin() as connection:
+        await connection.run_sync(SQLModel.metadata.create_all)
+
+    @asynccontextmanager
+    async def create_session() -> AsyncGenerator[AsyncSession, None]:
+        async with AsyncSession(engine, expire_on_commit=False) as session:
+            yield session
+
+    monkeypatch.setattr(wallet.db, "create_session", create_session)
+    try:
+        yield engine
+    finally:
+        await engine.dispose()
+
+
+async def _store(
+    engine: AsyncEngine,
+    *,
+    token: str,
+    created_at: int,
+    typ: str = "out",
+    collected: bool = False,
+    swept: bool = False,
+) -> None:
+    async with AsyncSession(engine, expire_on_commit=False) as session:
+        session.add(
+            CashuTransaction(
+                token=token,
+                amount=10,
+                unit="sat",
+                type=typ,
+                created_at=created_at,
+                collected=collected,
+                swept=swept,
+            )
+        )
+        await session.commit()
+
+
+async def _transactions(engine: AsyncEngine) -> dict[str, CashuTransaction]:
+    async with AsyncSession(engine, expire_on_commit=False) as session:
+        results = await session.exec(select(CashuTransaction))
+        return {transaction.token: transaction for transaction in results.all()}
+
+
+@pytest.mark.asyncio
+async def test_refund_sweep_only_processes_eligible_outbound_transactions(
+    refund_db: AsyncEngine, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    cutoff = 1_000
+    await _store(refund_db, token="eligible", created_at=cutoff - 1)
+    await _store(refund_db, token="too-new", created_at=cutoff)
+    await _store(
+        refund_db, token="collected", created_at=cutoff - 1, collected=True
+    )
+    await _store(refund_db, token="swept", created_at=cutoff - 1, swept=True)
+    await _store(refund_db, token="inbound", created_at=cutoff - 1, typ="in")
+    receive_token = AsyncMock()
+    monkeypatch.setattr(wallet, "recieve_token", receive_token)
+
+    await wallet._refund_sweep_once(cutoff)
+
+    receive_token.assert_awaited_once_with("eligible")
+    transactions = await _transactions(refund_db)
+    assert transactions["eligible"].swept is True
+    assert transactions["too-new"].swept is False
+    assert transactions["collected"].collected is True
+    assert transactions["collected"].swept is False
+    assert transactions["swept"].swept is True
+    assert transactions["inbound"].swept is False
+
+
+@pytest.mark.asyncio
+async def test_refund_sweep_persists_success_and_isolates_token_failures(
+    refund_db: AsyncEngine, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    cutoff = 1_000
+    for token in ("success", "already-spent", "temporary-failure"):
+        await _store(refund_db, token=token, created_at=cutoff - 1)
+
+    async def receive_token(token: str) -> None:
+        if token == "already-spent":
+            raise RuntimeError("Token already spent")
+        if token == "temporary-failure":
+            raise RuntimeError("mint unavailable")
+
+    receive = AsyncMock(side_effect=receive_token)
+    monkeypatch.setattr(wallet, "recieve_token", receive)
+
+    await wallet._refund_sweep_once(cutoff)
+
+    assert receive.await_count == 3
+    transactions = await _transactions(refund_db)
+    assert transactions["success"].swept is True
+    assert transactions["success"].collected is False
+    assert transactions["already-spent"].collected is True
+    assert transactions["already-spent"].swept is False
+    assert transactions["temporary-failure"].collected is False
+    assert transactions["temporary-failure"].swept is False

tests/unit/test_provider_slug_migration.py

@@ -0,0 +1,55 @@
+from __future__ import annotations
+
+import importlib.util
+from pathlib import Path
+
+import sqlalchemy as sa
+
+_MIGRATION_PATH = (
+    Path(__file__).resolve().parents[2]
+    / "migrations"
+    / "versions"
+    / "c6d7e8f9a0b1_add_slug_to_upstream_providers.py"
+)
+_spec = importlib.util.spec_from_file_location("provider_slug_migration", _MIGRATION_PATH)
+assert _spec is not None and _spec.loader is not None
+migration = importlib.util.module_from_spec(_spec)
+_spec.loader.exec_module(migration)
+
+
+def test_slug_migration_backfill_uses_api_safe_deterministic_slugs() -> None:
+    engine = sa.create_engine("sqlite:///:memory:")
+    with engine.begin() as conn:
+        conn.execute(
+            sa.text(
+                "CREATE TABLE upstream_providers ("
+                "id INTEGER PRIMARY KEY, "
+                "provider_type VARCHAR NOT NULL, "
+                "slug VARCHAR NULL"
+                ")"
+            )
+        )
+        conn.execute(
+            sa.text(
+                "INSERT INTO upstream_providers (id, provider_type, slug) VALUES "
+                "(1, 'OpenAI Compatible', NULL), "
+                "(2, 'OpenAI Compatible', ''), "
+                "(3, '123', NULL), "
+                "(4, 'x', NULL), "
+                "(5, 'anthropic', 'anthropic')"
+            )
+        )
+
+        migration._backfill_provider_slugs(conn)
+
+        rows = conn.execute(
+            sa.text("SELECT id, slug FROM upstream_providers ORDER BY id")
+        ).all()
+
+    assert rows == [
+        (1, "openai-compatible"),
+        (2, "openai-compatible-2"),
+        (3, "provider-123"),
+        (4, "x-provider"),
+        (5, "anthropic"),
+    ]

tests/unit/test_provider_slugs.py

@@ -0,0 +1,144 @@
+from __future__ import annotations
+
+import pytest
+from sqlalchemy.ext.asyncio import create_async_engine
+from sqlmodel import SQLModel, select
+from sqlmodel.ext.asyncio.session import AsyncSession
+
+from routstr.core.admin import _get_upstream_provider_by_ref
+from routstr.core.db import UpstreamProviderRow
+from routstr.core.provider_slugs import (
+    allocate_unique_provider_slug,
+    provider_slug_base,
+)
+from routstr.upstream.helpers import _seed_providers_from_settings
+
+
+@pytest.mark.asyncio
+async def test_allocate_unique_provider_slug_is_deterministic_with_suffixes() -> None:
+    engine = create_async_engine("sqlite+aiosqlite:///:memory:")
+    async with engine.begin() as conn:
+        await conn.run_sync(SQLModel.metadata.create_all)
+
+    async with AsyncSession(engine) as session:
+        session.add(
+            UpstreamProviderRow(
+                slug="openai",
+                provider_type="openai",
+                base_url="https://api.openai.com/v1",
+                api_key="key-1",
+            )
+        )
+        await session.commit()
+
+        assert await allocate_unique_provider_slug(session, "openai") == "openai-2"
+        assert (
+            await allocate_unique_provider_slug(session, "openai", {"openai-2"})
+            == "openai-3"
+        )
+
+    await engine.dispose()
+
+
+def test_provider_slug_base_sanitizes_provider_type() -> None:
+    assert provider_slug_base("OpenAI Compatible") == "openai-compatible"
+    assert provider_slug_base("!!!") == "provider"
+    assert provider_slug_base("AI") == "ai-provider"
+    assert provider_slug_base("123") == "provider-123"
+
+
+@pytest.mark.asyncio
+async def test_provider_ref_lookup_accepts_existing_numeric_ids_and_slugs() -> None:
+    engine = create_async_engine("sqlite+aiosqlite:///:memory:")
+    async with engine.begin() as conn:
+        await conn.run_sync(SQLModel.metadata.create_all)
+
+    async with AsyncSession(engine) as session:
+        provider = UpstreamProviderRow(
+            slug="openai",
+            provider_type="openai",
+            base_url="https://api.openai.com/v1",
+            api_key="key-1",
+        )
+        session.add(provider)
+        await session.commit()
+        await session.refresh(provider)
+
+        by_id = await _get_upstream_provider_by_ref(session, str(provider.id))
+        by_slug = await _get_upstream_provider_by_ref(session, "openai")
+
+    assert by_id.id == provider.id
+    assert by_slug.id == provider.id
+
+    await engine.dispose()
+
+
+@pytest.mark.asyncio
+async def test_seed_providers_from_settings_sets_deterministic_slug(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    engine = create_async_engine("sqlite+aiosqlite:///:memory:")
+    async with engine.begin() as conn:
+        await conn.run_sync(SQLModel.metadata.create_all)
+
+    monkeypatch.setenv("OPENAI_API_KEY", "seeded-openai-key")
+
+    class SettingsStub:
+        chat_completions_api_version: str | None = None
+        upstream_base_url: str | None = None
+        upstream_api_key: str = ""
+
+    async with AsyncSession(engine) as session:
+        await _seed_providers_from_settings(session, SettingsStub())  # type: ignore[arg-type]
+        await session.commit()
+
+        result = await session.exec(select(UpstreamProviderRow))
+        providers: list[UpstreamProviderRow] = list(result.all())
+
+    assert [(p.provider_type, p.slug) for p in providers] == [("openai", "openai")]
+
+    await engine.dispose()
+
+
+@pytest.mark.asyncio
+async def test_seed_providers_from_settings_keeps_slug_stable_on_reseed(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    engine = create_async_engine("sqlite+aiosqlite:///:memory:")
+    async with engine.begin() as conn:
+        await conn.run_sync(SQLModel.metadata.create_all)
+
+    monkeypatch.setenv("OPENAI_API_KEY", "seeded-openai-key")
+
+    class SettingsStub:
+        chat_completions_api_version: str | None = None
+        upstream_base_url: str | None = None
+        upstream_api_key: str = ""
+
+    async with AsyncSession(engine) as session:
+        session.add(
+            UpstreamProviderRow(
+                slug="openai",
+                provider_type="openai",
+                base_url="https://example.invalid/v1",
+                api_key="other-key",
+            )
+        )
+        await session.commit()
+
+        await _seed_providers_from_settings(session, SettingsStub())  # type: ignore[arg-type]
+        await session.commit()
+        await _seed_providers_from_settings(session, SettingsStub())  # type: ignore[arg-type]
+        await session.commit()
+
+        result = await session.exec(
+            select(UpstreamProviderRow).order_by(UpstreamProviderRow.slug)
+        )
+        providers: list[UpstreamProviderRow] = list(result.all())
+
+    assert [(p.provider_type, p.slug) for p in providers] == [
+        ("openai", "openai"),
+        ("openai", "openai-2"),
+    ]
+
+    await engine.dispose()

tests/unit/test_proxy_tinfoil_attestation_routing.py

@@ -0,0 +1,141 @@
+from __future__ import annotations
+
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+from fastapi import FastAPI
+from fastapi.responses import Response
+from httpx import ASGITransport, AsyncClient
+
+from routstr import proxy as proxy_module
+
+
+@pytest.fixture
+def proxy_app() -> FastAPI:
+    app = FastAPI()
+    app.include_router(proxy_module.proxy_router)
+    return app
+
+
+@pytest.mark.asyncio
+async def test_attestation_get_routes_directly_to_tinfoil_provider(
+    monkeypatch: pytest.MonkeyPatch, proxy_app: FastAPI
+) -> None:
+    non_tinfoil = MagicMock()
+    non_tinfoil.provider_type = "openai"
+    non_tinfoil.prepare_headers = MagicMock(return_value={})
+    non_tinfoil.forward_get_request = AsyncMock(
+        return_value=Response(status_code=404, content=b"wrong upstream")
+    )
+
+    tinfoil = MagicMock()
+    tinfoil.provider_type = "tinfoil"
+    tinfoil.prepare_headers = MagicMock(return_value={"accept": "application/json"})
+    tinfoil.forward_get_request = AsyncMock(
+        return_value=Response(status_code=200, content=b'{"attestation":true}')
+    )
+
+    monkeypatch.setattr(proxy_module, "_upstreams", [non_tinfoil, tinfoil])
+
+    async with AsyncClient(
+        transport=ASGITransport(app=proxy_app),  # type: ignore[arg-type]
+        base_url="http://test",
+    ) as client:
+        response = await client.get("/attestation")
+
+    assert response.status_code == 200
+    assert response.content == b'{"attestation":true}'
+    non_tinfoil.forward_get_request.assert_not_called()
+    tinfoil.forward_get_request.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_tee_attestation_get_routes_directly_to_tinfoil_provider(
+    monkeypatch: pytest.MonkeyPatch, proxy_app: FastAPI
+) -> None:
+    non_tinfoil = MagicMock()
+    non_tinfoil.provider_type = "openrouter"
+    non_tinfoil.prepare_headers = MagicMock(return_value={})
+    non_tinfoil.forward_get_request = AsyncMock(
+        return_value=Response(status_code=404, content=b"wrong upstream")
+    )
+
+    tinfoil = MagicMock()
+    tinfoil.provider_type = "tinfoil"
+    tinfoil.prepare_headers = MagicMock(return_value={"accept": "application/json"})
+    tinfoil.forward_get_request = AsyncMock(
+        return_value=Response(status_code=200, content=b'{"tee":true}')
+    )
+
+    monkeypatch.setattr(proxy_module, "_upstreams", [non_tinfoil, tinfoil])
+
+    async with AsyncClient(
+        transport=ASGITransport(app=proxy_app),  # type: ignore[arg-type]
+        base_url="http://test",
+    ) as client:
+        response = await client.get("/tee/attestation")
+
+    assert response.status_code == 200
+    assert response.content == b'{"tee":true}'
+    non_tinfoil.forward_get_request.assert_not_called()
+    tinfoil.forward_get_request.assert_awaited_once()
+
+
+@pytest.mark.parametrize("path", ["attestation/", "tee/attestation/"])
+@pytest.mark.asyncio
+async def test_attestation_trailing_slash_routes_directly_to_tinfoil(
+    monkeypatch: pytest.MonkeyPatch, proxy_app: FastAPI, path: str
+) -> None:
+    tinfoil = MagicMock()
+    tinfoil.provider_type = "tinfoil"
+    tinfoil.prepare_headers = MagicMock(return_value={})
+    tinfoil.forward_get_request = AsyncMock(return_value=Response(status_code=200))
+    monkeypatch.setattr(proxy_module, "_upstreams", [tinfoil])
+
+    async with AsyncClient(
+        transport=ASGITransport(app=proxy_app),  # type: ignore[arg-type]
+        base_url="http://test",
+    ) as client:
+        response = await client.get(f"/{path}")
+
+    assert response.status_code == 200
+    tinfoil.forward_get_request.assert_awaited_once()
+
+
+@pytest.mark.parametrize("path", ["attestation/foo", "attestationjunk"])
+@pytest.mark.asyncio
+async def test_non_attestation_prefix_does_not_bypass_authentication(
+    monkeypatch: pytest.MonkeyPatch, proxy_app: FastAPI, path: str
+) -> None:
+    tinfoil = MagicMock()
+    tinfoil.provider_type = "tinfoil"
+    tinfoil.forward_get_request = AsyncMock()
+    monkeypatch.setattr(proxy_module, "_upstreams", [tinfoil])
+
+    async with AsyncClient(
+        transport=ASGITransport(app=proxy_app),  # type: ignore[arg-type]
+        base_url="http://test",
+    ) as client:
+        response = await client.get(f"/{path}")
+
+    assert response.status_code == 400
+    assert response.json()["error"]["type"] == "invalid_model"
+    tinfoil.forward_get_request.assert_not_awaited()
+
+
+def test_attestation_upstream_selection_is_tinfoil_only() -> None:
+    non_tinfoil = MagicMock(provider_type="openai")
+    tinfoil = MagicMock(provider_type="tinfoil")
+
+    assert proxy_module._select_unauthenticated_get_upstreams(
+        "attestation", [non_tinfoil, tinfoil]
+    ) == [tinfoil]
+    assert proxy_module._select_unauthenticated_get_upstreams(
+        "tee/attestation", [non_tinfoil, tinfoil]
+    ) == [tinfoil]
+    assert proxy_module._select_unauthenticated_get_upstreams(
+        "attestation/", [non_tinfoil, tinfoil]
+    ) == [tinfoil]
+    assert proxy_module._select_unauthenticated_get_upstreams(
+        "attestationjunk", [non_tinfoil, tinfoil]
+    ) == [non_tinfoil, tinfoil]

