ruff fix

Introduce child key

examples/create_child_keys.py

@@ -0,0 +1,45 @@
+import json
+import sys
+
+import httpx
+
+
+def create_child_keys(base_url: str, api_key: str, count: int = 3) -> list[str]:
+    headers = {"Authorization": f"Bearer {api_key}"}
+
+    print(f"Requesting {count} child keys from {base_url}...")
+
+    child_keys = []
+
+    for i in range(count):
+        try:
+            response = httpx.post(f"{base_url}/v1/balance/child-key", headers=headers)
+            if response.status_code == 200:
+                data = response.json()
+                child_keys.append(data["api_key"])
+                print(
+                    f"  [{i + 1}] Created: {data['api_key']} (Cost: {data['cost_msats']} msats)"
+                )
+            else:
+                print(f"  [{i + 1}] Failed: {response.status_code} - {response.text}")
+        except Exception as e:
+            print(f"  [{i + 1}] Error: {str(e)}")
+
+    return child_keys
+
+
+if __name__ == "__main__":
+    if len(sys.argv) < 2:
+        print("Usage: python create_child_keys.py <api_key_or_cashu_token> [base_url]")
+        sys.exit(1)
+
+    auth_key = sys.argv[1]
+    base_url = sys.argv[2] if len(sys.argv) > 2 else "http://localhost:8000"
+
+    keys = create_child_keys(base_url, auth_key)
+
+    if keys:
+        print("\nSuccessfully created child keys:")
+        print(json.dumps(keys, indent=2))
+    else:
+        print("\nNo child keys were created.")

migrations/versions/a86e5348850b_.py

@@ -0,0 +1,42 @@
+"""
+
+Revision ID: a86e5348850b
+Revises: b9667ffc5701
+Create Date: 2026-01-10 18:57:48.475781
+"""
+
+import sqlalchemy as sa
+import sqlmodel
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "a86e5348850b"
+down_revision = "b9667ffc5701"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    # Use batch_alter_table for SQLite compatibility
+    with op.batch_alter_table("api_keys", schema=None) as batch_op:
+        batch_op.add_column(
+            sa.Column(
+                "parent_key_hash", sqlmodel.sql.sqltypes.AutoString(), nullable=True
+            )
+        )
+        batch_op.create_index(
+            batch_op.f("ix_api_keys_parent_key_hash"), ["parent_key_hash"], unique=False
+        )
+        batch_op.create_foreign_key(
+            "fk_api_keys_parent_key_hash",
+            "api_keys",
+            ["parent_key_hash"],
+            ["hashed_key"],
+        )
+
+
+def downgrade() -> None:
+    with op.batch_alter_table("api_keys", schema=None) as batch_op:
+        batch_op.drop_constraint("fk_api_keys_parent_key_hash", type_="foreignkey")
+        batch_op.drop_index(batch_op.f("ix_api_keys_parent_key_hash"))
+        batch_op.drop_column("parent_key_hash")

routstr/algorithm.py

@@ -84,93 +84,26 @@ def get_provider_penalty(provider: "BaseUpstreamProvider") -> float:
     return penalty
 
 
-def should_prefer_model(
-    candidate_model: "Model",
-    candidate_provider: "BaseUpstreamProvider",
-    current_model: "Model",
-    current_provider: "BaseUpstreamProvider",
-    alias: str,
-) -> bool:
-    """Determine if candidate model should replace current model for an alias.
-
-    This is the core decision function for model prioritization. It considers:
-    1. Alias matching quality (exact match vs. canonical slug match)
-    2. Model cost (lower is better)
-    3. Provider penalties (e.g., slight preference against OpenRouter)
-
-    Args:
-        candidate_model: The new model being considered
-        candidate_provider: Provider offering the candidate model
-        current_model: The currently selected model for this alias
-        current_provider: Provider offering the current model
-        alias: The model alias being mapped
-
-    Returns:
-        True if candidate should replace current, False otherwise
-    """
-
-    def get_base_model_id(model_id: str) -> str:
-        """Get base model ID by removing provider prefix."""
-        return model_id.split("/", 1)[1] if "/" in model_id else model_id
-
-    def alias_priority(model: "Model") -> int:
-        """Rank how strong the mapping of alias->model is.
-
-        Highest priority when alias exactly equals the model ID without provider prefix.
-        Next when alias equals canonical slug without prefix. Otherwise lowest.
-        """
-        model_base = get_base_model_id(model.id)
-        if model_base == alias:
-            return 3
-        if model.canonical_slug:
-            canonical_base = get_base_model_id(model.canonical_slug)
-            if canonical_base == alias:
-                return 2
-        return 1
-
-    candidate_alias_priority = alias_priority(candidate_model)
-    current_alias_priority = alias_priority(current_model)
-
-    # If candidate has better alias match, prefer it regardless of cost
-    if candidate_alias_priority > current_alias_priority:
-        return True
-
-    # If current has better alias match, keep it regardless of cost
-    if current_alias_priority > candidate_alias_priority:
-        return False
-
-    # Same alias priority - compare costs
-    candidate_cost = calculate_model_cost_score(candidate_model)
-    current_cost = calculate_model_cost_score(current_model)
-
-    # Apply provider penalties
-    candidate_adjusted = candidate_cost * get_provider_penalty(candidate_provider)
-    current_adjusted = current_cost * get_provider_penalty(current_provider)
-
-    # Prefer lower adjusted cost
-    should_replace = candidate_adjusted < current_adjusted
-
-    return should_replace
-
-
 def create_model_mappings(
     upstreams: list["BaseUpstreamProvider"],
     overrides_by_id: dict[str, tuple],
     disabled_model_ids: set[str],
-) -> tuple[dict[str, "Model"], dict[str, "BaseUpstreamProvider"], dict[str, "Model"]]:
+) -> tuple[
+    dict[str, "Model"], dict[str, list["BaseUpstreamProvider"]], dict[str, "Model"]
+]:
     """Create optimal model mappings based on cost and provider preferences.
 
     This is the main entry point for the algorithm. It processes all upstream providers
     and creates three mappings based on cost optimization:
 
     1. model_instances: alias -> Model (all model aliases mapped to their Model objects)
-    2. provider_map: alias -> UpstreamProvider (which provider to use for each alias)
+    2. provider_map: alias -> List[UpstreamProvider] (sorted list of providers for each alias)
     3. unique_models: base_id -> Model (unique models without provider prefixes)
 
     The algorithm:
     - Processes non-OpenRouter providers first (they're typically cheaper)
     - Then processes OpenRouter models (they can still win if cheaper)
-    - For each model alias, uses should_prefer_model() to select the best provider
+    - For each model alias, collects all candidates and sorts them by priority and cost.
 
     Args:
         upstreams: List of all upstream provider instances
@@ -183,8 +116,7 @@ def create_model_mappings(
     from .payment.models import _row_to_model
     from .upstream.helpers import resolve_model_alias
 
-    model_instances: dict[str, "Model"] = {}
-    provider_map: dict[str, "BaseUpstreamProvider"] = {}
+    candidates: dict[str, list[tuple["Model", "BaseUpstreamProvider"]]] = {}
     unique_models: dict[str, "Model"] = {}
 
     # Separate OpenRouter from other providers
@@ -202,24 +134,14 @@ def create_model_mappings(
         """Get base model ID by removing provider prefix."""
         return model_id.split("/", 1)[1] if "/" in model_id else model_id
 
-    def _maybe_set_alias(
+    def _add_candidate(
         alias: str, model: "Model", provider: "BaseUpstreamProvider"
     ) -> None:
-        """Set alias to model/provider if not set or if new model is preferred."""
+        """Add candidate model/provider for an alias."""
         alias_lower = alias.lower()
-        existing_model = model_instances.get(alias_lower)
-        if not existing_model:
-            # No existing mapping, set it
-            model_instances[alias_lower] = model
-            provider_map[alias_lower] = provider
-        else:
-            # Check if candidate should replace existing
-            existing_provider = provider_map[alias_lower]
-            if should_prefer_model(
-                model, provider, existing_model, existing_provider, alias
-            ):
-                model_instances[alias_lower] = model
-                provider_map[alias_lower] = provider
+        if alias_lower not in candidates:
+            candidates[alias_lower] = []
+        candidates[alias_lower].append((model, provider))
 
     def process_provider_models(
         upstream: "BaseUpstreamProvider", is_openrouter: bool = False
@@ -266,21 +188,55 @@ def create_model_mappings(
 
