Address EHBP review comments

Browser clients need these EHBP protocol headers visible to JavaScript
so the Tinfoil SDK can detect and decrypt encrypted responses. Without
them, CORS hides the headers, the SDK treats the response as a plaintext
proxy error, and users see 'The provider did not respond to this request.'

Node.js scripts are unaffected (no CORS enforcement).
See ../routstr-chat/TINFOIL_CORS_ISSUE.md for full root-cause analysis.

.env.example

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

.github/workflows/container.yml

@@ -14,6 +14,14 @@ jobs:
     steps:
       - name: Checkout code
         uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
+
+      - name: Resolve git metadata
+        id: gitmeta
+        run: |
+          echo "sha=$(git rev-parse --short=7 HEAD)" >> "$GITHUB_OUTPUT"
+          echo "tag=$(git describe --tags --exact-match HEAD 2>/dev/null || true)" >> "$GITHUB_OUTPUT"
 
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v2
@@ -29,7 +37,11 @@ jobs:
         uses: docker/build-push-action@v4
         with:
           context: .
+          file: Dockerfile.full
           push: true
+          build-args: |
+            GIT_COMMIT=${{ steps.gitmeta.outputs.sha }}
+            GIT_TAG=${{ steps.gitmeta.outputs.tag }}
           tags: |
             ghcr.io/routstr/proxy:latest
             ghcr.io/routstr/core:latest

.gitignore

@@ -38,3 +38,4 @@ proof_backups
 
 *.todo
 ui_out
+.worktrees

Dockerfile

@@ -1,26 +1,29 @@
-FROM ghcr.io/astral-sh/uv:python3.11-alpine
+FROM ghcr.io/astral-sh/uv:python3.11-bookworm-slim
 
-# Install system dependencies required for secp256k1
-RUN apk add --no-cache \
-    pkgconf \
-    build-base \
-    automake \
-    autoconf \
-    libtool \
-    m4 \
-    perl
-RUN apk add git
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends \
+        git \
+        build-essential \
+        pkg-config \
+        libsecp256k1-dev \
+        autoconf \
+        automake \
+        libtool \
+    && rm -rf /var/lib/apt/lists/*
 
 COPY uv.lock pyproject.toml ./
 RUN mkdir -p /routstr
 
-RUN uv add git+https://github.com/saschanaz/secp256k1-py.git#branch=upgrade060
-# RUN uv sync
+RUN uv sync --frozen --no-dev --no-install-project
 
 WORKDIR /app
 
 COPY . .
 
+ARG GIT_COMMIT=""
+ARG GIT_TAG=""
+ENV GIT_COMMIT=${GIT_COMMIT}
+ENV GIT_TAG=${GIT_TAG}
 ENV PORT=8000
 ENV PYTHONUNBUFFERED=1
 

Dockerfile.full

@@ -4,7 +4,7 @@ FROM node:23-alpine AS ui-builder
 WORKDIR /app/ui
 
 # Install pnpm
-RUN corepack enable pnpm && corepack prepare pnpm@latest --activate
+RUN corepack enable pnpm && corepack prepare pnpm@10.15.0 --activate
 
 # Copy UI source
 COPY ui/package.json ui/pnpm-lock.yaml* ./
@@ -16,35 +16,38 @@ ENV NEXT_TELEMETRY_DISABLED=1
 RUN pnpm run build
 
 # Stage 2: Build the Routstr Node
-FROM ghcr.io/astral-sh/uv:python3.11-alpine AS runner
+FROM ghcr.io/astral-sh/uv:python3.11-bookworm-slim AS runner
 
-# Install system dependencies
-RUN apk add --no-cache \
-    pkgconf \
-    build-base \
-    automake \
-    autoconf \
-    libtool \
-    m4 \
-    perl \
-    git
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends \
+        git \
+        build-essential \
+        pkg-config \
+        libsecp256k1-dev \
+        autoconf \
+        automake \
+        libtool \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY uv.lock pyproject.toml ./
+
+RUN uv sync --no-dev --no-install-project
 
 WORKDIR /app
 
-# Copy the rest of the application (required for uv sync to find the package)
 COPY . .
 
-# Install dependencies including the specific secp256k1 branch
-RUN uv add git+https://github.com/saschanaz/secp256k1-py.git#branch=upgrade060
-RUN uv sync --no-dev
-
 # Copy the built UI from the ui-builder stage
 COPY --from=ui-builder /app/ui/out ./ui_out
 
+ARG GIT_COMMIT=""
+ARG GIT_TAG=""
+ENV GIT_COMMIT=${GIT_COMMIT}
+ENV GIT_TAG=${GIT_TAG}
 ENV PORT=8000
 ENV PYTHONUNBUFFERED=1
 
 EXPOSE 8000
 
 # Run the application
-CMD ["/app/.venv/bin/fastapi", "run", "routstr", "--host", "0.0.0.0"]
+CMD ["/.venv/bin/fastapi", "run", "routstr", "--host", "0.0.0.0"]