tests/unit/test_refund_sweep.py

@@ -0,0 +1,118 @@
+from collections.abc import AsyncIterator
+from pathlib import Path
+from unittest.mock import AsyncMock, patch
+
+import pytest
+from sqlalchemy.ext.asyncio import async_sessionmaker, create_async_engine
+from sqlmodel import SQLModel, select
+from sqlmodel.ext.asyncio.session import AsyncSession
+
+from routstr.core.db import CashuTransaction
+from routstr.wallet import refund_sweep_once
+
+
+@pytest.fixture
+async def session_factory(
+    tmp_path: Path,
+) -> AsyncIterator[async_sessionmaker[AsyncSession]]:
+    engine = create_async_engine(f"sqlite+aiosqlite:///{tmp_path / 'refunds.db'}")
+    async with engine.begin() as connection:
+        await connection.run_sync(SQLModel.metadata.create_all)
+    factory = async_sessionmaker(engine, class_=AsyncSession, expire_on_commit=False)
+    yield factory
+    await engine.dispose()
+
+
+async def _insert(
+    factory: async_sessionmaker[AsyncSession], *rows: CashuTransaction
+) -> None:
+    async with factory() as session:
+        session.add_all(rows)
+        await session.commit()
+
+
+async def _load(
+    factory: async_sessionmaker[AsyncSession],
+) -> dict[str, CashuTransaction]:
+    async with factory() as session:
+        result = await session.exec(select(CashuTransaction))
+        return {row.token: row for row in result.all()}
+
+
+@pytest.mark.asyncio
+async def test_refund_sweep_only_processes_expired_eligible_outgoing_tokens(
+    session_factory: async_sessionmaker[AsyncSession],
+) -> None:
+    rows = [
+        CashuTransaction(
+            token="eligible", amount=1, unit="sat", type="out", created_at=800
+        ),
+        CashuTransaction(
+            token="fresh", amount=1, unit="sat", type="out", created_at=950
+        ),
+        CashuTransaction(
+            token="boundary", amount=1, unit="sat", type="out", created_at=900
+        ),
+        CashuTransaction(
+            token="collected",
+            amount=1,
+            unit="sat",
+            type="out",
+            created_at=800,
+            collected=True,
+        ),
+        CashuTransaction(
+            token="swept", amount=1, unit="sat", type="out", created_at=800, swept=True
+        ),
+        CashuTransaction(
+            token="incoming", amount=1, unit="sat", type="in", created_at=800
+        ),
+    ]
+    await _insert(session_factory, *rows)
+    receive = AsyncMock()
+    with (
+        patch("routstr.wallet.db.create_session", side_effect=session_factory),
+        patch("routstr.wallet.settings.refund_sweep_ttl_seconds", 100),
+        patch("routstr.wallet.time.time", return_value=1000),
+        patch("routstr.wallet.recieve_token", receive),
+    ):
+        await refund_sweep_once()
+
+    receive.assert_awaited_once_with("eligible")
+    loaded = await _load(session_factory)
+    assert loaded["eligible"].swept is True
+    assert loaded["fresh"].swept is False
+    assert loaded["boundary"].swept is False
+    assert loaded["collected"].swept is False
+    assert loaded["swept"].swept is True
+    assert loaded["incoming"].swept is False
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("error", "collected"),
+    [
+        (RuntimeError("token already spent"), True),
+        (RuntimeError("mint unavailable"), False),
+    ],
+)
+async def test_refund_sweep_records_terminal_but_not_transient_failures(
+    session_factory: async_sessionmaker[AsyncSession], error: Exception, collected: bool
+) -> None:
+    await _insert(
+        session_factory,
+        CashuTransaction(
+            token="refund", amount=1, unit="sat", type="out", created_at=800
+        ),
+    )
+    with (
+        patch("routstr.wallet.db.create_session", side_effect=session_factory),
+        patch("routstr.wallet.settings.refund_sweep_ttl_seconds", 100),
+        patch("routstr.wallet.time.time", return_value=1000),
+        patch("routstr.wallet.recieve_token", AsyncMock(side_effect=error)),
+    ):
+        await refund_sweep_once()
+
+    refund = (await _load(session_factory))["refund"]
+    assert refund.collected is collected
+    assert refund.swept is False

tests/unit/test_streaming_sse_providers.py

@@ -337,3 +337,75 @@ async def test_multiline_non_json_data_each_line_prefixed() -> None:
             continue
         assert line.startswith(b"data: "), f"bare line leaked to client: {line!r}"
     assert b"data: line one" in blob and b"data: line two" in blob
+
+
+@pytest.mark.asyncio
+async def test_crlf_delimiter_split_across_chunk_boundary() -> None:
+    """CRLF event delimiter straddling two TCP reads must not merge events.
+
+    Regression: a per-chunk ``replace(b"\\r\\n", b"\\n")`` left a stray ``\\r``
+    when a ``\\r\\n`` of the ``\\r\\n\\r\\n`` delimiter landed at the very end of
+    one ``aiter_bytes`` chunk and the matching ``\\n`` opened the next. The
+    ``\\n\\n`` split then missed the boundary, glued two events into one frame
+    with two ``data:`` lines, and the client's ``JSON.parse`` threw on the
+    concatenated payload (the "unexpected token"/"Extra data" crash).
+    """
+    e1 = b'data: {"id":"x","choices":[{"delta":{"content":"a"}}]}'
+    e2 = b'data: {"id":"x","choices":[{"delta":{"content":"b"}}]}'
+    chunks = [
+        e1 + b"\r\n\r",  # delimiter cut mid-CRLF
+        b"\n" + e2 + b"\r\n\r\n",
+        b"data: [DONE]\r\n\r\n",
+    ]
+    out = await _drive(chunks)
+
+    # Client-accurate check: a real SSE client concatenates all ``data:`` lines
+    # *within one event* (events are ``\n\n``-delimited) before parsing. A
+    # merged frame would surface here as two objects glued into one payload,
+    # which ``_assert_clean`` (per-line) would miss.
+    blob = b"".join(out)
+    contents: list[str] = []
+    for event in blob.split(b"\n\n"):
+        datas = [
+            ln[len(b"data: ") :]
+            for ln in event.split(b"\n")
+            if ln.startswith(b"data: ")
+        ]
+        if not datas:
+            continue
+        payload = b"".join(datas)
+        if payload.strip() == b"[DONE]":
+            continue
+        obj = json.loads(payload)  # raises if two events were merged into one
+        for c in obj.get("choices", []):
+            if "delta" in c:
+                contents.append(c["delta"]["content"])
+    assert contents == ["a", "b"]
+
+
+@pytest.mark.asyncio
+async def test_truncated_json_tail_on_connection_close() -> None:
+    """A stream that drops mid-event must not emit the partial JSON downstream.
+
+    Regression: the end-of-stream flush ran ``_process_event`` on the leftover
+    buffer unconditionally. When the upstream connection closed mid-event the
+    leftover was incomplete JSON, which fell through to the raw-forward path and
+    handed the client a ``data: {partial`` frame -> ``Unterminated string`` parse
+    error. The truncated tail must be dropped instead.
+    """
+    chunks = [
+        b'data: {"id":"x","choices":[{"delta":{"content":"ok"}}]}\n\n',
+        b'data: {"id":"x","choices":[{"delta":{"con',  # connection dies here
+    ]
+    out = await _drive(chunks)
+    objs = _assert_clean(out)  # raises if the partial tail leaked as a data frame
+    contents = [
+        c["delta"]["content"]
+        for o in objs
+        for c in o.get("choices", [])
+        if "delta" in c
+    ]
+    # The one complete chunk is delivered; the truncated fragment is dropped
+    # entirely (no second delta), and _assert_clean above guarantees nothing
+    # non-JSON ever reached the client.
+    assert contents == ["ok"]