             # Try to set each alias
             for alias in aliases:
-                _maybe_set_alias(alias, model_to_use, upstream)
+                _add_candidate(alias, model_to_use, upstream)
 
-    # Process non-OpenRouter providers first (they're typically cheaper)
+    # Process non-OpenRouter providers first
     for upstream in other_upstreams:
         process_provider_models(upstream, is_openrouter=False)
 
-    # Process OpenRouter last - models only win if they're cheaper or better matched
+    # Process OpenRouter last
     if openrouter:
         process_provider_models(openrouter, is_openrouter=True)
 
-    # Log provider distribution
+    # Sort candidates and build final maps
+    model_instances: dict[str, "Model"] = {}
+    provider_map: dict[str, list["BaseUpstreamProvider"]] = {}
+
+    def alias_priority(model: "Model", alias: str) -> int:
+        """Rank how strong the mapping of alias->model is."""
+        model_base = get_base_model_id(model.id)
+        if model_base == alias:
+            return 3
+        if model.canonical_slug:
+            canonical_base = get_base_model_id(model.canonical_slug)
+            if canonical_base == alias:
+                return 2
+        return 1
+
+    for alias, items in candidates.items():
+        # Sort key: (priority DESC, cost ASC)
+        # Using negative cost for DESC sort overall to keep high priority first
+        def sort_key(item: tuple["Model", "BaseUpstreamProvider"]) -> tuple[int, float]:
+            model, provider = item
+            priority = alias_priority(model, alias)
+            cost = calculate_model_cost_score(model)
+            penalty = get_provider_penalty(provider)
+            adjusted_cost = cost * penalty
+            return (priority, -adjusted_cost)
+
+        items.sort(key=sort_key, reverse=True)
+
+        best_model, best_provider = items[0]
+        model_instances[alias] = best_model
+        provider_map[alias] = [p for _, p in items]
+
+    # Log provider distribution (using top provider for stats)
     provider_counts: dict[str, int] = {}
-    for provider in provider_map.values():
-        provider_name = getattr(provider, "upstream_name", "unknown")
-        provider_counts[provider_name] = provider_counts.get(provider_name, 0) + 1
+    for providers in provider_map.values():
+        if providers:
+            provider = providers[0]
+            provider_name = getattr(provider, "upstream_name", "unknown")
+            provider_counts[provider_name] = provider_counts.get(provider_name, 0) + 1
 
     logger.debug(
         f"Updated model mappings with ({len(unique_models)} unique models and {len(model_instances)} aliases)",

routstr/auth.py

@@ -286,30 +286,55 @@ async def validate_bearer_key(
     )
 
 
+async def get_billing_key(key: ApiKey, session: AsyncSession) -> ApiKey:
+    """Returns the key that should be charged for the request."""
+    if key.parent_key_hash:
+        parent = await session.get(ApiKey, key.parent_key_hash)
+        if parent:
+            # We want to keep the total_requests and total_spent on the child key
+            # but use the balance and reserved_balance of the parent.
+            # However, pay_for_request updates reserved_balance and total_requests.
+            # To stay simple, we charge the parent's balance and update parent's total_requests.
+            return parent
+        else:
+            logger.error(
+                "Parent key not found for child key",
+                extra={
+                    "child_key_hash": key.hashed_key[:8] + "...",
+                    "parent_key_hash": key.parent_key_hash[:8] + "...",
+                },
+            )
+    return key
+
+
 async def pay_for_request(
     key: ApiKey, cost_per_request: int, session: AsyncSession
 ) -> int:
     """Process payment for a request."""
 
+    billing_key = await get_billing_key(key, session)
+
     logger.info(
         "Processing payment for request",
         extra={
             "key_hash": key.hashed_key[:8] + "...",
-            "current_balance": key.balance,
+            "billing_key_hash": billing_key.hashed_key[:8] + "...",
+            "current_balance": billing_key.balance,
             "required_cost": cost_per_request,
-            "sufficient_balance": key.balance >= cost_per_request,
+            "sufficient_balance": billing_key.balance >= cost_per_request,
         },
     )
 
-    if key.total_balance < cost_per_request:
+    if billing_key.total_balance < cost_per_request:
         logger.warning(
             "Insufficient balance for request",
             extra={
                 "key_hash": key.hashed_key[:8] + "...",
-                "balance": key.balance,
-                "reserved_balance": key.reserved_balance,
+                "billing_key_hash": billing_key.hashed_key[:8] + "...",
+                "balance": billing_key.balance,
+                "reserved_balance": billing_key.reserved_balance,
                 "required": cost_per_request,
-                "shortfall": cost_per_request - key.total_balance,
+                "shortfall": cost_per_request - billing_key.total_balance,
             },
         )
 
@@ -317,7 +342,7 @@ async def pay_for_request(
             status_code=402,
             detail={
                 "error": {
-                    "message": f"Insufficient balance: {cost_per_request} mSats required. {key.total_balance} available. (reserved: {key.reserved_balance})",
+                    "message": f"Insufficient balance: {cost_per_request} mSats required. {billing_key.total_balance} available. (reserved: {billing_key.reserved_balance})",
                     "type": "insufficient_quota",
                     "code": "insufficient_balance",
                 }
@@ -328,15 +353,16 @@ async def pay_for_request(
         "Charging base cost for request",
         extra={
             "key_hash": key.hashed_key[:8] + "...",
+            "billing_key_hash": billing_key.hashed_key[:8] + "...",
             "cost": cost_per_request,
-            "balance_before": key.balance,
+            "balance_before": billing_key.balance,
         },
     )
 
     # Charge the base cost for the request atomically to avoid race conditions
     stmt = (
         update(ApiKey)
-        .where(col(ApiKey.hashed_key) == key.hashed_key)
+        .where(col(ApiKey.hashed_key) == billing_key.hashed_key)
         .where(col(ApiKey.balance) - col(ApiKey.reserved_balance) >= cost_per_request)
         .values(
             reserved_balance=col(ApiKey.reserved_balance) + cost_per_request,
@@ -344,6 +370,16 @@ async def pay_for_request(
         )
     )
     result = await session.exec(stmt)  # type: ignore[call-overload]
+
+    # Also increment total_requests on the child key if it's different
+    if billing_key.hashed_key != key.hashed_key:
+        child_stmt = (
+            update(ApiKey)
+            .where(col(ApiKey.hashed_key) == key.hashed_key)
+            .values(total_requests=col(ApiKey.total_requests) + 1)
+        )
+        await session.exec(child_stmt)  # type: ignore[call-overload]
+
     await session.commit()
 
     if result.rowcount == 0:
@@ -351,8 +387,9 @@ async def pay_for_request(
             "Concurrent request depleted balance",
             extra={
                 "key_hash": key.hashed_key[:8] + "...",
+                "billing_key_hash": billing_key.hashed_key[:8] + "...",
                 "required_cost": cost_per_request,
-                "current_balance": key.balance,
+                "current_balance": billing_key.balance,
             },
         )
 
@@ -361,23 +398,26 @@ async def pay_for_request(
             status_code=402,
             detail={
                 "error": {
-                    "message": f"Insufficient balance: {cost_per_request} mSats required. {key.balance} available.",
+                    "message": f"Insufficient balance: {cost_per_request} mSats required. {billing_key.balance} available.",
                     "type": "insufficient_quota",
                     "code": "insufficient_balance",
                 }
             },
         )
 
