bump version

make cache prices configurable

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"

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

@@ -41,6 +41,28 @@ 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,
@@ -83,19 +105,7 @@ 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.get("usage") or {}
     if not isinstance(usage_data, dict):
@@ -109,6 +119,27 @@ async def calculate_cost(
     # 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 — "

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

@@ -97,6 +97,8 @@ 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:

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

@@ -6,13 +6,12 @@ import math
 import traceback
 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
@@ -92,6 +91,11 @@ 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] = {}
 
@@ -106,6 +110,7 @@ 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 = {}
 
@@ -123,10 +128,13 @@ 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
@@ -134,6 +142,19 @@ 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,
@@ -4878,14 +4899,11 @@ 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/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

@@ -45,7 +45,7 @@ class GenericUpstreamProvider(BaseUpstreamProvider):
         )
 
     @classmethod
-    def from_db_row(
+    def _build_from_row(
         cls, provider_row: "UpstreamProviderRow"
     ) -> "GenericUpstreamProvider":
         return cls(

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

@@ -216,9 +216,6 @@ 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",
@@ -413,9 +410,7 @@ 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/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

@@ -90,7 +90,7 @@ 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

@@ -46,7 +46,7 @@ class PPQAIUpstreamProvider(BaseUpstreamProvider):
         )
 
     @classmethod
-    def from_db_row(
+    def _build_from_row(
         cls, provider_row: "UpstreamProviderRow"
     ) -> "PPQAIUpstreamProvider":
         return cls(
@@ -229,17 +229,14 @@ 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/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/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,

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_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/unit/test_cost_calculation_caching.py

@@ -404,6 +404,67 @@ async def test_deepseek_malformed_hit_tokens_coerce_to_zero() -> None:
     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
+
+
 # ============================================================================
 # Test 13: Missing Usage Block
 # ============================================================================

ui/components/add-provider-model-dialog.tsx

@@ -68,6 +68,8 @@ const FormSchema = z.object({
   upstream_provider_id: z.string().default(''),
   input_cost: z.coerce.number().min(0).default(0),
   output_cost: z.coerce.number().min(0).default(0),
+  cache_read_cost: z.coerce.number().min(0).default(0),
+  cache_write_cost: z.coerce.number().min(0).default(0),
   request_cost: z.coerce.number().min(0).default(0),
   image_cost: z.coerce.number().min(0).default(0),
   web_search_cost: z.coerce.number().min(0).default(0),
@@ -125,6 +127,8 @@ export function AddProviderModelDialog({
       upstream_provider_id: '',
       input_cost: 0,
       output_cost: 0,
+      cache_read_cost: 0,
+      cache_write_cost: 0,
       request_cost: 0,
       image_cost: 0,
       web_search_cost: 0,
@@ -190,6 +194,8 @@ export function AddProviderModelDialog({
             : initialData.upstream_provider_id?.toString() || '',
         input_cost: pricing?.prompt ?? 0,
         output_cost: pricing?.completion ?? 0,
+        cache_read_cost: pricing?.input_cache_read ?? 0,
+        cache_write_cost: pricing?.input_cache_write ?? 0,
         request_cost: pricing?.request ?? 0,
         image_cost: pricing?.image ?? 0,
         web_search_cost: pricing?.web_search ?? 0,
@@ -231,6 +237,8 @@ export function AddProviderModelDialog({
         upstream_provider_id: '',
         input_cost: 0,
         output_cost: 0,
+        cache_read_cost: 0,
+        cache_write_cost: 0,
         request_cost: 0,
         image_cost: 0,
         web_search_cost: 0,
@@ -294,6 +302,8 @@ export function AddProviderModelDialog({
     );
     form.setValue('input_cost', pricing?.prompt ?? 0);
     form.setValue('output_cost', pricing?.completion ?? 0);
+    form.setValue('cache_read_cost', pricing?.input_cache_read ?? 0);
+    form.setValue('cache_write_cost', pricing?.input_cache_write ?? 0);
     form.setValue('request_cost', pricing?.request ?? 0);
     form.setValue('image_cost', pricing?.image ?? 0);
     form.setValue('web_search_cost', pricing?.web_search ?? 0);
@@ -365,6 +375,8 @@ export function AddProviderModelDialog({
         pricing: {
           prompt: data.input_cost,
           completion: data.output_cost,
+          input_cache_read: data.cache_read_cost,
+          input_cache_write: data.cache_write_cost,
           request: data.request_cost,
           image: data.image_cost,
           web_search: data.web_search_cost,
@@ -810,6 +822,38 @@ export function AddProviderModelDialog({
                     </FormItem>
                   )}
                 />
+                <FormField
+                  control={form.control}
+                  name='cache_read_cost'
+                  render={({ field }) => (
+                    <FormItem>
+                      <FormLabel>Cache Read Cost</FormLabel>
+                      <FormControl>
+                        <Input type='number' step='0.000001' {...field} />
+                      </FormControl>
+                      <FormDescription>
+                        Discounted cached-input read price per 1M tokens.
+                      </FormDescription>
+                      <FormMessage />
+                    </FormItem>
+                  )}
+                />
+                <FormField
+                  control={form.control}
+                  name='cache_write_cost'
+                  render={({ field }) => (
+                    <FormItem>
+                      <FormLabel>Cache Write Cost</FormLabel>
+                      <FormControl>
+                        <Input type='number' step='0.000001' {...field} />
+                      </FormControl>
+                      <FormDescription>
+                        Cached-input creation price per 1M tokens.
+                      </FormDescription>
+                      <FormMessage />
+                    </FormItem>
+                  )}
+                />
                 <FormField
                   control={form.control}
                   name='request_cost'

ui/lib/api/services/admin.ts

@@ -53,6 +53,8 @@ export const AdminModelPricingSchema = z.object({
   image: z.number().optional(),
   web_search: z.number().optional(),
   internal_reasoning: z.number().optional(),
+  input_cache_read: z.number().optional(),
+  input_cache_write: z.number().optional(),
 });
 
 export const AdminModelArchitectureSchema = z.object({
@@ -149,7 +151,7 @@ export class AdminService {
     if (!pricing) return pricing;
     const result = { ...pricing };
 
-    // Only prompt and completion are per-token and need scaling to per-1M
+    // Token-priced fields are stored per-token by the API and shown per-1M in the UI.
     const convertField = (field: string) => {
       const val = result[field];
       if (val !== undefined && val !== null) {
@@ -164,6 +166,8 @@ export class AdminService {
 
     convertField('prompt');
     convertField('completion');
+    convertField('input_cache_read');
+    convertField('input_cache_write');
 
     // Other fields (request, image, etc.) are already flat fees (per item)
     // so we do NOT scale them.
@@ -177,7 +181,7 @@ export class AdminService {
     if (!pricing) return pricing;
     const result = { ...pricing };
 
-    // Only prompt and completion are per-1M in UI and need scaling down to per-token
+    // Token-priced fields are per-1M in the UI and need scaling down to per-token.
     const convertField = (field: string) => {
       const val = result[field];
       if (val !== undefined && val !== null) {
@@ -190,6 +194,8 @@ export class AdminService {
 
     convertField('prompt');
     convertField('completion');
+    convertField('input_cache_read');
+    convertField('input_cache_write');
 
     // Other fields stay as flat fees