tests/unit/test_tinfoil_integration.py

@@ -0,0 +1,749 @@
+"""Unit tests for Tinfoil direct integration.
+
+Covers the EHBP usage-metrics header parser, the proxy header stripping, the
+enclave URL override, and the TinfoilUpstreamProvider model fetching/forwarding
+target logic.
+"""
+
+from __future__ import annotations
+
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+from routstr.upstream.ehbp import (
+    _PROXY_ONLY_HEADERS,
+    _compute_ehbp_actual_cost,
+    _prepare_ehbp_upstream_headers,
+    _resolve_ehbp_target_url,
+    _strip_proxy_headers,
+    parse_tinfoil_usage_metrics,
+)
+from routstr.upstream.tinfoil import (
+    TinfoilModel,
+    TinfoilUpstreamProvider,
+)
+
+# ---------------------------------------------------------------------------
+# parse_tinfoil_usage_metrics
+# ---------------------------------------------------------------------------
+
+
+class TestParseTinfoilUsageMetrics:
+    def test_full_header(self) -> None:
+        result = parse_tinfoil_usage_metrics("prompt=67,completion=42,total=109")
+        assert result == {
+            "prompt_tokens": 67,
+            "completion_tokens": 42,
+            "total_tokens": 109,
+        }
+
+    def test_without_total(self) -> None:
+        result = parse_tinfoil_usage_metrics("prompt=10,completion=5")
+        assert result == {"prompt_tokens": 10, "completion_tokens": 5}
+
+    def test_none(self) -> None:
+        assert parse_tinfoil_usage_metrics(None) is None
+
+    def test_empty(self) -> None:
+        assert parse_tinfoil_usage_metrics("") is None
+
+    def test_malformed(self) -> None:
+        assert parse_tinfoil_usage_metrics("garbage") is None
+
+    def test_missing_completion(self) -> None:
+        assert parse_tinfoil_usage_metrics("prompt=10") is None
+
+    def test_extra_whitespace(self) -> None:
+        result = parse_tinfoil_usage_metrics(
+            "prompt = 100 , completion = 200 , total = 300"
+        )
+        assert result == {
+            "prompt_tokens": 100,
+            "completion_tokens": 200,
+            "total_tokens": 300,
+        }
+
+    def test_with_model_field(self) -> None:
+        result = parse_tinfoil_usage_metrics(
+            "prompt=42,completion=10,total=52,model=llama3-3-70b"
+        )
+        assert result == {
+            "prompt_tokens": 42,
+            "completion_tokens": 10,
+            "total_tokens": 52,
+            "model": "llama3-3-70b",
+        }
+
+    def test_with_model_no_total(self) -> None:
+        result = parse_tinfoil_usage_metrics(
+            "prompt=67,completion=42,model=gpt-oss-120b"
+        )
+        assert result == {
+            "prompt_tokens": 67,
+            "completion_tokens": 42,
+            "model": "gpt-oss-120b",
+        }
+
+    def test_model_with_dashes_and_numbers(self) -> None:
+        result = parse_tinfoil_usage_metrics(
+            "prompt=1,completion=1,total=2,model=kimi-k2-6"
+        )
+        assert result is not None
+        assert result["model"] == "kimi-k2-6"
+
+    def test_model_with_extra_fields(self) -> None:
+        result = parse_tinfoil_usage_metrics(
+            "prompt=69,completion=20,total=89,"
+            "cached_prompt_tokens=64,uncached_prompt_tokens=5,"
+            "model=kimi-k2-6"
+        )
+        assert result is not None
+        assert result["prompt_tokens"] == 69
+        assert result["completion_tokens"] == 20
+        assert result["total_tokens"] == 89
+        assert result["model"] == "kimi-k2-6"
+
+    def test_old_format_still_works(self) -> None:
+        """Headers without the model field (pre-PR #385) still parse."""
+        result = parse_tinfoil_usage_metrics(
+            "prompt=67,completion=42,total=109"
+        )
+        assert result == {
+            "prompt_tokens": 67,
+            "completion_tokens": 42,
+            "total_tokens": 109,
+        }
+        assert "model" not in result
+
+
+# ---------------------------------------------------------------------------
+# _strip_proxy_headers
+# ---------------------------------------------------------------------------
+
+
+class TestStripProxyHeaders:
+    def test_strips_all_proxy_only(self) -> None:
+        headers = {
+            "x-routstr-model": "tinfoil-llama3-3-70b",
+            "X-Tinfoil-Enclave-Url": "https://inference.tinfoil.sh",
+            "X-Tinfoil-Request-Usage-Metrics": "true",
+            "Authorization": "Bearer secret",
+            "Ehbp-Encapsulated-Key": "abc123",
+        }
+        clean = _strip_proxy_headers(headers)
+        assert "x-routstr-model" not in clean
+        assert "X-Tinfoil-Enclave-Url" not in clean
+        assert "X-Tinfoil-Request-Usage-Metrics" not in clean
+        assert clean["Authorization"] == "Bearer secret"
+        assert clean["Ehbp-Encapsulated-Key"] == "abc123"
+
+    def test_all_proxy_only_headers_covered(self) -> None:
+        assert _PROXY_ONLY_HEADERS == {
+            "x-routstr-model",
+            "x-tinfoil-enclave-url",
+            "x-tinfoil-request-usage-metrics",
+        }
+
+
+class TestPrepareEHBPUpstreamHeaders:
+    def test_strips_client_proxy_headers_before_merging_target_headers(self) -> None:
+        headers = {
+            "x-routstr-model": "tinfoil-llama3-3-70b",
+            "X-Tinfoil-Enclave-Url": "https://enclave.tinfoil.sh",
+            "X-Tinfoil-Request-Usage-Metrics": "false",
+            "Authorization": "Bearer upstream-key",
+            "Ehbp-Encapsulated-Key": "abc123",
+        }
+        target_headers = {"X-Tinfoil-Request-Usage-Metrics": "true"}
+
+        clean = _prepare_ehbp_upstream_headers(headers, target_headers)
+
+        assert "x-routstr-model" not in clean
+        assert "X-Tinfoil-Enclave-Url" not in clean
+        assert clean["Authorization"] == "Bearer upstream-key"
+        assert clean["Ehbp-Encapsulated-Key"] == "abc123"
+        assert clean["X-Tinfoil-Request-Usage-Metrics"] == "true"
+
+
+# ---------------------------------------------------------------------------
+# _resolve_ehbp_target_url
+# ---------------------------------------------------------------------------
+
+
+class TestResolveEhbpTargetUrl:
+    def test_override_with_enclave_url_for_tinfoil(self) -> None:
+        result = _resolve_ehbp_target_url(
+            "https://default.example.com/v1/chat/completions",
+            "v1/chat/completions",
+            {"X-Tinfoil-Enclave-Url": "https://enclave.tinfoil.sh"},
+            "tinfoil",
+        )
+        assert result == "https://enclave.tinfoil.sh/v1/chat/completions"
+
+    def test_override_lowercase_header_for_tinfoil(self) -> None:
+        result = _resolve_ehbp_target_url(
+            "https://default.example.com/v1/chat/completions",
+            "v1/chat/completions",
+            {"x-tinfoil-enclave-url": "https://enclave.tinfoil.sh"},
+            "tinfoil",
+        )
+        assert result == "https://enclave.tinfoil.sh/v1/chat/completions"
+
+    def test_no_override(self) -> None:
+        default = "https://inference.tinfoil.sh/v1/chat/completions"
+        result = _resolve_ehbp_target_url(
+            default,
+            "v1/chat/completions",
+            {},
+            "tinfoil",
+        )
+        assert result == default
+
+    def test_non_tinfoil_provider_ignores_enclave_url(self) -> None:
+        default = "https://api.ppq.ai/private/v1/chat/completions"
+        result = _resolve_ehbp_target_url(
+            default,
+            "v1/chat/completions",
+            {"X-Tinfoil-Enclave-Url": "https://enclave.tinfoil.sh"},
+            "ppqai",
+        )
+        assert result == default
+
+    @pytest.mark.parametrize(
+        "bad_url",
+        [
+            "http://enclave.tinfoil.sh",
+            "https://attacker.example",
+            "https://tinfoil.sh.attacker.example",
+            "https://127.0.0.1",
+            "https://enclave.tinfoil.sh:8443",
+            "https://user:pass@enclave.tinfoil.sh",
+        ],
+    )
+    def test_tinfoil_rejects_unsafe_enclave_url(self, bad_url: str) -> None:
+        from routstr.core.exceptions import UpstreamError
+
+        with pytest.raises(UpstreamError):
+            _resolve_ehbp_target_url(
+                "https://default.example.com/v1/chat/completions",
+                "v1/chat/completions",
+                {"X-Tinfoil-Enclave-Url": bad_url},
+                "tinfoil",
+            )
+
+
+# ---------------------------------------------------------------------------
+# _compute_ehbp_actual_cost
+# ---------------------------------------------------------------------------
+
+
+class TestComputeEhbpActualCost:
+    @pytest.mark.asyncio
+    async def test_no_usage_falls_back_to_max_cost(self) -> None:
+        model_obj = MagicMock()
+        model_obj.id = "llama3-3-70b"
+        model_obj.forwarded_model_id = "llama3-3-70b"
+        result = await _compute_ehbp_actual_cost(None, model_obj, 100_000)
+        assert result["total_msats"] == 100_000
+        assert result["input_tokens"] == 0
+        assert result["output_tokens"] == 0
+
+    @pytest.mark.asyncio
+    async def test_usage_parsed_and_clamped(self) -> None:
+        model_obj = MagicMock()
+        model_obj.id = "llama3-3-70b"
+        model_obj.forwarded_model_id = "llama3-3-70b"
+        # The actual cost from calculate_cost will be small; we just verify
+        # it's clamped to min_request_msat at minimum.
+        with patch(
+            "routstr.upstream.ehbp.calculate_cost",
+            new_callable=AsyncMock,
+        ) as mock_calc:
+            from routstr.payment.cost_calculation import CostData
+
+            mock_calc.return_value = CostData(
+                base_msats=0,
+                input_msats=10,
+                output_msats=20,
+                total_msats=30,
+                total_usd=0.0001,
+                input_tokens=67,
+                output_tokens=42,
+            )
+            result = await _compute_ehbp_actual_cost(
+                "prompt=67,completion=42,total=109",
+                model_obj,
+                100_000,
+            )
+            assert result["total_msats"] == 30
+            assert result["total_msats"] <= 100_000
+            assert result["input_tokens"] == 67
+            assert result["output_tokens"] == 42
+            assert result["total_tokens"] == 109
+            assert result["input_msats"] == 10
+            assert result["output_msats"] == 20
+
+    @pytest.mark.asyncio
+    async def test_max_cost_data_falls_back(self) -> None:
+        model_obj = MagicMock()
+        model_obj.id = "llama3-3-70b"
+        model_obj.forwarded_model_id = "llama3-3-70b"
+        with patch(
+            "routstr.upstream.ehbp.calculate_cost",
+            new_callable=AsyncMock,
+        ) as mock_calc:
+            from routstr.payment.cost_calculation import MaxCostData
+
+            mock_calc.return_value = MaxCostData(
+                base_msats=0,
+                input_msats=0,
+                output_msats=0,
+                total_msats=0,
+                total_usd=0.0,
+                input_tokens=0,
+                output_tokens=0,
+            )
+            result = await _compute_ehbp_actual_cost(
+                "prompt=0,completion=0",
+                model_obj,
+                50_000,
+            )
+            assert result["total_msats"] == 50_000
+            assert result["input_tokens"] == 0
+            assert result["output_tokens"] == 0
+
+    @pytest.mark.asyncio
+    async def test_model_match_no_actual_model_key(self) -> None:
+        """When the served model matches the requested one, no actual_model key."""
+        model_obj = MagicMock()
+        model_obj.id = "llama3-3-70b"
+        model_obj.forwarded_model_id = "llama3-3-70b"
+        with patch(
+            "routstr.upstream.ehbp.calculate_cost",
+            new_callable=AsyncMock,
+        ) as mock_calc:
+            from routstr.payment.cost_calculation import CostData
+
+            mock_calc.return_value = CostData(
+                base_msats=0,
+                input_msats=5,
+                output_msats=10,
+                total_msats=15,
+                total_usd=0.0,
+                input_tokens=42,
+                output_tokens=10,
+            )
+            result = await _compute_ehbp_actual_cost(
+                "prompt=42,completion=10,total=52,model=llama3-3-70b",
+                model_obj,
+                100_000,
+            )
+            assert "actual_model" not in result
+            # calculate_cost called with requested model
+            call_args = mock_calc.call_args
+            assert call_args[0][0]["model"] == "llama3-3-70b"
+
+    @pytest.mark.asyncio
+    async def test_alias_match_no_actual_model_key(self) -> None:
+        """When the served upstream model matches forwarded_model_id through
+        a client-facing alias, no actual_model key is set."""
+        model_obj = MagicMock()
+        model_obj.id = "tinfoil-glm-5-2"  # client-facing alias
+        model_obj.forwarded_model_id = "glm-5-2"  # actual upstream ID
+        with patch(
+            "routstr.upstream.ehbp.calculate_cost",
+            new_callable=AsyncMock,
+        ) as mock_calc:
+            from routstr.payment.cost_calculation import CostData
+
+            mock_calc.return_value = CostData(
+                base_msats=0,
+                input_msats=5,
+                output_msats=10,
+                total_msats=15,
+                total_usd=0.0,
+                input_tokens=42,
+                output_tokens=10,
+            )
+            # Tinfoil header returns the actual upstream model ID
+            result = await _compute_ehbp_actual_cost(
+                "prompt=42,completion=10,total=52,model=glm-5-2",
+                model_obj,
+                100_000,
+            )
+            assert "actual_model" not in result
+            # calculate_cost called with the client-facing model ID (whose
+            # pricing includes the correct upstream rates)
+            call_args = mock_calc.call_args
+            assert call_args[0][0]["model"] == "tinfoil-glm-5-2"
+
+    @pytest.mark.asyncio
+    async def test_real_mismatch_uses_actual_model_for_pricing(self) -> None:
+        """When the served model differs from the expected upstream model,
+        the actual model's pricing is used."""
+        model_obj = MagicMock()
+        model_obj.id = "tinfoil-gpt-oss-120b"  # client-facing alias