-    await session.refresh(key)
+    await session.refresh(billing_key)
+    if billing_key.hashed_key != key.hashed_key:
+        await session.refresh(key)
 
     logger.info(
         "Payment processed successfully",
         extra={
             "key_hash": key.hashed_key[:8] + "...",
+            "billing_key_hash": billing_key.hashed_key[:8] + "...",
             "charged_amount": cost_per_request,
-            "new_balance": key.balance,
-            "total_spent": key.total_spent,
-            "total_requests": key.total_requests,
+            "new_balance": billing_key.balance,
+            "total_spent": billing_key.total_spent,
+            "total_requests": billing_key.total_requests,
         },
     )
 
@@ -387,9 +427,11 @@ async def pay_for_request(
 async def revert_pay_for_request(
     key: ApiKey, session: AsyncSession, cost_per_request: int
 ) -> None:
+    billing_key = await get_billing_key(key, session)
+
     stmt = (
         update(ApiKey)
-        .where(col(ApiKey.hashed_key) == key.hashed_key)
+        .where(col(ApiKey.hashed_key) == billing_key.hashed_key)
         .values(
             reserved_balance=col(ApiKey.reserved_balance) - cost_per_request,
             total_requests=col(ApiKey.total_requests) - 1,
@@ -397,27 +439,40 @@ async def revert_pay_for_request(
     )
 
     result = await session.exec(stmt)  # type: ignore[call-overload]
+
+    # Also decrement total_requests on the child key if it's different
+    if billing_key.hashed_key != key.hashed_key:
+        child_stmt = (
+            update(ApiKey)
+            .where(col(ApiKey.hashed_key) == key.hashed_key)
+            .values(total_requests=col(ApiKey.total_requests) - 1)
+        )
+        await session.exec(child_stmt)  # type: ignore[call-overload]
+
     await session.commit()
     if result.rowcount == 0:
         logger.error(
             "Failed to revert payment - insufficient reserved balance",
             extra={
                 "key_hash": key.hashed_key[:8] + "...",
+                "billing_key_hash": billing_key.hashed_key[:8] + "...",
                 "cost_to_revert": cost_per_request,
-                "current_reserved_balance": key.reserved_balance,
+                "current_reserved_balance": billing_key.reserved_balance,
             },
         )
         raise HTTPException(
             status_code=402,
             detail={
                 "error": {
-                    "message": f"failed to revert request payment: {cost_per_request} mSats required. {key.balance} available.",
+                    "message": f"failed to revert request payment: {cost_per_request} mSats required. {billing_key.balance} available.",
                     "type": "payment_error",
                     "code": "payment_error",
                 }
             },
         )
-    await session.refresh(key)
+    await session.refresh(billing_key)
+    if billing_key.hashed_key != key.hashed_key:
+        await session.refresh(key)
 
 
 async def adjust_payment_for_tokens(
@@ -428,15 +483,17 @@ async def adjust_payment_for_tokens(
     This is called after the initial payment and the upstream request is complete.
     Returns cost data to be included in the response.
     """
+    billing_key = await get_billing_key(key, session)
     model = response_data.get("model", "unknown")
 
     logger.debug(
         "Starting payment adjustment for tokens",
         extra={
             "key_hash": key.hashed_key[:8] + "...",
+            "billing_key_hash": billing_key.hashed_key[:8] + "...",
             "model": model,
             "deducted_max_cost": deducted_max_cost,
-            "current_balance": key.balance,
+            "current_balance": billing_key.balance,
             "has_usage": "usage" in response_data,
         },
     )
@@ -446,8 +503,10 @@ async def adjust_payment_for_tokens(
         try:
             release_stmt = (
                 update(ApiKey)
-                .where(col(ApiKey.hashed_key) == key.hashed_key)
-                .values(reserved_balance=col(ApiKey.reserved_balance) - deducted_max_cost)
+                .where(col(ApiKey.hashed_key) == billing_key.hashed_key)
+                .values(
+                    reserved_balance=col(ApiKey.reserved_balance) - deducted_max_cost
+                )
             )
             await session.exec(release_stmt)  # type: ignore[call-overload]
             await session.commit()
@@ -455,13 +514,18 @@ async def adjust_payment_for_tokens(
                 "Released reservation without charging (fallback)",
                 extra={
                     "key_hash": key.hashed_key[:8] + "...",
+                    "billing_key_hash": billing_key.hashed_key[:8] + "...",
                     "deducted_max_cost": deducted_max_cost,
                 },
             )
         except Exception as e:
             logger.error(
                 "Failed to release reservation in fallback",
-                extra={"error": str(e), "key_hash": key.hashed_key[:8] + "..."},
+                extra={
+                    "error": str(e),
+                    "key_hash": key.hashed_key[:8] + "...",
+                    "billing_key_hash": billing_key.hashed_key[:8] + "...",
+                },
             )
 
     match await calculate_cost(response_data, deducted_max_cost, session):
@@ -470,6 +534,7 @@ async def adjust_payment_for_tokens(
                 "Using max cost data (no token adjustment)",
                 extra={
                     "key_hash": key.hashed_key[:8] + "...",
+                    "billing_key_hash": billing_key.hashed_key[:8] + "...",
                     "model": model,
                     "max_cost": cost.total_msats,
                 },
@@ -477,7 +542,7 @@ async def adjust_payment_for_tokens(
             # Finalize by releasing reservation and charging max cost
             finalize_stmt = (
                 update(ApiKey)
-                .where(col(ApiKey.hashed_key) == key.hashed_key)
+                .where(col(ApiKey.hashed_key) == billing_key.hashed_key)
                 .values(
                     reserved_balance=col(ApiKey.reserved_balance) - deducted_max_cost,
                     balance=col(ApiKey.balance) - cost.total_msats,
@@ -485,27 +550,41 @@ async def adjust_payment_for_tokens(
                 )
             )
             result = await session.exec(finalize_stmt)  # type: ignore[call-overload]
+
+            # Also update total_spent on the child key if it's different
+            if billing_key.hashed_key != key.hashed_key:
+                child_stmt = (
+                    update(ApiKey)
+                    .where(col(ApiKey.hashed_key) == key.hashed_key)
+                    .values(total_spent=col(ApiKey.total_spent) + cost.total_msats)
+                )
+                await session.exec(child_stmt)  # type: ignore[call-overload]
+
             await session.commit()
             if result.rowcount == 0:
                 logger.error(
                     "Failed to finalize max-cost payment - retrying reservation release",
                     extra={
                         "key_hash": key.hashed_key[:8] + "...",
+                        "billing_key_hash": billing_key.hashed_key[:8] + "...",
                         "deducted_max_cost": deducted_max_cost,
-                        "current_reserved_balance": key.reserved_balance,
+                        "current_reserved_balance": billing_key.reserved_balance,
                         "total_cost": cost.total_msats,
                         "model": model,
                     },
                 )
                 await release_reservation_only()
             else:
-                await session.refresh(key)
+                await session.refresh(billing_key)
+                if billing_key.hashed_key != key.hashed_key:
+                    await session.refresh(key)
                 logger.info(
                     "Max cost payment finalized",
                     extra={
                         "key_hash": key.hashed_key[:8] + "...",
+                        "billing_key_hash": billing_key.hashed_key[:8] + "...",
                         "charged_amount": cost.total_msats,
-                        "new_balance": key.balance,
+                        "new_balance": billing_key.balance,
                         "model": model,
                     },
                 )
@@ -521,6 +600,7 @@ async def adjust_payment_for_tokens(
                 "Calculated token-based cost",
                 extra={
                     "key_hash": key.hashed_key[:8] + "...",
+                    "billing_key_hash": billing_key.hashed_key[:8] + "...",
                     "model": model,
                     "token_cost": cost.total_msats,
                     "deducted_max_cost": deducted_max_cost,
@@ -533,11 +613,15 @@ async def adjust_payment_for_tokens(
             if cost_difference == 0:
                 logger.debug(
                     "Finalizing with exact reserved cost",
-                    extra={"key_hash": key.hashed_key[:8] + "...", "model": model},
+                    extra={
+                        "key_hash": key.hashed_key[:8] + "...",
+                        "billing_key_hash": billing_key.hashed_key[:8] + "...",
+                        "model": model,
+                    },
                 )
                 finalize_stmt = (
                     update(ApiKey)
-                    .where(col(ApiKey.hashed_key) == key.hashed_key)
+                    .where(col(ApiKey.hashed_key) == billing_key.hashed_key)
                     .values(
                         reserved_balance=col(ApiKey.reserved_balance)
                         - deducted_max_cost,
@@ -546,8 +630,20 @@ async def adjust_payment_for_tokens(
                     )
                 )
                 await session.exec(finalize_stmt)  # type: ignore[call-overload]
+
+                # Also update total_spent on the child key if it's different
+                if billing_key.hashed_key != key.hashed_key:
+                    child_stmt = (
+                        update(ApiKey)
+                        .where(col(ApiKey.hashed_key) == key.hashed_key)
+                        .values(total_spent=col(ApiKey.total_spent) + total_cost_msats)
+                    )
+                    await session.exec(child_stmt)  # type: ignore[call-overload]
+
                 await session.commit()
-                await session.refresh(key)
+                await session.refresh(billing_key)
+                if billing_key.hashed_key != key.hashed_key:
+                    await session.refresh(key)
                 return cost.dict()
 