+        model_obj.forwarded_model_id = "gpt-oss-120b"  # expected upstream
+
+        actual_model_obj = MagicMock()
+        actual_model_obj.id = "tinfoil-llama3-3-70b"  # client-facing of actual
+        actual_model_obj.forwarded_model_id = "llama3-3-70b"
+
+        with patch(
+            "routstr.proxy.get_model_instance",
+            return_value=actual_model_obj,
+        ), patch(
+            "routstr.upstream.ehbp.calculate_cost",
+            new_callable=AsyncMock,
+        ) as mock_calc:
+            from routstr.payment.cost_calculation import CostData
+
+            mock_calc.return_value = CostData(
+                base_msats=0,
+                input_msats=20,
+                output_msats=40,
+                total_msats=60,
+                total_usd=0.0,
+                input_tokens=42,
+                output_tokens=10,
+            )
+            # Tinfoil served llama3-3-70b instead of gpt-oss-120b
+            result = await _compute_ehbp_actual_cost(
+                "prompt=42,completion=10,total=52,model=llama3-3-70b",
+                model_obj,
+                100_000,
+            )
+            assert result["actual_model"] == "llama3-3-70b"
+            assert result["total_msats"] == 60
+            # calculate_cost called with the actual model's client-facing ID
+            call_args = mock_calc.call_args
+            assert call_args[0][0]["model"] == "tinfoil-llama3-3-70b"
+
+    @pytest.mark.asyncio
+    async def test_model_mismatch_unknown_model_falls_back(self) -> None:
+        """When the served model is not in the registry, use requested model."""
+        model_obj = MagicMock()
+        model_obj.id = "gpt-oss-120b"
+        model_obj.forwarded_model_id = "gpt-oss-120b"
+
+        with patch(
+            "routstr.proxy.get_model_instance",
+            return_value=None,
+        ), patch(
+            "routstr.upstream.ehbp.calculate_cost",
+            new_callable=AsyncMock,
+        ) as mock_calc:
+            from routstr.payment.cost_calculation import CostData
+
+            mock_calc.return_value = CostData(
+                base_msats=0,
+                input_msats=5,
+                output_msats=10,
+                total_msats=15,
+                total_usd=0.0,
+                input_tokens=42,
+                output_tokens=10,
+            )
+            result = await _compute_ehbp_actual_cost(
+                "prompt=42,completion=10,total=52,model=nonexistent",
+                model_obj,
+                100_000,
+            )
+            assert "actual_model" not in result
+            # calculate_cost called with the requested model (fallback)
+            call_args = mock_calc.call_args
+            assert call_args[0][0]["model"] == "gpt-oss-120b"
+
+    @pytest.mark.asyncio
+    async def test_old_format_no_model_uses_requested(self) -> None:
+        """Old format without model field uses requested model for pricing."""
+        model_obj = MagicMock()
+        model_obj.id = "llama3-3-70b"
+        model_obj.forwarded_model_id = "llama3-3-70b"
+        with patch(
+            "routstr.upstream.ehbp.calculate_cost",
+            new_callable=AsyncMock,
+        ) as mock_calc:
+            from routstr.payment.cost_calculation import CostData
+
+            mock_calc.return_value = CostData(
+                base_msats=0,
+                input_msats=5,
+                output_msats=10,
+                total_msats=15,
+                total_usd=0.0,
+                input_tokens=67,
+                output_tokens=42,
+            )
+            result = await _compute_ehbp_actual_cost(
+                "prompt=67,completion=42,total=109",
+                model_obj,
+                100_000,
+            )
+            assert "actual_model" not in result
+            call_args = mock_calc.call_args
+            assert call_args[0][0]["model"] == "llama3-3-70b"
+
+    @pytest.mark.asyncio
+    async def test_case_insensitive_model_match(self) -> None:
+        """Casing differences between the header and forwarded_model_id
+        should not trigger a spurious mismatch."""
+        model_obj = MagicMock()
+        model_obj.id = "tinfoil-glm-5-2"
+        model_obj.forwarded_model_id = "glm-5-2"  # lowercase
+        with patch(
+            "routstr.proxy.get_model_instance"
+        ) as mock_get_model, patch(
+            "routstr.upstream.ehbp.calculate_cost",
+            new_callable=AsyncMock,
+        ) as mock_calc:
+            from routstr.payment.cost_calculation import CostData
+
+            mock_calc.return_value = CostData(
+                base_msats=0,
+                input_msats=5,
+                output_msats=10,
+                total_msats=15,
+                total_usd=0.0,
+                input_tokens=42,
+                output_tokens=10,
+            )
+            # Header returns uppercase — same model, different casing
+            result = await _compute_ehbp_actual_cost(
+                "prompt=42,completion=10,total=52,model=GLM-5-2",
+                model_obj,
+                100_000,
+            )
+            assert "actual_model" not in result
+            # No mismatch: requested model pricing used
+            call_args = mock_calc.call_args
+            assert call_args[0][0]["model"] == "tinfoil-glm-5-2"
+            mock_get_model.assert_not_called()
+
+    @pytest.mark.asyncio
+    async def test_date_versioned_alias_resolves_to_requested(self) -> None:
+        """When the served model is a date-versioned alias that resolves back
+        to the requested model, no mismatch is propagated."""
+        model_obj = MagicMock()
+        model_obj.id = "tinfoil-glm-5-2"
+        model_obj.forwarded_model_id = "glm-5-2"
+
+        resolved_model_obj = MagicMock()
+        resolved_model_obj.id = "other-provider-glm-5-2"
+        resolved_model_obj.forwarded_model_id = "glm-5-2"
+
+        with patch(
+            "routstr.proxy.get_model_instance",
+            return_value=resolved_model_obj,
+        ) as mock_get_model, patch(
+            "routstr.upstream.ehbp.calculate_cost",
+            new_callable=AsyncMock,
+        ) as mock_calc:
+            from routstr.payment.cost_calculation import CostData
+
+            mock_calc.return_value = CostData(
+                base_msats=0,
+                input_msats=5,
+                output_msats=10,
+                total_msats=15,
+                total_usd=0.0,
+                input_tokens=42,
+                output_tokens=10,
+            )
+            # Tinfoil returns a date-versioned ID with different casing.
+            result = await _compute_ehbp_actual_cost(
+                "prompt=42,completion=10,total=52,model=GLM-5-2-20260415",
+                model_obj,
+                100_000,
+            )
+            # Registry resolution, rather than unconditional suffix removal,
+            # establishes that this alias represents the expected model.
+            assert "actual_model" not in result
+            assert mock_calc.call_args[0][0]["model"] == "tinfoil-glm-5-2"
+            mock_get_model.assert_called_once_with("GLM-5-2-20260415")
+
+    @pytest.mark.asyncio
+    async def test_configured_date_version_is_preserved_as_identity(self) -> None:
+        """A date suffix in forwarded_model_id is meaningful and preserved."""
+        model_obj = MagicMock()
+        model_obj.id = "tinfoil-glm-5-2-20260415"
+        model_obj.forwarded_model_id = "glm-5-2-20260415"
+
+        with patch(
+            "routstr.proxy.get_model_instance"
+        ) as mock_get_model, patch(
+            "routstr.upstream.ehbp.calculate_cost",
+            new_callable=AsyncMock,
+        ) as mock_calc:
+            from routstr.payment.cost_calculation import CostData
+
+            mock_calc.return_value = CostData(
+                base_msats=0,
+                input_msats=5,
+                output_msats=10,
+                total_msats=15,
+                total_usd=0.0,
+                input_tokens=42,
+                output_tokens=10,
+            )
+            result = await _compute_ehbp_actual_cost(
+                "prompt=42,completion=10,total=52,model=GLM-5-2-20260415",
+                model_obj,
+                100_000,
+            )
+
+            assert "actual_model" not in result
+            assert (
+                mock_calc.call_args[0][0]["model"]
+                == "tinfoil-glm-5-2-20260415"
+            )
+            mock_get_model.assert_not_called()
+
+    @pytest.mark.asyncio
+    async def test_different_client_alias_same_upstream_identity(self) -> None:
+        """A global alias winner from another provider is not a failover when
+        its forwarded model ID matches the requested upstream identity."""
+        model_obj = MagicMock()
+        model_obj.id = "tinfoil-glm-5-2"
+        model_obj.forwarded_model_id = "glm-5-2"
+
+        resolved_model_obj = MagicMock()
+        resolved_model_obj.id = "other-provider-glm-5-2"
+        resolved_model_obj.forwarded_model_id = "GLM-5-2"
+
+        with patch(
+            "routstr.proxy.get_model_instance",
+            return_value=resolved_model_obj,
+        ) as mock_get_model, patch(
+            "routstr.upstream.ehbp.calculate_cost",
+            new_callable=AsyncMock,
+        ) as mock_calc:
+            from routstr.payment.cost_calculation import CostData
+
+            mock_calc.return_value = CostData(
+                base_msats=0,
+                input_msats=5,
+                output_msats=10,
+                total_msats=15,
+                total_usd=0.0,
+                input_tokens=42,
+                output_tokens=10,
+            )
+            result = await _compute_ehbp_actual_cost(
+                "prompt=42,completion=10,total=52,model=provider-alias",
+                model_obj,
+                100_000,
+            )
+
+            mock_get_model.assert_called_once_with("provider-alias")
+            assert "actual_model" not in result
+            assert mock_calc.call_args[0][0]["model"] == "tinfoil-glm-5-2"
+
+
+# ---------------------------------------------------------------------------
+# TinfoilUpstreamProvider
+# ---------------------------------------------------------------------------
+
+
+class TestTinfoilUpstreamProvider:
+    def test_provider_type_and_defaults(self) -> None:
+        assert TinfoilUpstreamProvider.provider_type == "tinfoil"
+        assert (
+            TinfoilUpstreamProvider.default_base_url
+            == "https://inference.tinfoil.sh"
+        )
+        assert TinfoilUpstreamProvider.supports_ehbp is True
+
+    def test_transform_model_name(self) -> None:
+        provider = TinfoilUpstreamProvider(api_key="test")
+        assert provider.transform_model_name("tinfoil/llama3-3-70b") == "llama3-3-70b"
+        assert provider.transform_model_name("llama3-3-70b") == "llama3-3-70b"
+
+    def test_get_ehbp_forwarding_target_includes_usage_header(self) -> None:
+        provider = TinfoilUpstreamProvider(api_key="test")
+        model_obj = MagicMock()
+        model_obj.id = "llama3-3-70b"
+        model_obj.forwarded_model_id = "llama3-3-70b"
+        target = provider.get_ehbp_forwarding_target("v1/chat/completions", model_obj)
+        assert (
+            target.headers["X-Tinfoil-Request-Usage-Metrics"] == "true"
+        )
+        assert "v1/chat/completions" in target.url
+
+    def test_get_provider_metadata(self) -> None:
+        meta = TinfoilUpstreamProvider.get_provider_metadata()
+        assert meta["id"] == "tinfoil"
+        assert meta["name"] == "Tinfoil"
+        assert meta["fixed_base_url"] is True
+
+    def test_tinfoil_model_pricing_parses(self) -> None:
+        data = {
+            "id": "llama3-3-70b",
+            "context_window": 128000,
+            "created": 1721764788,
+            "pricing": {
+                "inputTokenPricePer1M": 1.75,
+                "outputTokenPricePer1M": 2.75,
+                "requestPrice": 0,
+            },
+            "endpoints": ["/v1/chat/completions"],
+            "type": "chat",
+        }
+        tf = TinfoilModel.parse_obj(data)
+        assert tf.id == "llama3-3-70b"
+        assert tf.pricing.inputTokenPricePer1M == 1.75
+        assert tf.pricing.outputTokenPricePer1M == 2.75
+
+    @pytest.mark.asyncio
+    async def test_fetch_models_parses_response(self) -> None:
+        provider = TinfoilUpstreamProvider(api_key="test")
+        mock_response = MagicMock()
+        mock_response.status_code = 200
+        mock_response.raise_for_status = MagicMock()
+        mock_response.json.return_value = {
+            "data": [
+                {
+                    "id": "llama3-3-70b",
+                    "context_window": 128000,
+                    "created": 1721764788,
+                    "multimodal": False,
+                    "pricing": {
+                        "inputTokenPricePer1M": 1.75,
+                        "outputTokenPricePer1M": 2.75,
+                        "requestPrice": 0,
+                    },
+                    "endpoints": ["/v1/chat/completions"],
+                    "type": "chat",
+                }
+            ]
+        }
+
+        with patch("routstr.upstream.tinfoil.httpx.AsyncClient") as mock_client_cls:
+            mock_client = MagicMock()
+            mock_client.__aenter__ = AsyncMock(return_value=mock_client)
+            mock_client.__aexit__ = AsyncMock(return_value=None)
+            mock_client.get = AsyncMock(return_value=mock_response)
+            mock_client_cls.return_value = mock_client
+
+            models = await provider.fetch_models()
+
+        assert len(models) == 1
+        assert models[0].id == "llama3-3-70b"
+        assert models[0].pricing.prompt == 1.75 / 1_000_000
+        assert models[0].pricing.completion == 2.75 / 1_000_000
+        assert models[0].context_length == 128000
+
+    @pytest.mark.asyncio
+    async def test_fetch_models_handles_error(self) -> None:
+        provider = TinfoilUpstreamProvider(api_key="test")
+        with patch("routstr.upstream.tinfoil.httpx.AsyncClient") as mock_client_cls:
+            mock_client = MagicMock()
+            mock_client.__aenter__ = AsyncMock(return_value=mock_client)
+            mock_client.__aexit__ = AsyncMock(return_value=None)
+            mock_client.get = AsyncMock(side_effect=Exception("network error"))
+            mock_client_cls.return_value = mock_client
+
+            models = await provider.fetch_models()
+
+        assert models == []