             # this should never happen why do we handle this???
@@ -557,16 +653,17 @@ async def adjust_payment_for_tokens(
                     "Additional charge required for token usage",
                     extra={
                         "key_hash": key.hashed_key[:8] + "...",
+                        "billing_key_hash": billing_key.hashed_key[:8] + "...",
                         "additional_charge": cost_difference,
-                        "current_balance": key.balance,
-                        "sufficient_balance": key.balance >= cost_difference,
+                        "current_balance": billing_key.balance,
+                        "sufficient_balance": billing_key.balance >= cost_difference,
                         "model": model,
                     },
                 )
 
                 finalize_stmt = (
                     update(ApiKey)
-                    .where(col(ApiKey.hashed_key) == key.hashed_key)
+                    .where(col(ApiKey.hashed_key) == billing_key.hashed_key)
                     .values(
                         reserved_balance=col(ApiKey.reserved_balance)
                         - deducted_max_cost,
@@ -575,18 +672,31 @@ async def adjust_payment_for_tokens(
                     )
                 )
                 result = await session.exec(finalize_stmt)  # type: ignore[call-overload]
+
+                # Also update total_spent on the child key if it's different
+                if billing_key.hashed_key != key.hashed_key:
+                    child_stmt = (
+                        update(ApiKey)
+                        .where(col(ApiKey.hashed_key) == key.hashed_key)
+                        .values(total_spent=col(ApiKey.total_spent) + total_cost_msats)
+                    )
+                    await session.exec(child_stmt)  # type: ignore[call-overload]
+
                 await session.commit()
 
                 if result.rowcount:
                     cost.total_msats = total_cost_msats
-                    await session.refresh(key)
+                    await session.refresh(billing_key)
+                    if billing_key.hashed_key != key.hashed_key:
+                        await session.refresh(key)
 
                     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": key.balance,
+                            "new_balance": billing_key.balance,
                             "model": model,
                         },
                     )
@@ -595,6 +705,7 @@ async def adjust_payment_for_tokens(
                         "Failed to finalize additional charge - releasing reservation",
                         extra={
                             "key_hash": key.hashed_key[:8] + "...",
+                            "billing_key_hash": billing_key.hashed_key[:8] + "...",
                             "attempted_charge": total_cost_msats,
                             "model": model,
                         },
@@ -607,15 +718,16 @@ async def adjust_payment_for_tokens(
                     "Refunding excess payment",
                     extra={
                         "key_hash": key.hashed_key[:8] + "...",
+                        "billing_key_hash": billing_key.hashed_key[:8] + "...",
                         "refund_amount": refund,
-                        "current_balance": key.balance,
+                        "current_balance": billing_key.balance,
                         "model": model,
                     },
                 )
 
                 refund_stmt = (
                     update(ApiKey)
-                    .where(col(ApiKey.hashed_key) == key.hashed_key)
+                    .where(col(ApiKey.hashed_key) == billing_key.hashed_key)
                     .values(
                         reserved_balance=col(ApiKey.reserved_balance)
                         - deducted_max_cost,
@@ -624,6 +736,16 @@ async def adjust_payment_for_tokens(
                     )
                 )
                 result = await session.exec(refund_stmt)  # type: ignore[call-overload]
+
+                # Also update total_spent on the child key if it's different
+                if billing_key.hashed_key != key.hashed_key:
+                    child_stmt = (
+                        update(ApiKey)
+                        .where(col(ApiKey.hashed_key) == key.hashed_key)
+                        .values(total_spent=col(ApiKey.total_spent) + total_cost_msats)
+                    )
+                    await session.exec(child_stmt)  # type: ignore[call-overload]
+
                 await session.commit()
 
                 if result.rowcount == 0:
@@ -631,8 +753,9 @@ async def adjust_payment_for_tokens(
                         "Failed to finalize payment - releasing reservation",
                         extra={
                             "key_hash": key.hashed_key[:8] + "...",
+                            "billing_key_hash": billing_key.hashed_key[:8] + "...",
                             "deducted_max_cost": deducted_max_cost,
-                            "current_reserved_balance": key.reserved_balance,
+                            "current_reserved_balance": billing_key.reserved_balance,
                             "total_cost": total_cost_msats,
                             "model": model,
                         },
@@ -640,14 +763,17 @@ async def adjust_payment_for_tokens(
                     await release_reservation_only()
                 else:
                     cost.total_msats = total_cost_msats
-                    await session.refresh(key)
+                    await session.refresh(billing_key)
+                    if billing_key.hashed_key != key.hashed_key:
+                        await session.refresh(key)
 
                     logger.info(
                         "Refund processed successfully",
                         extra={
                             "key_hash": key.hashed_key[:8] + "...",
+                            "billing_key_hash": billing_key.hashed_key[:8] + "...",
                             "refunded_amount": refund,
-                            "new_balance": key.balance,
+                            "new_balance": billing_key.balance,
                             "final_cost": cost.total_msats,
                             "model": model,
                         },

routstr/balance.py

@@ -32,16 +32,30 @@ async def get_key_from_header(
     )
 
 
-# TODO: remove this endpoint when frontend is updated
-@router.get("/", include_in_schema=False)
-async def account_info(key: ApiKey = Depends(get_key_from_header)) -> dict:
+async def get_balance_info(key: ApiKey, session: AsyncSession) -> dict:
+    from .auth import get_billing_key
+
+    billing_key = await get_billing_key(key, session)
     return {
         "api_key": "sk-" + key.hashed_key,
-        "balance": key.balance,
-        "reserved": key.reserved_balance,
+        "balance": billing_key.balance,
+        "reserved": billing_key.reserved_balance,
+        "is_child": key.parent_key_hash is not None,
+        "parent_key": "sk-" + key.parent_key_hash if key.parent_key_hash else None,
+        "total_requests": key.total_requests,
+        "total_spent": key.total_spent,
     }
 
 
+# TODO: remove this endpoint when frontend is updated
+@router.get("/", include_in_schema=False)
+async def account_info(
+    key: ApiKey = Depends(get_key_from_header),
+    session: AsyncSession = Depends(get_session),
+) -> dict:
+    return await get_balance_info(key, session)
+
+
 # TODO: Implement POST /v1/wallet/create endpoint
 # This endpoint should accept:
 # - cashu_token (required): The eCash token to deposit
@@ -66,12 +80,11 @@ async def create_balance(
 