tests/unit/test_tinfoil_trailer.py

@@ -0,0 +1,138 @@
+from __future__ import annotations
+
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+
+from routstr.upstream.tinfoil_trailer import forward_with_trailer
+
+
+class FakeReader:
+    def __init__(self, chunks: list[bytes]) -> None:
+        self._chunks = chunks
+
+    async def read(self, _size: int) -> bytes:
+        if self._chunks:
+            return self._chunks.pop(0)
+        return b""
+
+
+class FakeWriter:
+    def __init__(self) -> None:
+        self.written = b""
+        self.drain = AsyncMock()
+        self.wait_closed = AsyncMock()
+        self.close = MagicMock()
+
+    def write(self, data: bytes) -> None:
+        self.written += data
+
+
+@pytest.mark.asyncio
+async def test_forward_with_trailer_captures_usage_trailer(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    response = (
+        b"HTTP/1.1 200 OK\r\n"
+        b"Transfer-Encoding: chunked\r\n"
+        b"Trailer: X-Tinfoil-Usage-Metrics\r\n"
+        b"\r\n"
+        b"5\r\nhello\r\n"
+        b"0\r\n"
+        b"X-Tinfoil-Usage-Metrics: prompt=1,completion=2,total=3\r\n"
+        b"\r\n"
+    )
+    reader = FakeReader([response])
+    writer = FakeWriter()
+    open_connection = AsyncMock(return_value=(reader, writer))
+    monkeypatch.setattr(
+        "routstr.upstream.tinfoil_trailer.asyncio.open_connection", open_connection
+    )
+
+    result = await forward_with_trailer(
+        method="POST",
+        url="https://enclave.tinfoil.sh/v1/chat/completions?stream=true",
+        headers={"Authorization": "Bearer upstream"},
+        body=b"opaque",
+    )
+
+    assert result.status_code == 200
+    assert result.body == b"hello"
+    assert result.trailers == [
+        ("x-tinfoil-usage-metrics", "prompt=1,completion=2,total=3")
+    ]
+    assert b"Connection: close" in writer.written
+    writer.close.assert_called_once()
+    writer.wait_closed.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_forward_with_trailer_strips_hop_by_hop_headers(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    response = b"HTTP/1.1 200 OK\r\nContent-Length: 2\r\n\r\nok"
+    reader = FakeReader([response])
+    writer = FakeWriter()
+    monkeypatch.setattr(
+        "routstr.upstream.tinfoil_trailer.asyncio.open_connection",
+        AsyncMock(return_value=(reader, writer)),
+    )
+
+    await forward_with_trailer(
+        method="POST",
+        url="https://enclave.tinfoil.sh/v1/chat/completions",
+        headers={
+            "Authorization": "Bearer upstream",
+            "Connection": "keep-alive, X-Client-Hop",
+            "Keep-Alive": "timeout=5",
+            "Proxy-Authenticate": "Basic",
+            "Proxy-Authorization": "Basic secret",
+            "TE": "trailers",
+            "Trailer": "X-Usage",
+            "Transfer-Encoding": "chunked",
+            "Upgrade": "websocket",
+            "X-Client-Hop": "remove-me",
+            "X-End-To-End": "preserve-me",
+        },
+        body=b"opaque",
+    )
+
+    serialized_headers = writer.written.split(b"\r\n\r\n", 1)[0].lower()
+    for name in (
+        b"keep-alive",
+        b"proxy-authenticate",
+        b"proxy-authorization",
+        b"te:",
+        b"trailer:",
+        b"transfer-encoding",
+        b"upgrade:",
+        b"x-client-hop",
+    ):
+        assert name not in serialized_headers
+    assert b"connection: close" in serialized_headers
+    assert b"content-length: 6" in serialized_headers
+    assert b"x-end-to-end: preserve-me" in serialized_headers
+
+
+@pytest.mark.asyncio
+async def test_forward_with_trailer_enforces_response_size_limit(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    response = b"HTTP/1.1 200 OK\r\nContent-Length: 5\r\n\r\nhello"
+    reader = FakeReader([response])
+    writer = FakeWriter()
+    monkeypatch.setattr(
+        "routstr.upstream.tinfoil_trailer.asyncio.open_connection",
+        AsyncMock(return_value=(reader, writer)),
+    )
+
+    with pytest.raises(ValueError, match="EHBP response exceeded"):
+        await forward_with_trailer(
+            method="POST",
+            url="https://enclave.tinfoil.sh/v1/chat/completions",
+            headers={},
+            body=b"opaque",
+            max_response_bytes=4,
+        )
+
+    writer.close.assert_called_once()

tests/unit/test_upstream_generic.py

@@ -0,0 +1,524 @@
+"""Unit tests for ``GenericUpstreamProvider.fetch_models`` price/metadata resolution.
+
+A generic upstream is any OpenAI-compatible API. Most (DeepSeek, OpenAI,
+Groq, ...) answer ``/models`` with bare ``{id, object, owned_by}`` entries that
+carry *no* pricing. The provider must not fabricate a price for those: it
+resolves through native ``model_spec`` (Venice's bespoke schema) → litellm's
+bundled cost map → the OpenRouter feed, and only when every source misses does
+it import the model **disabled** with a warning rather than invent a number.
+
+These tests drive that behaviour through the public ``fetch_models`` API. The
+``/models`` HTTP call is faked at ``httpx.AsyncClient``; the OpenRouter feed is
+patched at its source (``routstr.payment.models.async_fetch_openrouter_models``)
+so the resolver's lazy import picks up the stub. litellm's real bundled cost map
+is used unmocked — the DeepSeek rates it ships are the assertion's ground truth.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import Any
+from unittest.mock import AsyncMock, patch
+
+import pytest
+
+from routstr.upstream.generic import GenericUpstreamProvider
+
+
+class _FakeResponse:
+    def __init__(self, payload: dict[str, Any]) -> None:
+        self._payload = payload
+
+    def raise_for_status(self) -> None:
+        return None
+
+    def json(self) -> dict[str, Any]:
+        return self._payload
+
+
+class _FakeAsyncClient:
+    """Stand-in for ``httpx.AsyncClient`` returning a canned ``/models`` body."""
+
+    def __init__(self, payload: dict[str, Any]) -> None:
+        self._payload = payload
+
+    async def __aenter__(self) -> "_FakeAsyncClient":
+        return self
+
+    async def __aexit__(self, *exc: object) -> bool:
+        return False
+
+    async def get(self, url: str, headers: dict[str, str] | None = None) -> _FakeResponse:
+        return _FakeResponse(self._payload)
+
+
+def _patch_models_endpoint(payload: dict[str, Any]) -> Any:
+    """Patch the provider's ``httpx.AsyncClient`` to serve ``payload``."""
+    return patch(
+        "routstr.upstream.generic.httpx.AsyncClient",
+        lambda *args, **kwargs: _FakeAsyncClient(payload),
+    )
+
+
+def _model_by_id(models: list[Any], model_id: str) -> Any:
+    return next(m for m in models if m.id == model_id)
+
+
+# ---------------------------------------------------------------------------
+# native model_spec (Venice) — must keep resolving, and capture its metadata
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_native_model_spec_resolves_and_captures_metadata() -> None:
+    """A Venice-shaped ``model_spec`` is authoritative: prices/context come
+    straight from it and vision capability becomes an image input modality."""
+    payload = {
+        "data": [
+            {
+                "id": "venice-llama",
+                "owned_by": "venice",
+                "model_spec": {
+                    "name": "Venice Llama",
+                    "availableContextTokens": 65536,
+                    "pricing": {
+                        "input": {"usd": 0.5},
+                        "output": {"usd": 1.5},
+                    },
+                    "capabilities": {"supportsVision": True},
+                },
+            }
+        ]
+    }
+
+    with _patch_models_endpoint(payload):
+        or_feed = AsyncMock(return_value=[])
+        with patch(
+            "routstr.payment.models.async_fetch_openrouter_models", or_feed
+        ):
+            models = await GenericUpstreamProvider(base_url="http://x").fetch_models()
+
+    model = _model_by_id(models, "venice-llama")
+    assert model.enabled is True
+    assert model.pricing.prompt == pytest.approx(0.5 / 1_000_000)
+    assert model.pricing.completion == pytest.approx(1.5 / 1_000_000)
+    assert model.context_length == 65536
+    assert "image" in model.architecture.input_modalities
+    # Vision capability must be reflected in the combined modality string, not
+    # flattened to "text->text".
+    assert model.architecture.modality == "text+image->text"
+    # A native price never needs the OpenRouter feed.
+    or_feed.assert_not_awaited()
+
+
+# ---------------------------------------------------------------------------
+# native model_spec validation — a bogus native price is not authoritative
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_native_both_zero_price_falls_through_to_litellm() -> None:
+    """A native ``model_spec`` that prices both tokens at 0 is not a real price
+    (the same free-tier trap the litellm/OpenRouter rungs already reject). It
+    must not be treated as authoritative and served free; the resolver falls
+    through, so a litellm-known model lands on litellm's real rate instead."""
+    payload = {
+        "data": [
+            {
+                "id": "deepseek-chat",
+                "owned_by": "deepseek",
+                "model_spec": {
+                    "pricing": {"input": {"usd": 0}, "output": {"usd": 0}},
+                },
+            }
+        ]
+    }
+
+    with _patch_models_endpoint(payload):
+        or_feed = AsyncMock(return_value=[])
+        with patch("routstr.payment.models.async_fetch_openrouter_models", or_feed):
+            models = await GenericUpstreamProvider(base_url="http://x").fetch_models()
+
+    model = _model_by_id(models, "deepseek-chat")
+    assert model.enabled is True
+    assert model.pricing.prompt == pytest.approx(2.8e-07)
+    assert model.pricing.completion == pytest.approx(4.2e-07)
+
+
+@pytest.mark.asyncio
+async def test_native_negative_price_falls_through_to_litellm() -> None:
+    """A negative native price would credit the caller's balance on every
+    request (a fund drain, not a discount). Reject it like any other unusable
+    price and fall through to the chain."""
+    payload = {
+        "data": [
+            {
+                "id": "deepseek-chat",
+                "owned_by": "deepseek",
+                "model_spec": {
+                    "pricing": {"input": {"usd": -0.5}, "output": {"usd": -1.5}},
+                },
+            }
+        ]
+    }
+
+    with _patch_models_endpoint(payload):
+        or_feed = AsyncMock(return_value=[])
+        with patch("routstr.payment.models.async_fetch_openrouter_models", or_feed):
+            models = await GenericUpstreamProvider(base_url="http://x").fetch_models()
+
+    model = _model_by_id(models, "deepseek-chat")
+    assert model.enabled is True
+    assert model.pricing.prompt == pytest.approx(2.8e-07)
+    assert model.pricing.completion == pytest.approx(4.2e-07)
+
+
+@pytest.mark.asyncio
+async def test_native_non_numeric_price_does_not_break_catalog() -> None:
+    """A non-numeric native price (``"free"``) must not raise while parsing —
+    an unguarded ``"free" / 1_000_000`` throws and the outer catch drops the
+    *entire* provider catalog. It has to fail closed for that one model while
+    every other model in the same response still resolves."""
+    payload = {
+        "data": [
+            {
+                "id": "broken-price",
+                "owned_by": "mystery",
+                "model_spec": {
+                    "pricing": {"input": {"usd": "free"}, "output": {"usd": "free"}},
+                },
+            },
+            {"id": "deepseek-chat", "object": "model", "owned_by": "deepseek"},
+        ]
+    }
+
+    with _patch_models_endpoint(payload):
+        or_feed = AsyncMock(return_value=[])
+        with patch("routstr.payment.models.async_fetch_openrouter_models", or_feed):
+            models = await GenericUpstreamProvider(base_url="http://x").fetch_models()
+
+    # One malformed entry must not empty the catalog.
+    assert {m.id for m in models} == {"broken-price", "deepseek-chat"}
+    broken = _model_by_id(models, "broken-price")
+    assert broken.enabled is False
+    healthy = _model_by_id(models, "deepseek-chat")
+    assert healthy.enabled is True
+    assert healthy.pricing.prompt == pytest.approx(2.8e-07)
+
+
+# ---------------------------------------------------------------------------
+# litellm rescue — the money-critical case (DeepSeek bare /models)
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_bare_deepseek_resolves_via_litellm() -> None:
+    """DeepSeek's ``/models`` carries no pricing. The old code fabricated
+    ``$0.001`` + ctx 4096; the resolver must instead pull DeepSeek's real
+    rates from litellm's bundled cost map (``$0.28``/``$0.42`` per 1M, ctx
+    131072) and keep the model enabled."""
+    payload = {
+        "data": [
+            {"id": "deepseek-chat", "object": "model", "owned_by": "deepseek"},
+        ]
+    }
+
+    with _patch_models_endpoint(payload):
+        or_feed = AsyncMock(return_value=[])
+        with patch(
+            "routstr.payment.models.async_fetch_openrouter_models", or_feed
+        ):
+            models = await GenericUpstreamProvider(base_url="http://x").fetch_models()
+
+    model = _model_by_id(models, "deepseek-chat")
+    assert model.enabled is True
+    assert model.pricing.prompt == pytest.approx(2.8e-07)
+    assert model.pricing.completion == pytest.approx(4.2e-07)
+    assert model.context_length == 131072
+    # Richer metadata than the two base prices is captured too.
+    assert model.pricing.input_cache_read == pytest.approx(2.8e-08)
+    assert model.top_provider is not None
+    assert model.top_provider.max_completion_tokens == 8192
+    # litellm answered, so the OpenRouter feed is never consulted.
+    or_feed.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_litellm_zero_price_entry_fails_closed(
+    caplog: pytest.LogCaptureFixture,
+) -> None:
+    """A litellm entry that lists a model but prices it at 0/0 (free-tier
+    moderation/rerank models do this) is not a real price — treating it as one
+    would silently serve the model for free. The resolver must reject a both-zero
+    litellm hit and fall through, so the model imports disabled, not at $0."""
+    payload = {
+        "data": [
+            {"id": "omni-moderation-latest", "object": "model", "owned_by": "openai"},
+        ]
+    }
+
+    gen_logger = logging.getLogger("routstr.upstream.generic")
+    gen_logger.addHandler(caplog.handler)
+    try:
+        with _patch_models_endpoint(payload):
+            or_feed = AsyncMock(return_value=[])
+            with patch(
+                "routstr.payment.models.async_fetch_openrouter_models", or_feed
+            ):
+                models = await GenericUpstreamProvider(
+                    base_url="http://x"
+                ).fetch_models()
+    finally:
+        gen_logger.removeHandler(caplog.handler)
+
+    model = _model_by_id(models, "omni-moderation-latest")
+    assert model.enabled is False
+    assert model.pricing.prompt == 0.0
+    assert model.pricing.completion == 0.0
+    assert any(
+        "omni-moderation-latest" in rec.getMessage()
+        for rec in caplog.records
+        if rec.levelno >= logging.WARNING
+    )
+
+
+@pytest.mark.asyncio
+async def test_litellm_output_cap_not_used_as_context() -> None:
+    """litellm's ``max_tokens`` is the completion cap, not the context window
+    (it tracks ``max_output_tokens`` for ~94% of models). When a model reports
+    no ``max_input_tokens``, the resolver must not smuggle the output cap in as
+    the context window; it falls back to the id-based estimate instead, while
+    ``max_tokens`` still feeds the completion limit."""
+    payload = {
+        "data": [
+            {
+                "id": "gemini/gemini-gemma-2-9b-it",
+                "object": "model",
+                "owned_by": "google",
+            },
+        ]
+    }
+
+    with _patch_models_endpoint(payload):
+        or_feed = AsyncMock(return_value=[])
+        with patch(
+            "routstr.payment.models.async_fetch_openrouter_models", or_feed
+        ):
+            models = await GenericUpstreamProvider(base_url="http://x").fetch_models()
+
+    model = _model_by_id(models, "gemini/gemini-gemma-2-9b-it")
+    assert model.enabled is True
+    # litellm gives this model max_input_tokens=None, max_tokens=8192 (an output
+    # cap). Context must come from the estimate (4096), never the 8192 cap.
+    assert model.context_length == 4096
+    # The 8192 output cap still lands where it belongs: the completion limit.
+    assert model.top_provider is not None
+    assert model.top_provider.max_completion_tokens == 8192
+
+
+# ---------------------------------------------------------------------------
+# OpenRouter fallback — litellm misses, OR carries a full payload
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_unknown_to_litellm_resolves_via_openrouter() -> None:
+    """A model litellm has never heard of still resolves if the OpenRouter
+    feed lists it, pulling price + context from that entry."""
+    payload = {
+        "data": [
+            {"id": "exotic/model-9000", "object": "model", "owned_by": "exotic"},
+        ]
+    }
+    or_entry = {
+        "id": "exotic/model-9000",
+        "name": "Exotic 9000",
+        "context_length": 65536,
+        "architecture": {
+            "modality": "text+image->text",
+            "input_modalities": ["text", "image"],
+            "output_modalities": ["text"],
+            "tokenizer": "Other",
+            "instruct_type": None,
+        },
+        "pricing": {"prompt": "0.000005", "completion": "0.000010"},
+        "top_provider": {
+            "context_length": 65536,
+            "max_completion_tokens": 4096,
+            "is_moderated": False,
+        },
+    }
+
+    with _patch_models_endpoint(payload):
+        or_feed = AsyncMock(return_value=[or_entry])
+        with patch(
+            "routstr.payment.models.async_fetch_openrouter_models", or_feed
+        ):
+            models = await GenericUpstreamProvider(base_url="http://x").fetch_models()
+
+    model = _model_by_id(models, "exotic/model-9000")
+    assert model.enabled is True
+    assert model.pricing.prompt == pytest.approx(5e-06)
+    assert model.pricing.completion == pytest.approx(1e-05)
+    assert model.context_length == 65536
+    # The feed's own modality string is carried through verbatim, not recomputed.
+    assert model.architecture.modality == "text+image->text"
+    or_feed.assert_awaited()
+
+
+@pytest.mark.asyncio
+async def test_openrouter_bare_tail_collision_picks_highest_price() -> None:
+    """When a bare model id matches several OpenRouter entries by tail
+    (``model`` ↔ ``a/model``, ``b/model``), the match must be deterministic and
+    money-safe: pick the highest-priced candidate regardless of feed order, so
+    ordering can never leave the node charging below true cost. (The live feed
+    has zero such collisions today; this guards the latent case.)"""
+    payload = {
+        "data": [
+            {"id": "zzz-phantom-model", "object": "model", "owned_by": "mystery"},
+        ]
+    }
+    # Same bare tail, different resellers; the pricier one is listed *second*
+    # so a first-wins match would pick the cheaper (undercharging) entry.
+    or_feed = AsyncMock(
+        return_value=[
+            {
+                "id": "cheapco/zzz-phantom-model",
+                "context_length": 8192,
+                "pricing": {"prompt": "0.000001", "completion": "0.000002"},
+            },
+            {
+                "id": "premiumco/zzz-phantom-model",
+                "context_length": 8192,
+                "pricing": {"prompt": "0.000009", "completion": "0.000010"},
+            },
+        ]
+    )
+
+    with _patch_models_endpoint(payload):
+        with patch("routstr.payment.models.async_fetch_openrouter_models", or_feed):
+            models = await GenericUpstreamProvider(base_url="http://x").fetch_models()
+
+    model = _model_by_id(models, "zzz-phantom-model")
+    assert model.pricing.prompt == pytest.approx(9e-06)
+    assert model.pricing.completion == pytest.approx(1e-05)
+
+
+@pytest.mark.asyncio
+async def test_openrouter_bare_tail_tie_breaks_on_combined_cost() -> None:
+    """The bare-tail tie-break must weigh *both* rates, not prompt alone.
+    Given two colliding entries where one is cheaper on prompt but far dearer
+    on completion, ranking by prompt would pick the entry that undercharges
+    output-heavy traffic. Pick the highest *combined* per-token cost so the
+    money-safe choice holds whichever way the traffic leans."""
+    payload = {
+        "data": [
+            {"id": "yyy-phantom-model", "object": "model", "owned_by": "mystery"},
+        ]
+    }
+    # dear-overall is listed first with the *lower* prompt, so a prompt-only max
+    # would wrongly pick the second (cheaper-overall) entry.
+    or_feed = AsyncMock(
+        return_value=[
+            {
+                "id": "dearco/yyy-phantom-model",
+                "context_length": 8192,
+                "pricing": {"prompt": "0.000001", "completion": "0.000100"},
+            },
+            {
+                "id": "cheapco/yyy-phantom-model",
+                "context_length": 8192,
+                "pricing": {"prompt": "0.000009", "completion": "0.000002"},
+            },
+        ]
+    )
+
+    with _patch_models_endpoint(payload):
+        with patch("routstr.payment.models.async_fetch_openrouter_models", or_feed):
+            models = await GenericUpstreamProvider(base_url="http://x").fetch_models()
+
+    model = _model_by_id(models, "yyy-phantom-model")
+    assert model.pricing.prompt == pytest.approx(1e-06)
+    assert model.pricing.completion == pytest.approx(1e-04)
+
+
+@pytest.mark.asyncio
+async def test_openrouter_feed_fetched_once_per_discovery() -> None:
+    """Two models both missing litellm must share a single OpenRouter fetch —
+    the feed is not re-downloaded per model."""
+    payload = {
+        "data": [
+            {"id": "exotic/model-a", "object": "model", "owned_by": "exotic"},
+            {"id": "exotic/model-b", "object": "model", "owned_by": "exotic"},
+        ]
+    }
+    or_feed = AsyncMock(
+        return_value=[
+            {
+                "id": "exotic/model-a",
+                "context_length": 8192,
+                "pricing": {"prompt": "0.000001", "completion": "0.000002"},
+            },
+            {
+                "id": "exotic/model-b",
+                "context_length": 8192,
+                "pricing": {"prompt": "0.000003", "completion": "0.000004"},
+            },
+        ]
+    )
+
+    with _patch_models_endpoint(payload):
+        with patch("routstr.payment.models.async_fetch_openrouter_models", or_feed):
+            models = await GenericUpstreamProvider(base_url="http://x").fetch_models()
+
+    assert {m.id for m in models} == {"exotic/model-a", "exotic/model-b"}
+    assert or_feed.await_count == 1
+
+
+# ---------------------------------------------------------------------------
+# fail closed — no source resolves → disabled + warned, never fabricated
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_unresolvable_model_fails_closed(
+    caplog: pytest.LogCaptureFixture,
+) -> None:
+    """When native, litellm and OpenRouter all miss, the model is imported
+    disabled with a warning naming it — and no price is invented (the old
+    ``$0.001`` placeholder is gone)."""
+    payload = {
+        "data": [
+            {
+                "id": "nobody-has-priced-this-xyz",
+                "object": "model",
+                "owned_by": "mystery",
+            },
+        ]
+    }
+
+    # routstr loggers set propagate=False, so caplog's root handler misses
+    # them; attach its handler to the provider logger directly.
+    gen_logger = logging.getLogger("routstr.upstream.generic")
+    gen_logger.addHandler(caplog.handler)
+    try:
+        with _patch_models_endpoint(payload):
+            or_feed = AsyncMock(return_value=[])
+            with patch(
+                "routstr.payment.models.async_fetch_openrouter_models", or_feed
+            ):
+                models = await GenericUpstreamProvider(
+                    base_url="http://x"
+                ).fetch_models()
+    finally:
+        gen_logger.removeHandler(caplog.handler)
+
+    model = _model_by_id(models, "nobody-has-priced-this-xyz")
+    assert model.enabled is False
+    assert model.pricing.prompt == 0.0
+    assert model.pricing.completion == 0.0
+    assert any(
+        "nobody-has-priced-this-xyz" in rec.getMessage()
+        for rec in caplog.records
+        if rec.levelno >= logging.WARNING
+    )