 
 @router.get("/info")
-async def wallet_info(key: ApiKey = Depends(get_key_from_header)) -> dict:
-    return {
-        "api_key": "sk-" + key.hashed_key,
-        "balance": key.balance,
-        "reserved": key.reserved_balance,
-    }
+async def wallet_info(
+    key: ApiKey = Depends(get_key_from_header),
+    session: AsyncSession = Depends(get_session),
+) -> dict:
+    return await get_balance_info(key, session)
 
 
 class TopupRequest(BaseModel):
@@ -85,6 +98,10 @@ async def topup_wallet_endpoint(
     key: ApiKey = Depends(get_key_from_header),
     session: AsyncSession = Depends(get_session),
 ) -> dict[str, int]:
+    from .auth import get_billing_key
+
+    billing_key = await get_billing_key(key, session)
+
     if topup_request is not None:
         cashu_token = topup_request.cashu_token
     if cashu_token is None:
@@ -94,7 +111,7 @@ async def topup_wallet_endpoint(
     if len(cashu_token) < 10 or "cashu" not in cashu_token:
         raise HTTPException(status_code=400, detail="Invalid token format")
     try:
-        amount_msats = await credit_balance(cashu_token, key, session)
+        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():
@@ -155,6 +172,12 @@ async def refund_wallet_endpoint(
 
     key: ApiKey = await validate_bearer_key(bearer_value, session)
 
+    if key.parent_key_hash:
+        raise HTTPException(
+            status_code=400,
+            detail="Cannot refund child key. Please refund the parent key instead.",
+        )
+
     remaining_balance_msats: int = key.total_balance
 
     if key.refund_currency == "sat":
@@ -228,6 +251,53 @@ async def donate(token: str, ref: str | None = None) -> str:
         return "Invalid token."
 
 
+@router.post("/child-key")
+async def create_child_key(
+    key: ApiKey = Depends(get_key_from_header),
+    session: AsyncSession = Depends(get_session),
+) -> dict:
+    """Creates a child API key that uses the parent's balance."""
+    # Check if this is already a child key
+    if key.parent_key_hash:
+        raise HTTPException(
+            status_code=400,
+            detail="Cannot create a child key for another child key.",
+        )
+
+    cost = settings.child_key_cost
+
+    if key.total_balance < cost:
+        raise HTTPException(
+            status_code=402,
+            detail=f"Insufficient balance to create child key. {cost} mSats required.",
+        )
+
+    # Deduct cost from parent
+    key.balance -= cost
+    key.total_spent += cost
+    session.add(key)
+
+    # Generate new key
+    import secrets
+
+    new_key_raw = secrets.token_hex(32)
+    new_key_hash = new_key_raw  # We use the raw key as the hash for sk- keys
+
+    child_key = ApiKey(
+        hashed_key=new_key_hash,
+        balance=0,
+        parent_key_hash=key.hashed_key,
+    )
+    session.add(child_key)
+    await session.commit()
+
+    return {
+        "api_key": "sk-" + new_key_hash,
+        "cost_msats": cost,
+        "parent_balance": key.balance,
+    }
+
+
 @router.api_route(
     "/{path:path}",
     methods=["GET", "POST", "PUT", "DELETE"],

routstr/core/admin.py

@@ -124,7 +124,7 @@ async def partial_apikeys(request: Request) -> str:
 
     rows = "".join(
         [
-            f"<tr><td>{key.hashed_key}</td><td>{key.balance}</td><td>{key.total_spent}</td><td>{key.total_requests}</td><td>{key.refund_address}</td><td>{fmt_time(key.key_expiry_time)}</td></tr>"
+            f"<tr><td>{key.hashed_key}{' <br><small>(Child of ' + key.parent_key_hash[:8] + '...)</small>' if key.parent_key_hash else ''}</td><td>{key.balance}</td><td>{key.total_spent}</td><td>{key.total_requests}</td><td>{key.refund_address}</td><td>{fmt_time(key.key_expiry_time)}</td></tr>"
             for key in api_keys
         ]
     )
@@ -158,6 +158,7 @@ async def get_temporary_balances_api(request: Request) -> list[dict[str, object]
             "total_requests": key.total_requests,
             "refund_address": key.refund_address,
             "key_expiry_time": key.key_expiry_time,
+            "parent_key_hash": key.parent_key_hash,
         }
         for key in api_keys
     ]

routstr/core/db.py

@@ -47,6 +47,9 @@ class ApiKey(SQLModel, table=True):  # type: ignore
         default=None,
         description="Currency of the cashu-token",
     )
+    parent_key_hash: str | None = Field(
+        default=None, foreign_key="api_keys.hashed_key", index=True
+    )
 
     @property
     def total_balance(self) -> int:

routstr/core/exceptions.py

@@ -6,6 +6,15 @@ from .logging import get_logger
 logger = get_logger(__name__)
 
 
+class UpstreamError(Exception):
+    """Exception raised when an upstream provider fails."""
+
+    def __init__(self, message: str, status_code: int = 502):
+        self.message = message
+        self.status_code = status_code
+        super().__init__(message)
+
+
 async def http_exception_handler(request: Request, exc: Exception) -> JSONResponse:
     """Handle HTTP exceptions and include request ID in response."""
     request_id = getattr(request.state, "request_id", "unknown")

routstr/core/main.py

@@ -13,10 +13,7 @@ from starlette.exceptions import HTTPException
 from ..balance import balance_router, deprecated_wallet_router
 from ..discovery import providers_cache_refresher, providers_router
 from ..nip91 import announce_provider
-from ..payment.models import (
-    models_router,
-    update_sats_pricing,
-)
+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 ..wallet import periodic_payout

routstr/core/settings.py

@@ -52,6 +52,7 @@ class Settings(BaseSettings):
     exchange_fee: float = Field(default=1.005, env="EXCHANGE_FEE")
     upstream_provider_fee: float = Field(default=1.05, env="UPSTREAM_PROVIDER_FEE")
     tolerance_percentage: float = Field(default=1.0, env="TOLERANCE_PERCENTAGE")
+    child_key_cost: int = Field(default=1000, env="CHILD_KEY_COST")
     # Minimum per-request charge in millisatoshis when model pricing is free/zero
     min_request_msat: int = Field(default=1, env="MIN_REQUEST_MSAT")
     reset_reserved_balance_on_startup: bool = Field(

routstr/proxy.py

@@ -16,6 +16,7 @@ from .core.db import (
     create_session,
     get_session,
 )
+from .core.exceptions import UpstreamError
 from .payment.helpers import (
     calculate_discounted_max_cost,
     check_token_balance,
@@ -31,7 +32,9 @@ proxy_router = APIRouter()
 
 _upstreams: list[BaseUpstreamProvider] = []
 _model_instances: dict[str, Model] = {}  # All aliases -> Model
-_provider_map: dict[str, BaseUpstreamProvider] = {}  # All aliases -> Provider
+_provider_map: dict[
+    str, list[BaseUpstreamProvider]
+] = {}  # All aliases -> List[Provider]
 _unique_models: dict[str, Model] = {}  # Unique model.id -> Model (no duplicates)
 
 
@@ -68,8 +71,8 @@ def get_model_instance(model_id: str) -> Model | None:
     return _model_instances.get(model_id.lower())
 
 
-def get_provider_for_model(model_id: str) -> BaseUpstreamProvider | None:
-    """Get UpstreamProvider for model ID from global cache."""
+def get_provider_for_model(model_id: str) -> list[BaseUpstreamProvider] | None:
+    """Get UpstreamProvider list for model ID from global cache."""
     return _provider_map.get(model_id.lower())
 
 
@@ -154,8 +157,8 @@ async def proxy(
             "invalid_model", f"Model '{model_id}' not found", 400, request=request
         )
 
-    upstream = get_provider_for_model(model_id)
-    if not upstream:
+    upstreams = get_provider_for_model(model_id)
+    if not upstreams:
         return create_error_response(
             "invalid_model",
             f"No provider found for model '{model_id}'",
@@ -163,6 +166,10 @@ async def proxy(
             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]
+
     _max_cost_for_model = await get_max_cost_for_model(
         model=model_id, session=session, model_obj=model_obj
     )
@@ -172,14 +179,31 @@ async def proxy(
     check_token_balance(headers, request_body_dict, max_cost_for_model)
 
     if x_cashu := headers.get("x-cashu", None):
-        if is_responses_api:
-            return await upstream.handle_x_cashu_responses(
-                request, x_cashu, path, max_cost_for_model, model_obj
-            )
-        else:
-            return await upstream.handle_x_cashu(
-                request, x_cashu, path, max_cost_for_model, model_obj
-            )
+        last_error = None
+        for i, upstream in enumerate(upstreams):
+            try:
+                if is_responses_api:
+                    return await upstream.handle_x_cashu_responses(
+                        request, x_cashu, path, max_cost_for_model, model_obj
+                    )
+                else:
+                    return await upstream.handle_x_cashu(
+                        request, x_cashu, path, max_cost_for_model, model_obj
+                    )
+            except UpstreamError as e:
+                logger.warning(
+                    f"Upstream {upstream.provider_type} failed (x-cashu): {e}"
+                )
+                if i == len(upstreams) - 1:
+                    last_error = e
+                continue
+
+        return create_error_response(
+            "upstream_error",
+            str(last_error) if last_error else "All upstreams failed",
+            502,
+            request=request,
+        )
 
     elif auth := headers.get("authorization", None):
         key = await get_bearer_token_key(headers, path, session, auth)
@@ -194,56 +218,152 @@ async def proxy(
             )
 
         logger.debug("Processing unauthenticated GET request", extra={"path": path})
-        headers = upstream.prepare_headers(dict(request.headers))
-        return await upstream.forward_get_request(request, path, headers)
+
+        last_error_response = None
+        for i, upstream in enumerate(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(upstreams) - 1:
+                    error_message = ""
+                    try:
+                        if hasattr(response, "body"):
+                            body_bytes = response.body
+                            data = json.loads(body_bytes)
+                            if "error" in data:
+                                error_data = data["error"]
+                                if isinstance(error_data, dict):
+                                    error_message = error_data.get("message", "")
+                                elif isinstance(error_data, str):
+                                    error_message = error_data
+                    except Exception:
+                        pass
+
+                    await upstream.on_upstream_error_redirect(
+                        response.status_code, error_message
+                    )
+
+                    logger.warning(
+                        f"Upstream {upstream.provider_type} returned {response.status_code} (GET), trying next provider",
+                        extra={
+                            "status_code": response.status_code,
+                            "upstream": upstream.provider_type,
+                        },
+                    )
+                    continue
+                return response
+            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
+                    )
+                continue
+        return last_error_response or create_error_response(
+            "upstream_error", "All upstreams failed", 502, request=request
+        )
 
     if request_body_dict:
         await pay_for_request(key, max_cost_for_model, session)
 
-    headers = upstream.prepare_headers(dict(request.headers))
+    for i, upstream in enumerate(upstreams):
+        headers = upstream.prepare_headers(dict(request.headers))
 
-    if is_responses_api:
-        response = await upstream.forward_responses_request(
-            request,
-            path,
-            headers,
-            request_body,
-            key,
-            max_cost_for_model,
-            session,
-            model_obj,
-        )
-    else:
-        response = await upstream.forward_request(
-            request,
-            path,
-            headers,
-            request_body,
-            key,
-            max_cost_for_model,
-            session,
-            model_obj,
-        )
+        try:
+            if is_responses_api:
+                response = await upstream.forward_responses_request(
+                    request,
+                    path,
+                    headers,
+                    request_body,
+                    key,
+                    max_cost_for_model,
+                    session,
+                    model_obj,
+                )
+            else:
+                response = await upstream.forward_request(
+                    request,
+                    path,
+                    headers,
+                    request_body,
+                    key,
+                    max_cost_for_model,
+                    session,
+                    model_obj,
+                )
 
-    if response.status_code != 200:
-        await revert_pay_for_request(key, session, max_cost_for_model)
-        logger.warning(
-            "Upstream request failed, revert payment",
-            extra={
-                "status_code": response.status_code,
-                "path": path,
-                "key_hash": key.hashed_key[:8] + "...",
-                "key_balance": key.balance,
-                "max_cost_for_model": max_cost_for_model,
-                "upstream_headers": response.headers
-                if hasattr(response, "headers")
-                else None,
-            },
-        )
-        # Return the mapped error response generated earlier rather than masking with 502
-        return response
+            if response.status_code != 200:
+                # Check if we should retry (502 Upstream Error or 429 Rate Limit)
+                should_retry = response.status_code in [502, 429, 400, 401, 403, 404]
+                if should_retry and i < len(upstreams) - 1:
+                    error_message = ""
+                    try:
+                        if hasattr(response, "body"):
+                            body_bytes = response.body
+                            data = json.loads(body_bytes)
+                            if "error" in data:
+                                error_data = data["error"]
+                                if isinstance(error_data, dict):
+                                    error_message = error_data.get("message", "")
+                                elif isinstance(error_data, str):
+                                    error_message = error_data
+                    except Exception:
+                        pass
 
-    return response
+                    await upstream.on_upstream_error_redirect(
+                        response.status_code, error_message
+                    )
+
+                    logger.warning(
+                        f"Upstream {upstream.provider_type} returned {response.status_code}, trying next provider",
+                        extra={
+                            "status_code": response.status_code,
+                            "upstream": upstream.provider_type,
+                        },
+                    )
+                    continue
+
+                # 4xx error (user error), or other non-retryable error, or last provider failed
+                await revert_pay_for_request(key, session, max_cost_for_model)
+                logger.warning(
+                    "Upstream request failed, revert payment",
+                    extra={
+                        "status_code": response.status_code,
+                        "path": path,
+                        "key_hash": key.hashed_key[:8] + "...",
+                        "key_balance": key.balance,
+                        "max_cost_for_model": max_cost_for_model,
+                        "upstream_headers": response.headers
+                        if hasattr(response, "headers")
+                        else None,
+                    },
+                )
+                return response
+
+            return response
+
+        except UpstreamError as e:
+            logger.warning(
+                f"Upstream {upstream.provider_type} failed: {e}",
+                extra={"retry": i < len(upstreams) - 1},
+            )
+
+            # 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
+                )
+
+            # Otherwise loop continues to next provider
+            continue
+
+    # Should not be reached given logic above
+    return create_error_response(
+        "upstream_error", "All upstreams failed", 502, request=request
+    )
 
 
 async def get_bearer_token_key(

routstr/upstream/base.py

@@ -15,6 +15,7 @@ from pydantic import BaseModel
 from ..auth import adjust_payment_for_tokens
 from ..core import get_logger
 from ..core.db import ApiKey, AsyncSession, create_session
+from ..core.exceptions import UpstreamError
 
 if TYPE_CHECKING:
     from ..core.db import UpstreamProviderRow
@@ -340,6 +341,20 @@ class BaseUpstreamProvider:
                 message = preview[:500]
         return message, upstream_code
 
+    async def on_upstream_error_redirect(
+        self, status_code: int, error_message: str
+    ) -> None:
+        """Hook called when the proxy redirects to another provider due to an error.
+
+        Subclasses can implement this to perform actions like disabling the provider
+        if it's out of balance.
+
+        Args:
+            status_code: The HTTP status code returned by the upstream
+            error_message: The error message extracted from the upstream response
+        """
+        pass
+
     async def map_upstream_error_response(
         self, request: Request, path: str, upstream_response: httpx.Response
     ) -> Response:
@@ -1148,6 +1163,14 @@ class BaseUpstreamProvider:
             )
 
             if response.status_code != 200:
+                if response.status_code >= 500:
+                    await response.aclose()
+                    await client.aclose()
+                    raise UpstreamError(
+                        f"Upstream returned status {response.status_code}",
+                        status_code=response.status_code,
+                    )
+
                 try:
                     mapped_error = await self.map_upstream_error_response(
                         request, path, response
@@ -1238,6 +1261,9 @@ class BaseUpstreamProvider:
                 background=background_tasks,
             )
 
+        except UpstreamError:
+            raise
+
         except httpx.RequestError as exc:
             await client.aclose()
             error_type = type(exc).__name__
@@ -1265,9 +1291,7 @@ class BaseUpstreamProvider:
             else:
                 error_message = f"Error connecting to upstream service: {error_type}"
 