tests/unit/test_upstream_rate_limit.py

@@ -0,0 +1,396 @@
+"""Tests for upstream rate-limit detection, classification, and org-ID redaction.
+
+Covers issue #555: upstream OpenAI-compatible providers return rate-limit
+errors that embed a sensitive organization ID. The proxy must classify these
+distinctly (``UPSTREAM_RATE_LIMIT``), preserve useful debugging fields, and
+never emit a raw ``org-*`` identifier in logs, errors, or returned bodies.
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+from unittest.mock import AsyncMock, MagicMock, Mock, patch
+
+import httpx
+import pytest
+
+from routstr.core.redaction import redact_org_ids
+from routstr.upstream.base import BaseUpstreamProvider
+from routstr.upstream.rate_limit import (
+    UPSTREAM_RATE_LIMIT,
+    RateLimitInfo,
+    classify_rate_limit,
+)
+
+# The exact scenario from the issue, with a realistic (fake) org identifier.
+RAW_ORG_ID = "org-abc123XYZ456def"
+RATE_LIMIT_MESSAGE = (
+    f"Rate limit reached for gpt-5.5-2026-04-23 (for limit gpt-5.5) in "
+    f"organization {RAW_ORG_ID} on tokens per min (TPM): Limit 180000000, "
+    f"Used 180000000, Requested 8929. Please try again in 2ms. Visit "
+    f"https://platform.openai.com/account/rate-limits to learn more."
+)
+
+
+def _make_request(request_id: str = "req-123") -> Mock:
+    request = Mock(spec=["method", "state"])
+    request.method = "POST"
+    request.state = Mock()
+    request.state.request_id = request_id
+    return request
+
+
+def _make_upstream_response(
+    *,
+    body: bytes,
+    status_code: int = 429,
+    content_type: str | None = "application/json",
+    extra_headers: dict[str, str] | None = None,
+) -> httpx.Response:
+    headers: dict[str, str] = {}
+    if content_type is not None:
+        headers["content-type"] = content_type
+    if extra_headers:
+        headers.update(extra_headers)
+    return httpx.Response(status_code=status_code, headers=headers, content=body)
+
+
+@pytest.fixture
+def provider() -> BaseUpstreamProvider:
+    return BaseUpstreamProvider(
+        base_url="https://privateprovider.xyz", api_key="k", provider_fee=1.0
+    )
+
+
+# --------------------------------------------------------------------------- #
+# Redaction
+# --------------------------------------------------------------------------- #
+
+
+def test_redact_org_ids_replaces_identifier() -> None:
+    assert RAW_ORG_ID not in redact_org_ids(RATE_LIMIT_MESSAGE)
+    assert "org-[REDACTED]" in redact_org_ids(RATE_LIMIT_MESSAGE)
+
+
+def test_redact_org_ids_is_idempotent() -> None:
+    once = redact_org_ids(RATE_LIMIT_MESSAGE)
+    assert redact_org_ids(once) == once
+
+
+def test_redact_org_ids_leaves_unrelated_text() -> None:
+    assert redact_org_ids("organize the org-chart") == "organize the org-chart"
+    assert redact_org_ids("") == ""
+
+
+# --------------------------------------------------------------------------- #
+# Classification
+# --------------------------------------------------------------------------- #
+
+
+def test_classify_exact_scenario() -> None:
+    info = classify_rate_limit(429, RATE_LIMIT_MESSAGE)
+    assert isinstance(info, RateLimitInfo)
+    assert info.code == UPSTREAM_RATE_LIMIT
+    assert info.model == "gpt-5.5-2026-04-23"
+    assert info.limit_name == "gpt-5.5"
+    assert info.metric == "tokens per min (TPM)"
+    assert info.limit == 180000000
+    assert info.used == 180000000
+    assert info.requested == 8929
+    assert info.retry_after_seconds == pytest.approx(0.002)
+    # Redaction-safe: no raw org id survives into the structured view.
+    assert RAW_ORG_ID not in info.message
+    assert RAW_ORG_ID not in json.dumps(info.as_details())
+
+
+def test_classify_by_status_code_without_marker() -> None:
+    info = classify_rate_limit(429, "slow down")
+    assert info is not None
+    assert info.code == UPSTREAM_RATE_LIMIT
+
+
+def test_classify_by_message_marker_without_429() -> None:
+    info = classify_rate_limit(400, "rate_limit_exceeded for this key")
+    assert info is not None
+
+
+def test_retry_after_header_takes_precedence() -> None:
+    info = classify_rate_limit(429, RATE_LIMIT_MESSAGE, {"Retry-After": "12"})
+    assert info is not None
+    assert info.retry_after_seconds == pytest.approx(12.0)
+
+
+def test_non_rate_limit_error_is_not_classified() -> None:
+    assert classify_rate_limit(400, "invalid request: missing field 'model'") is None
+    assert classify_rate_limit(500, "internal server error") is None
+
+
+# --------------------------------------------------------------------------- #
+# forward_upstream_error_response integration
+# --------------------------------------------------------------------------- #
+
+
+@pytest.mark.asyncio
+async def test_json_rate_limit_body_is_redacted_and_forwarded(
+    provider: BaseUpstreamProvider,
+) -> None:
+    body = json.dumps(
+        {"error": {"message": RATE_LIMIT_MESSAGE, "type": "rate_limit_exceeded"}}
+    ).encode()
+    upstream = _make_upstream_response(body=body, status_code=429)
+
+    response = await provider.forward_upstream_error_response(
+        _make_request(), "v1/chat/completions", upstream
+    )
+
+    assert response.status_code == 429
+    raw = bytes(response.body).decode()
+    # No raw organization id may survive in the forwarded body.
+    assert RAW_ORG_ID not in raw
+    assert "org-[REDACTED]" in raw
+    # Body remains valid JSON; the original type is preserved while a stable
+    # rate-limit code is injected so callers can switch on it.
+    payload: dict[str, Any] = json.loads(raw)
+    assert payload["error"]["type"] == "rate_limit_exceeded"
+    assert payload["error"]["code"] == UPSTREAM_RATE_LIMIT
+    assert payload["error"]["details"]["model"] == "gpt-5.5-2026-04-23"
+    # A retry hint extracted from the message is surfaced as a header.
+    assert "retry-after" in {k.lower() for k in response.headers}
+
+
+@pytest.mark.asyncio
+async def test_non_json_rate_limit_envelope_uses_stable_code(
+    provider: BaseUpstreamProvider,
+) -> None:
+    upstream = _make_upstream_response(
+        body=RATE_LIMIT_MESSAGE.encode(),
+        status_code=429,
+        content_type="text/plain",
+    )
+
+    response = await provider.forward_upstream_error_response(
+        _make_request(), "v1/chat/completions", upstream
+    )
+
+    payload: dict[str, Any] = json.loads(bytes(response.body))
+    assert payload["error"]["code"] == UPSTREAM_RATE_LIMIT
+    assert payload["error"]["details"]["model"] == "gpt-5.5-2026-04-23"
+    serialized = json.dumps(payload)
+    assert RAW_ORG_ID not in serialized
+    assert "org-[REDACTED]" in serialized
+
+
+@pytest.mark.asyncio
+async def test_non_rate_limit_json_error_unchanged(
+    provider: BaseUpstreamProvider,
+) -> None:
+    body = json.dumps(
+        {"error": {"message": "missing field 'model'", "type": "invalid_request"}}
+    ).encode()
+    upstream = _make_upstream_response(body=body, status_code=400)
+
+    response = await provider.forward_upstream_error_response(
+        _make_request(), "v1/chat/completions", upstream
+    )
+
+    assert response.status_code == 400
+    payload: dict[str, Any] = json.loads(bytes(response.body))
+    assert payload["error"]["type"] == "invalid_request"
+    assert "retry-after" not in {k.lower() for k in response.headers}
+
+
+# --------------------------------------------------------------------------- #
+# UpstreamError -> proxy response (preserves code/details/status)
+# --------------------------------------------------------------------------- #
+
+
+def test_create_upstream_error_response_preserves_structure() -> None:
+    from routstr.core.exceptions import UpstreamError
+    from routstr.payment.helpers import create_upstream_error_response
+
+    info = classify_rate_limit(429, RATE_LIMIT_MESSAGE)
+    assert info is not None
+    err = UpstreamError(
+        f"Upstream error via litellm: {RATE_LIMIT_MESSAGE}",
+        status_code=429,
+        code=info.code,
+        details=info.as_details(),
+    )
+
+    response = create_upstream_error_response(err, _make_request())
+
+    # Original upstream status is preserved (not flattened to 502).
+    assert response.status_code == 429
+    payload: dict[str, Any] = json.loads(bytes(response.body))
+    assert payload["error"]["type"] == "upstream_error"
+    assert payload["error"]["code"] == UPSTREAM_RATE_LIMIT
+    assert payload["error"]["details"]["requested"] == 8929
+    serialized = json.dumps(payload)
+    assert RAW_ORG_ID not in serialized
+    assert "org-[REDACTED]" in serialized
+
+
+def test_generic_upstream_error_still_defaults_to_502() -> None:
+    from routstr.core.exceptions import UpstreamError
+    from routstr.payment.helpers import create_upstream_error_response
+
+    err = UpstreamError("connection refused")  # status_code defaults to 502
+
+    response = create_upstream_error_response(err, _make_request())
+
+    assert response.status_code == 502
+    payload: dict[str, Any] = json.loads(bytes(response.body))
+    assert payload["error"]["type"] == "upstream_error"
+    assert payload["error"]["code"] == 502
+    assert "details" not in payload["error"]
+
+
+# --------------------------------------------------------------------------- #
+# Structured log-extra redaction
+# --------------------------------------------------------------------------- #
+
+
+def test_security_filter_redacts_org_id_in_extra() -> None:
+    import logging
+
+    from routstr.core.logging import SecurityFilter
+
+    record = logging.LogRecord(
+        name="test",
+        level=logging.ERROR,
+        pathname=__file__,
+        lineno=1,
+        msg="upstream failed",
+        args=(),
+        exc_info=None,
+    )
+    # Simulate an ``extra={"body_preview": ...}`` field carrying an org id.
+    setattr(record, "body_preview", RATE_LIMIT_MESSAGE)
+
+    assert SecurityFilter().filter(record) is True
+    redacted: str = getattr(record, "body_preview")
+    assert RAW_ORG_ID not in redacted
+    assert "org-[REDACTED]" in redacted
+
+
+def test_security_filter_redacts_org_id_in_nested_extra() -> None:
+    import logging
+
+    from routstr.core.logging import SecurityFilter
+
+    record = logging.LogRecord(
+        name="test",
+        level=logging.ERROR,
+        pathname=__file__,
+        lineno=1,
+        msg="upstream failed",
+        args=(),
+        exc_info=None,
+    )
+    # Nested structures: dict containing a list containing the org id.
+    setattr(record, "body", {"error": {"messages": [RATE_LIMIT_MESSAGE]}})
+
+    assert SecurityFilter().filter(record) is True
+    serialized = json.dumps(getattr(record, "body"))
+    assert RAW_ORG_ID not in serialized
+    assert "org-[REDACTED]" in serialized
+
+
+# --------------------------------------------------------------------------- #
+# 5xx-wrapped rate limit through forward_upstream_error_response
+# --------------------------------------------------------------------------- #
+
+
+@pytest.mark.asyncio
+async def test_5xx_wrapped_rate_limit_is_classified(
+    provider: BaseUpstreamProvider,
+) -> None:
+    # Some providers wrap a rate-limit in a 5xx envelope; classification must
+    # key off the message marker, not only the 429 status.
+    body = json.dumps({"error": {"message": RATE_LIMIT_MESSAGE}}).encode()
+    upstream = _make_upstream_response(body=body, status_code=500)
+
+    response = await provider.forward_upstream_error_response(
+        _make_request(), "v1/chat/completions", upstream
+    )
+
+    assert response.status_code == 500
+    payload: dict[str, Any] = json.loads(bytes(response.body))
+    assert payload["error"]["code"] == UPSTREAM_RATE_LIMIT
+    serialized = json.dumps(payload)
+    assert RAW_ORG_ID not in serialized
+    assert "org-[REDACTED]" in serialized
+
+
+# --------------------------------------------------------------------------- #
+# Real proxy loop: structured error surfaced + reservation reverted once
+# --------------------------------------------------------------------------- #
+
+
+@pytest.mark.asyncio
+async def test_proxy_loop_surfaces_rate_limit_and_reverts_once() -> None:
+    from routstr import proxy as proxy_module
+    from routstr.core.db import ApiKey
+    from routstr.core.exceptions import UpstreamError
+
+    info = classify_rate_limit(429, RATE_LIMIT_MESSAGE)
+    assert info is not None
+
+    key = ApiKey(hashed_key="rlkey", balance=10_000)
+
+    request = MagicMock()
+    request.method = "POST"
+    request.headers = {"authorization": "Bearer sk-rlkey"}
+    request.body = AsyncMock(return_value=b'{"model": "test-model"}')
+    request.state = MagicMock()
+    request.state.request_id = "req-rl"
+
+    upstream = MagicMock()
+    upstream.provider_type = "test"
+    upstream.prepare_headers = MagicMock(side_effect=lambda h: h)
+    upstream.forward_request = AsyncMock(
+        side_effect=UpstreamError(
+            f"Upstream error via litellm: {RATE_LIMIT_MESSAGE}",
+            status_code=429,
+            code=info.code,
+            details=info.as_details(),
+        )
+    )
+
+    session = MagicMock()
+    revert_mock = AsyncMock(return_value=True)
+
+    with (
+        patch.object(proxy_module, "get_model_instance", return_value=MagicMock()),
+        patch.object(proxy_module, "get_provider_for_model", return_value=[upstream]),
+        patch.object(
+            proxy_module, "get_max_cost_for_model", AsyncMock(return_value=1_000)
+        ),
+        patch.object(
+            proxy_module,
+            "calculate_discounted_max_cost",
+            AsyncMock(return_value=1_000),
+        ),
+        patch.object(proxy_module, "check_token_balance", MagicMock()),
+        patch.object(
+            proxy_module, "get_bearer_token_key", AsyncMock(return_value=key)
+        ),
+        patch.object(proxy_module, "pay_for_request", AsyncMock(return_value=1_000)),
+        patch.object(proxy_module, "revert_pay_for_request", revert_mock),
+    ):
+        response = await proxy_module.proxy(
+            request, "v1/chat/completions", session=session
+        )
+
+    # Original 429 status and the stable code/details survive to the client.
+    assert response.status_code == 429
+    payload: dict[str, Any] = json.loads(bytes(response.body))
+    assert payload["error"]["type"] == "upstream_error"
+    assert payload["error"]["code"] == UPSTREAM_RATE_LIMIT
+    assert payload["error"]["details"]["requested"] == 8929
+    serialized = json.dumps(payload)
+    assert RAW_ORG_ID not in serialized
+    assert "org-[REDACTED]" in serialized
+    # Single upstream failed -> reservation reverted exactly once (no double-charge).
+    revert_mock.assert_awaited_once_with(key, session, 1_000)

tests/unit/test_usage_normalization.py

@@ -0,0 +1,140 @@
+"""Tests for vendor-agnostic usage normalization.
+
+Specifies the seam that keeps vendor usage dialects out of generic billing
+code: a canonical ``NormalizedUsage`` shape produced by
+``routstr.payment.usage.normalize_usage`` (union parser for the known,
+non-colliding dialects). ``calculate_cost`` normalizes the response's usage
+object with this parser and needs no vendor knowledge of its own.
+"""
+
+import os
+
+os.environ.setdefault("UPSTREAM_BASE_URL", "http://test")
+os.environ.setdefault("UPSTREAM_API_KEY", "test")
+os.environ.setdefault("LIGHTNING_ADDRESS", "test@stm.to")
+
+import pytest
+
+from routstr.payment.usage import NormalizedUsage, normalize_usage
+
+# ============================================================================
+# The union parser: one canonical shape for all known dialects
+# ============================================================================
+
+
+@pytest.mark.parametrize(
+    "usage,expected",
+    [
+        # OpenAI: cached_tokens included in prompt_tokens → subtracted
+        (
+            {
+                "prompt_tokens": 2000,
+                "completion_tokens": 100,
+                "prompt_tokens_details": {"cached_tokens": 800},
+            },
+            NormalizedUsage(
+                input_tokens=1200,
+                output_tokens=100,
+                cache_read_tokens=800,
+                cache_write_tokens=0,
+            ),
+        ),
+        # DeepSeek: hit/miss fields, prompt_tokens = hit + miss → hit subtracted
+        (
+            {
+                "prompt_tokens": 10000,
+                "completion_tokens": 500,
+                "prompt_cache_hit_tokens": 9000,
+                "prompt_cache_miss_tokens": 1000,
+            },
+            NormalizedUsage(
+                input_tokens=1000,
+                output_tokens=500,
+                cache_read_tokens=9000,
+                cache_write_tokens=0,
+            ),
+        ),
+        # Anthropic: cache fields additive, input_tokens NOT reduced
+        (
+            {
+                "input_tokens": 300,
+                "output_tokens": 100,
+                "cache_read_input_tokens": 500,
+                "cache_creation_input_tokens": 2000,
+            },
+            NormalizedUsage(
+                input_tokens=300,
+                output_tokens=100,
+                cache_read_tokens=500,
+                cache_write_tokens=2000,
+            ),
+        ),
+        # Plain OpenAI without caching
+        (
+            {"prompt_tokens": 100, "completion_tokens": 50},
+            NormalizedUsage(input_tokens=100, output_tokens=50),
+        ),
+        # OpenRouter: cache writes nested as prompt_tokens_details.cache_write_tokens,
+        # both reads and writes included in prompt_tokens → both subtracted
+        (
+            {
+                "prompt_tokens": 10000,
+                "completion_tokens": 60,
+                "prompt_tokens_details": {
+                    "cached_tokens": 5000,
+                    "cache_write_tokens": 2000,
+                },
+            },
+            NormalizedUsage(
+                input_tokens=3000,
+                output_tokens=60,
+                cache_read_tokens=5000,
+                cache_write_tokens=2000,
+            ),
+        ),
+        # litellm-normalized Anthropic: prompt_tokens is the grand total and the
+        # write field is named cache_creation_tokens; top-level fields mirror it.
+        # prompt_tokens present → both subtracted (NOT additive like native).
+        (
+            {
+                "prompt_tokens": 10000,
+                "completion_tokens": 100,
+                "cache_read_input_tokens": 5000,
+                "cache_creation_input_tokens": 2000,
+                "prompt_tokens_details": {
+                    "cached_tokens": 5000,
+                    "cache_creation_tokens": 2000,
+                },
+            },
+            NormalizedUsage(
+                input_tokens=3000,
+                output_tokens=100,
+                cache_read_tokens=5000,
+                cache_write_tokens=2000,
+            ),
+        ),
+    ],
+)
+def test_normalize_usage_dialects(usage: dict, expected: NormalizedUsage) -> None:
+    """Each known vendor dialect maps onto the same canonical shape."""
+    assert normalize_usage(usage) == expected
+
+
+def test_normalize_usage_absent_usage() -> None:
+    """Missing/invalid usage yields None so callers can bill at max cost."""
+    assert normalize_usage(None) is None
+    assert normalize_usage("not a dict") is None  # type: ignore[arg-type]
+
+
+def test_normalize_usage_never_negative() -> None:
+    """Buggy upstreams reporting more cached than prompt tokens clamp to 0."""
+    result = normalize_usage(
+        {
+            "prompt_tokens": 100,
+            "completion_tokens": 50,
+            "prompt_cache_hit_tokens": 150,
+        }
+    )
+    assert result is not None
+    assert result.input_tokens == 0
+    assert result.cache_read_tokens == 150