-            return create_error_response(
-                "upstream_error", error_message, 502, request=request
-            )
+            raise UpstreamError(error_message, status_code=502)
 
         except Exception as exc:
             await client.aclose()
@@ -1380,6 +1404,14 @@ class BaseUpstreamProvider:
             )
 
             if response.status_code != 200:
+                if response.status_code >= 500:
+                    await response.aclose()
+                    await client.aclose()
+                    raise UpstreamError(
+                        f"Upstream returned status {response.status_code}",
+                        status_code=response.status_code,
+                    )
+
                 try:
                     mapped_error = await self.map_upstream_error_response(
                         request, path, response
@@ -1447,6 +1479,9 @@ class BaseUpstreamProvider:
                 background=background_tasks,
             )
 
+        except UpstreamError:
+            raise
+
         except httpx.RequestError as exc:
             await client.aclose()
             error_type = type(exc).__name__
@@ -1474,9 +1509,7 @@ class BaseUpstreamProvider:
             else:
                 error_message = f"Error connecting to upstream service: {error_type}"
 
-            return create_error_response(
-                "upstream_error", error_message, 502, request=request
-            )
+            raise UpstreamError(error_message, status_code=502)
 
         except Exception as exc:
             await client.aclose()

routstr/upstream/ppqai.py

@@ -196,6 +196,37 @@ class PPQAIUpstreamProvider(BaseUpstreamProvider):
             )
             return []
 
+    async def on_upstream_error_redirect(
+        self, status_code: int, error_message: str
+    ) -> None:
+        if "insufficient balance" in error_message.lower():
+            logger.warning(
+                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
+                    and UpstreamProviderRow.api_key == self.api_key
+                )
+                result = await session.exec(statement)
+                provider = result.first()
+
+                if provider:
+                    provider.enabled = False
+                    session.add(provider)
+                    await session.commit()
+
+                    # Trigger re-initialization of providers
+                    # Import here to avoid circular dependency
+                    from ..proxy import reinitialize_upstreams
+
+                    await reinitialize_upstreams()
+
     async def create_account(self) -> dict[str, object]:
         """Create a new PPQ.AI account.
 

tests/integration/test_child_keys.py

@@ -0,0 +1,137 @@
+import secrets
+from typing import Any
+
+import pytest
+from fastapi import HTTPException
+from sqlmodel.ext.asyncio.session import AsyncSession
+
+from routstr.auth import adjust_payment_for_tokens, pay_for_request
+from routstr.balance import create_child_key
+from routstr.core.db import ApiKey
+from routstr.core.settings import settings
+
+
+@pytest.mark.asyncio
+async def test_child_key_flow(integration_session: AsyncSession) -> None:
+    # 1. Create a parent key with balance
+    parent_raw = "parent_test_key_" + secrets.token_hex(4)
+    parent_key = ApiKey(
+        hashed_key=parent_raw,
+        balance=10000,  # 10 sats
+    )
+    integration_session.add(parent_key)
+    await integration_session.commit()
+    await integration_session.refresh(parent_key)
+
+    # Mock settings
+    settings.child_key_cost = 1000  # 1 sat
+
+    # 2. Call create_child_key
+    result = await create_child_key(parent_key, integration_session)
+
+    assert "api_key" in result
+    assert result["cost_msats"] == 1000
+    assert result["parent_balance"] == 9000
+
+    child_key_raw = result["api_key"][3:]  # remove sk-
+
+    # 3. Verify child key exists in DB
+    child_key_db = await integration_session.get(ApiKey, child_key_raw)
+    assert child_key_db is not None
+    assert child_key_db.parent_key_hash == parent_key.hashed_key
+    assert child_key_db.balance == 0
+
+    # 4. Test payment with child key
+    cost = 500
+    await pay_for_request(child_key_db, cost, integration_session)
+
+    # Refresh keys
+    await integration_session.refresh(parent_key)
+    await integration_session.refresh(child_key_db)
+
+    # Parent should be charged
+    assert parent_key.reserved_balance == 500
+    assert parent_key.total_requests == 1
+
+    # Child should have total_requests incremented
+    assert child_key_db.total_requests == 1
+
+    # 5. Test adjustment
+    response_data = {"model": "test-model", "usage": {"total_tokens": 10}}
+
+    # Mock calculate_cost
+    import routstr.auth
+    from routstr.payment.cost_calculation import CostData
+
+    async def mock_calculate_cost(*args: Any, **kwargs: Any) -> CostData:
+        return CostData(
+            base_msats=0, input_msats=200, output_msats=200, total_msats=400
+        )
+
+    # Patch calculate_cost
+    original_calculate_cost = routstr.auth.calculate_cost
+    routstr.auth.calculate_cost = mock_calculate_cost
+
+    try:
+        adjustment = await adjust_payment_for_tokens(
+            child_key_db, response_data, integration_session, 500
+        )
+        assert adjustment["total_msats"] == 400
+
+        # Refresh keys
+        await integration_session.refresh(parent_key)
+        await integration_session.refresh(child_key_db)
+
+        # Parent should have updated balance and total_spent
+        assert parent_key.reserved_balance == 0
+        assert parent_key.balance == 9000 - 400
+        assert (
+            parent_key.total_spent == 1400
+        )  # 1000 for child key creation + 400 for request
+
+        # Child should also have total_spent updated
+        assert child_key_db.total_spent == 400
+
+    finally:
+        routstr.auth.calculate_cost = original_calculate_cost
+
+
+@pytest.mark.asyncio
+async def test_child_key_insufficient_balance(
+    integration_session: AsyncSession,
+) -> None:
+    parent_key = ApiKey(
+        hashed_key="poor_parent_" + secrets.token_hex(4),
+        balance=500,
+    )
+    integration_session.add(parent_key)
+    await integration_session.commit()
+    await integration_session.refresh(parent_key)
+
+    settings.child_key_cost = 1000
+
+    with pytest.raises(HTTPException) as exc:
+        await create_child_key(parent_key, integration_session)
+    assert exc.value.status_code == 402
+
+
+@pytest.mark.asyncio
+async def test_child_key_cannot_create_child(integration_session: AsyncSession) -> None:
+    parent_key = ApiKey(
+        hashed_key="parent_" + secrets.token_hex(4),
+        balance=10000,
+    )
+    child_key = ApiKey(
+        hashed_key="child_" + secrets.token_hex(4),
+        balance=0,
+        parent_key_hash=parent_key.hashed_key,
+    )
+    integration_session.add(parent_key)
+    integration_session.add(child_key)
+    await integration_session.commit()
+    await integration_session.refresh(child_key)
+
+    with pytest.raises(HTTPException) as exc:
+        await create_child_key(child_key, integration_session)
+    assert exc.value.status_code == 400
+    assert "Cannot create a child key for another child key" in str(exc.value.detail)

tests/unit/test_algorithm.py

@@ -10,7 +10,6 @@ os.environ["UPSTREAM_API_KEY"] = "test"
 from routstr.algorithm import (  # noqa: E402
     calculate_model_cost_score,
     get_provider_penalty,
-    should_prefer_model,
 )
 from routstr.payment.models import Architecture, Model, Pricing  # noqa: E402
 
@@ -100,100 +99,3 @@ def test_get_provider_penalty_openrouter() -> None:
     provider = create_test_provider("openrouter", "https://openrouter.ai/api/v1")
     penalty = get_provider_penalty(provider)
     assert penalty == 1.001
-
-
-def test_should_prefer_model_cheaper_wins() -> None:
-    """Test that cheaper model is preferred."""
-    cheap_model = create_test_model("cheap", prompt_price=0.001, completion_price=0.002)
-    expensive_model = create_test_model(
-        "expensive", prompt_price=0.03, completion_price=0.06
-    )
-
-    provider1 = create_test_provider("provider1")
-    provider2 = create_test_provider("provider2")
-
-    # Cheaper model should win
-    assert should_prefer_model(
-        cheap_model, provider1, expensive_model, provider2, "test-alias"
-    )
-
-    # More expensive model should not win
-    assert not should_prefer_model(
-        expensive_model, provider2, cheap_model, provider1, "test-alias"
-    )
-
-
-def test_should_prefer_model_exact_match_wins() -> None:
-    """Test that exact alias match beats cheaper price."""
-    # Make model IDs match the alias differently
-    exact_match = create_test_model(
-        "test-model", prompt_price=0.03, completion_price=0.06
-    )
-    no_match = create_test_model(
-        "other-model", prompt_price=0.001, completion_price=0.002
-    )
-
-    provider1 = create_test_provider("provider1")
-    provider2 = create_test_provider("provider2")
-
-    # Exact match should win even though it's more expensive
-    assert should_prefer_model(
-        exact_match, provider1, no_match, provider2, "test-model"
-    )
-
-
-def test_should_prefer_model_openrouter_slight_penalty() -> None:
-    """Test that OpenRouter has slight penalty compared to other providers."""
-    model1 = create_test_model("model1", prompt_price=0.001, completion_price=0.002)
-    model2 = create_test_model("model2", prompt_price=0.001, completion_price=0.002)
-
-    regular_provider = create_test_provider("regular", "http://provider.com")
-    openrouter_provider = create_test_provider(
-        "openrouter", "https://openrouter.ai/api/v1"
-    )
-
-    # Regular provider should be preferred over OpenRouter at same cost
-    assert should_prefer_model(
-        model1, regular_provider, model2, openrouter_provider, "test-alias"
-    )
-
-    # OpenRouter should not replace regular provider at same cost
-    assert not should_prefer_model(
-        model2, openrouter_provider, model1, regular_provider, "test-alias"
-    )
-
-
-def test_should_prefer_model_openrouter_can_win_if_cheaper() -> None:
-    """Test that OpenRouter can still win if significantly cheaper."""
-    cheap_model = create_test_model(
-        "cheap", prompt_price=0.0001, completion_price=0.0002
-    )
-    expensive_model = create_test_model(
-        "expensive", prompt_price=0.03, completion_price=0.06
-    )
-
-    regular_provider = create_test_provider("regular", "http://provider.com")
-    openrouter_provider = create_test_provider(
-        "openrouter", "https://openrouter.ai/api/v1"
-    )
-
-    # OpenRouter should win if it's much cheaper (even with penalty)
-    assert should_prefer_model(
-        cheap_model,
-        openrouter_provider,
-        expensive_model,
-        regular_provider,
-        "test-alias",
-    )
-
-
-def test_should_prefer_model_same_cost_first_wins() -> None:
-    """Test that when costs are identical, current model is kept."""
-    model1 = create_test_model("model1", prompt_price=0.001, completion_price=0.002)
-    model2 = create_test_model("model2", prompt_price=0.001, completion_price=0.002)
-
-    provider1 = create_test_provider("provider1")
-    provider2 = create_test_provider("provider2")
-
-    # When costs are equal, should not replace
-    assert not should_prefer_model(model2, provider2, model1, provider1, "test-alias")

ui/components/temporary-balances.tsx

@@ -60,7 +60,11 @@ export function TemporaryBalances({
     let totalRequests = 0;
 
     balances.forEach((balance) => {
-      totalBalance += balance.balance || 0;
+      // Only count parents for total balance to avoid double counting
+      // since child keys use parent balance
+      if (!balance.parent_key_hash) {
+        totalBalance += balance.balance || 0;
+      }
       totalSpent += balance.total_spent || 0;
       totalRequests += balance.total_requests || 0;
     });
@@ -72,6 +76,34 @@ export function TemporaryBalances({
     ? calculateTotals(data)
     : { totalBalance: 0, totalSpent: 0, totalRequests: 0 };
 
+  // Group parents and children
+  const hierarchicalData = (() => {
+    if (!data) return [];
+
+    const parents = filteredData.filter((item) => !item.parent_key_hash);
+    const result: (TemporaryBalance & { isChild?: boolean })[] = [];
+
+    parents.forEach((parent) => {
+      result.push(parent);
+      const children = data.filter(
+        (item) => item.parent_key_hash === parent.hashed_key
+      );
+      children.forEach((child) => {
+        result.push({ ...child, isChild: true });
+      });
+    });
+
+    // Add children whose parents didn't match the search or aren't in the list
+    const orphans = filteredData.filter(
+      (item) =>
+        item.parent_key_hash &&
+        !result.some((r) => r.hashed_key === item.hashed_key)
+    );
+    result.push(...orphans.map((o) => ({ ...o, isChild: true })));
+
+    return result;
+  })();
+
   return (
     <>
       <Card className='h-full w-full shadow-sm'>
@@ -182,22 +214,37 @@ export function TemporaryBalances({
                   <div className='text-right'>Expiry Time</div>
                 </div>
 
-                {filteredData.length > 0 ? (
-                  filteredData.map((balance, index) => (
+                {hierarchicalData.length > 0 ? (
+                  hierarchicalData.map((balance, index) => (
                     <div
                       key={index}
                       className={cn(
                         'hover:bg-muted/50 border-t p-3 text-sm transition-colors',
-                        balance.balance === 0 && 'opacity-60'
+                        balance.balance === 0 &&
+                          !balance.isChild &&
+                          'opacity-60',
+                        balance.isChild &&
+                          'ml-4 border-l-2 border-l-blue-200 bg-blue-50/30'
                       )}
                     >
                       {/* Desktop Layout */}
                       <div className='hidden grid-cols-6 gap-2 md:grid'>
-                        <div className='max-w-32 truncate font-mono text-xs break-all'>
+                        <div className='flex max-w-48 items-center gap-2 truncate font-mono text-xs break-all'>
+                          {balance.isChild && (
+                            <span className='rounded bg-blue-100 px-1 py-0.5 text-[10px] font-bold text-blue-700 uppercase'>
+                              Child
+                            </span>
+                          )}
                           {balance.hashed_key}
                         </div>
                         <div className='text-right font-mono'>
-                          {formatBalance(balance.balance)}
+                          {balance.isChild ? (
+                            <span className='text-muted-foreground italic'>
+                              (Parent)
+                            </span>
+                          ) : (
+                            formatBalance(balance.balance)
+                          )}
                         </div>
                         <div className='text-right font-mono'>
                           {formatBalance(balance.total_spent)}
@@ -226,13 +273,20 @@ export function TemporaryBalances({
 
                       {/* Mobile Layout */}
                       <div className='space-y-3 md:hidden'>
-                        <div className='space-y-1'>
-                          <span className='text-muted-foreground text-xs font-medium'>
-                            Key
-                          </span>
-                          <div className='font-mono text-xs break-all'>
-                            {balance.hashed_key}
+                        <div className='flex items-center justify-between'>
+                          <div className='space-y-1'>
+                            <span className='text-muted-foreground text-xs font-medium'>
+                              {balance.isChild ? 'Child Key' : 'Key'}
+                            </span>
+                            <div className='font-mono text-xs break-all'>
+                              {balance.hashed_key}
+                            </div>
                           </div>
+                          {balance.isChild && (
+                            <span className='rounded bg-blue-100 px-1.5 py-0.5 text-[10px] font-bold text-blue-700 uppercase'>
+                              Child
+                            </span>
+                          )}
                         </div>
 
                         <div className='grid grid-cols-2 gap-3'>
@@ -241,7 +295,13 @@ export function TemporaryBalances({
                               Balance
                             </div>
                             <div className='truncate font-mono text-sm'>
-                              {formatBalance(balance.balance)}
+                              {balance.isChild ? (
+                                <span className='text-muted-foreground text-xs italic'>
+                                  (Uses Parent)
+                                </span>
+                              ) : (
+                                formatBalance(balance.balance)
+                              )}
                             </div>
                           </div>
                           <div className='space-y-1'>

ui/lib/api/services/admin.ts

@@ -926,6 +926,7 @@ export const TemporaryBalanceSchema = z.object({
   total_requests: z.number(),
   refund_address: z.string().nullable(),
   key_expiry_time: z.number().nullable(),
+  parent_key_hash: z.string().nullable().optional(),
 });
 
 export type TemporaryBalance = z.infer<typeof TemporaryBalanceSchema>;