tests/unit/test_wallet_money_paths.py

@@ -0,0 +1,160 @@
+"""Additional money-path coverage tests for wallet.py (86% → target 90%).
+
+Tests error classification, periodic task structure, and token operations.
+"""
+
+from unittest.mock import AsyncMock, Mock, patch
+
+import pytest
+
+# ===========================================================================
+# is_mint_connection_error
+# ===========================================================================
+
+def test_is_mint_connection_error_true() -> None:
+    """Connection errors are detected."""
+    from routstr.wallet import is_mint_connection_error
+
+    assert is_mint_connection_error(ConnectionRefusedError("refused")) is True
+    assert is_mint_connection_error(TimeoutError("timeout")) is True
+
+
+def test_is_mint_connection_error_false() -> None:
+    """Non-connection errors are not flagged."""
+    from routstr.wallet import is_mint_connection_error
+
+    assert is_mint_connection_error(ValueError("bad data")) is False
+    assert is_mint_connection_error(KeyError("missing key")) is False
+    assert is_mint_connection_error(RuntimeError("something broke")) is False
+    assert is_mint_connection_error(AttributeError("no attr")) is False
+    # OSError is NOT a connection error unless it's a subclass
+    assert is_mint_connection_error(OSError("generic")) is False
+
+
+# ===========================================================================
+# classify_redemption_error
+# ===========================================================================
+
+def test_classify_redemption_error_token_consumed() -> None:
+    """Token already spent returns token_consumed classification."""
+    from routstr.wallet import TokenConsumedError, classify_redemption_error
+
+    result = classify_redemption_error(
+        TokenConsumedError("Token was already redeemed")
+    )
+    assert result is not None
+    assert result[0] == "token_consumed"
+    assert result[1] == 500
+
+
+def test_classify_redemption_error_mint_connection() -> None:
+    """Mint connection error is classified correctly."""
+    from routstr.wallet import classify_redemption_error
+
+    result = classify_redemption_error(
+        ConnectionRefusedError("Connection refused")
+    )
+    assert result is not None
+    # Should classify as mint_connection or return error tuple
+    assert isinstance(result, tuple)
+    assert len(result) >= 3
+
+
+def test_classify_redemption_error_unclassified() -> None:
+    """Generic errors are classified as cashu_error with 400 status."""
+    from routstr.wallet import classify_redemption_error
+
+    result = classify_redemption_error(ValueError("unexpected"))
+    # classify_redemption_error classifies all unrecognized errors
+    # as cashu_error with a generic message
+    assert result is not None
+    assert result[0] == "cashu_error"
+    assert result[1] == 400
+
+
+# ===========================================================================
+# Store readiness: store_cashu_transaction succeeds
+# ===========================================================================
+
+@pytest.mark.asyncio
+async def test_store_cashu_transaction_succeeds_normally() -> None:
+    """Normal store_cashu_transaction returns True on success."""
+    from routstr.core.db import store_cashu_transaction
+
+    with patch("routstr.core.db.create_session") as mock_create:
+        mock_session = AsyncMock()
+        mock_session.commit = AsyncMock()
+        mock_session.__aenter__ = AsyncMock(return_value=mock_session)
+        mock_session.__aexit__ = AsyncMock(return_value=None)
+        mock_create.return_value = mock_session
+
+        result = await store_cashu_transaction(
+            token="cashuAtest",
+            amount=1000,
+            unit="sat",
+            typ="in",
+            request_id="req-test",
+        )
+
+        assert result is True
+
+
+# ===========================================================================
+# get_balance
+# ===========================================================================
+
+@pytest.mark.asyncio
+async def test_get_balance_returns_integer() -> None:
+    """get_balance returns an integer balance from wallet."""
+    from routstr.wallet import get_balance
+
+    mock_wallet = Mock()
+    mock_wallet.available_balance = Mock(amount=50000)
+    mock_wallet.load_mint = AsyncMock()
+    mock_wallet.load_proofs = AsyncMock()
+
+    with (
+        patch("routstr.wallet._wallets", {}),
+        patch("routstr.wallet.Wallet.with_db", return_value=mock_wallet),
+    ):
+        balance = await get_balance("sat")
+        assert isinstance(balance, int)
+        assert balance == 50000
+
+
+# ===========================================================================
+# Periodic task structure verification
+# ===========================================================================
+
+def test_periodic_payout_has_loop_and_error_handling() -> None:
+    """periodic_payout runs in a loop with error handling."""
+    import inspect
+
+    from routstr import wallet
+
+    source = inspect.getsource(wallet.periodic_payout)
+    assert "while True" in source
+    assert "except" in source, "Must have error handling"
+
+
+def test_periodic_refund_sweep_has_error_handling() -> None:
+    """Refund sweep catches errors to stay alive."""
+    import inspect
+
+    from routstr import wallet
+
+    source = inspect.getsource(wallet.periodic_refund_sweep)
+    assert "while True" in source
+    assert "except" in source, "Must have error handling"
+
+
+def test_periodic_routstr_fee_payout_structure() -> None:
+    """Fee payout loop handles missing LN address gracefully."""
+    import inspect
+
+    from routstr import wallet
+
+    source = inspect.getsource(wallet.periodic_routstr_fee_payout)
+    # Returns early if ROUTSTR_LN_ADDRESS not set
+    assert "ROUTSTR_LN_ADDRESS" in source
+    assert "return" in source or "skip" in source.lower()

ui/Dockerfile

@@ -4,15 +4,15 @@ FROM base AS deps
 RUN apk add --no-cache libc6-compat
 WORKDIR /app
 
-RUN corepack enable pnpm && corepack prepare pnpm@latest --activate
+RUN corepack enable pnpm && corepack prepare pnpm@10.15.0 --activate
 
-COPY package.json pnpm-lock.yaml* ./
+COPY package.json pnpm-lock.yaml* pnpm-workspace.yaml* ./
 RUN pnpm install --frozen-lockfile
 
 FROM base AS builder
 WORKDIR /app
 
-RUN corepack enable pnpm && corepack prepare pnpm@latest --activate
+RUN corepack enable pnpm && corepack prepare pnpm@10.15.0 --activate
 
 COPY --from=deps /app/node_modules ./node_modules
 COPY . .

ui/Dockerfile.build

@@ -6,14 +6,14 @@ RUN apk add --no-cache libc6-compat
 WORKDIR /app
 
 # Copy package files
-COPY package.json pnpm-lock.yaml* ./
-RUN corepack enable pnpm && corepack prepare pnpm@latest --activate
+COPY package.json pnpm-lock.yaml* pnpm-workspace.yaml* ./
+RUN corepack enable pnpm && corepack prepare pnpm@10.15.0 --activate
 RUN pnpm install --frozen-lockfile
 
 # Build the UI
 FROM base AS builder
 WORKDIR /app
-RUN corepack enable pnpm && corepack prepare pnpm@latest --activate
+RUN corepack enable pnpm && corepack prepare pnpm@10.15.0 --activate
 COPY --from=deps /app/node_modules ./node_modules
 COPY . .
 

ui/README.md

@@ -1,36 +1,44 @@
-This is a [Next.js](https://nextjs.org) project bootstrapped with [`create-next-app`](https://nextjs.org/docs/app/api-reference/cli/create-next-app).
+# Routstr node admin UI
 
-## Getting Started
+A [Next.js](https://nextjs.org) app (App Router, **static export**) that provides the
+admin dashboard for a `routstr-core` node: login, settings, providers, balances,
+transactions, usage, and logs.
 
-First, run the development server:
+There is no separate web server in production. `next build` produces a fully static
+export (`next.config.ts` sets `output: 'export'`), and the FastAPI backend serves it
+directly from `../ui_out/` (see `routstr/core/main.py`). So the UI and the API are
+served from the **same origin** in production.
 
-```bash
-npm run dev
-# or
-yarn dev
-# or
-pnpm dev
-# or
-bun dev
-```
+## Developing the UI (hot reload)
 
-Open [http://localhost:3000](http://localhost:3000) with your browser to see the result.
+The everyday loop runs two processes side by side — you do **not** rebuild the static
+export while developing:
 
-You can start editing the page by modifying `app/page.tsx`. The page auto-updates as you edit the file.
+1. Start the backend on `:8000` — from the repo root: `make docker-up` (or
+   `uvicorn routstr.core.main:app --reload`).
+2. Start the Next.js dev server on `:3000` — from the repo root: `make ui-dev`
+   (or `cd ui && pnpm dev`). Edits hot-reload instantly.
 
-This project uses [`next/font`](https://nextjs.org/docs/app/building-your-application/optimizing/fonts) to automatically optimize and load [Geist](https://vercel.com/font), a new font family for Vercel.
+Open http://localhost:3000. With no `NEXT_PUBLIC_API_URL` set, the UI falls back to
+`http://127.0.0.1:8000` in development (see `lib/api/services/configuration.ts`), so it
+talks to the local backend out of the box.
 
-## Learn More
+Because dev is cross-origin (`:3000` → `:8000`), it relies on the backend's CORS
+allowing the UI origin. The default `cors_origins` is `["*"]`; if you tighten CORS,
+keep `http://localhost:3000` allowed for development.
 
-To learn more about Next.js, take a look at the following resources:
+## Building the integrated/static UI (what production serves)
 
-- [Next.js Documentation](https://nextjs.org/docs) - learn about Next.js features and API.
-- [Learn Next.js](https://nextjs.org/learn) - an interactive Next.js tutorial.
+To produce the bundle that FastAPI serves from `../ui_out/`:
 
-You can check out [the Next.js GitHub repository](https://github.com/vercel/next.js) - your feedback and contributions are welcome!
+- `make ui-build` — builds with local Node/pnpm (`scripts/build-ui.sh`), then moves
+  `ui/out/*` to `../ui_out/`.
+- `make ui-build-docker` — same, but inside Docker (no local Node needed).
 
-## Deploy on Vercel
+`NEXT_PUBLIC_*` variables are read from the repo-root `.env` at build time and baked in.
+For a same-origin deployment leave `NEXT_PUBLIC_API_URL` empty (relative paths); the UI
+uses `window.location.origin` at runtime. After building, start the backend and open
+http://localhost:8000 — the dashboard is served at `/` and `/admin`.
 
-The easiest way to deploy your Next.js app is to use the [Vercel Platform](https://vercel.com/new?utm_medium=default-template&filter=next.js&utm_source=create-next-app&utm_campaign=create-next-app-readme) from the creators of Next.js.
-
-Check out our [Next.js deployment documentation](https://nextjs.org/docs/app/building-your-application/deploying) for more details.
+If `../ui_out/` does not exist, the backend logs a warning at startup and serves the API
+only (hitting a UI route returns a small JSON fallback instead of the dashboard).

ui/app/providers/page.tsx

@@ -297,16 +297,13 @@ export default function ProvidersPage() {
   };
 
   const toggleProviderExpansion = (providerId: number) => {
-    const newExpanded = new Set(expandedProviders);
-    if (newExpanded.has(providerId)) {
-      newExpanded.delete(providerId);
-    } else {
-      newExpanded.add(providerId);
-    }
-    setExpandedProviders(newExpanded);
-    if (!newExpanded.has(providerId)) {
+    if (expandedProviders.has(providerId)) {
+      setExpandedProviders(new Set());
       setViewingModels(null);
     } else {
+      // Accordion: only one provider open at a time so switching to another
+      // provider's models auto-collapses the previously expanded one.
+      setExpandedProviders(new Set([providerId]));
       setViewingModels(providerId);
     }
   };