Makefile

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

compose.yml

@@ -8,8 +8,9 @@ services:
       args:
         # NEXT_PUBLIC_API_URL: ${NEXT_PUBLIC_API_URL:-http://127.0.0.1:8000}
         NEXT_PUBLIC_ADMIN_API_KEY: ${NEXT_PUBLIC_ADMIN_API_KEY:-}
+    user: root
     volumes:
-      - ./ui_out:/output
+      - ./ui_out:/output:z
     command:
       ["sh", "-c", "mkdir -p /output && cp -r /app/built/. /output/ && echo 'UI build copied to mounted volume' && ls -la /output/ && echo 'UI built and ready' && tail -f /dev/null"]
 
@@ -18,10 +19,10 @@ services:
     depends_on:
       - ui
     volumes:
-      - .:/app
-      - ./logs:/app/logs
+      - .:/app:z
+      - ./logs:/app/logs:z
       - tor-data:/var/lib/tor:ro
-      - ./ui_out:/app/ui_out:ro
+      - ./ui_out:/app/ui_out:ro,z
     env_file:
       - .env
     environment:

docs/api/endpoints.md

@@ -396,6 +396,8 @@ Authorization: Bearer sk-...
 }
 ```
 
+`balance` is the spendable balance used by request admission.
+
 ### Check Balance
 
 Get current wallet balance.

docs/api/overview.md

@@ -155,7 +155,7 @@ The response includes your change in the same header:
 X-Cashu: cashuA7k2mNp4...
 ```
 
-This is fully stateless—no session, no `/v1/balance/refund` call needed. However, **streaming does not work with `X-Cashu`** because the refund can only be calculated after the full response is generated.
+This is fully stateless—no session, no `/v1/balance/refund` call needed. However, **streaming does not work with `X-Cashu`** because the refund can only be calculated after the full response is generated. If you lose the `X-Cashu` response header before claiming your change, you can reclaim the refund via `POST /v1/wallet/refund` by supplying the original payment token in the `x-cashu` header.
 
 ## Response Headers
 

docs/client/payments.md

@@ -53,9 +53,11 @@ If your balance runs low, you don't need a new key. You can top up the existing
 
 ### Via Lightning
 
-`POST /lightning/invoice` with `{"amount_sats": 1000, "purpose": "topup", "api_key": "sk-..."}`.
+`POST /lightning/invoice` with `Authorization: Bearer sk-...` header and body `{"amount_sats": 1000, "purpose": "topup"}`.
 *Once paid, the funds are added to your existing key.*
 
+> Legacy: the endpoint is also exposed at `/v1/balance/lightning/invoice`, and accepts an `api_key` field in the body as a fallback for older clients. New integrations should use the RIP-08 path with the `Authorization` header.
+
 ### Via Cashu
 
 `POST /v1/balance/topup` with `{"cashu_token": "..."}` and `Authorization: Bearer sk-...`.

docs/ehbp-proxy-support.md

@@ -0,0 +1,158 @@
+# EHBP Proxy Support for Tinfoil Models
+
+## Problem
+
+The SDK's `SecureClient.fetch` encrypts request bodies with HPKE (EHBP protocol)
+and sends them to the Routstr provider. The Routstr proxy had no EHBP handling:
+
+1. It tried to `json.loads()` the binary HPKE-sealed body → failed with a 400
+2. The upstream PPQ.AI public endpoint (`/v1/chat/completions`) doesn't speak
+   EHBP, so the response had no `Ehbp-Response-Nonce` header
+3. `SecureClient` threw `Missing Ehbp-Response-Nonce header` because it expects
+   every response from an EHBP-configured `baseURL` to carry that header
+
+## Root cause
+
+PPQ.AI exposes EHBP-aware inference at `/private/`, separate from the public
+`/v1/` endpoint. The Routstr proxy was forwarding to `/v1/` (the public
+endpoint) instead of `/private/` (the enclave endpoint). The public endpoint
+can't decrypt the body, returns a normal HTTP response, and the SDK can't
+decrypt it because there's no nonce header.
+
+The PPQ private-mode proxy (`ppq-private-mode-proxy/lib/proxy.ts`) shows the
+correct pattern: `SecureClient` talks to `api.ppq.ai/private/v1/chat/completions`,
+which decrypts inside the attested enclave and returns an EHBP-encrypted
+response with the `Ehbp-Response-Nonce` header.
+
+## What was changed
+
+### `routstr/proxy.py`
+
+Detects EHBP requests by checking for the `Ehbp-Encapsulated-Key` header (set
+by the EHBP transport on every encrypted request). For EHBP requests:
+
+- Skips JSON body parsing (the body is binary ciphertext, not JSON)
+- Reads the model ID from the `X-Routstr-Model` header (set by the SDK) instead
+  of from `body.model`
+- Routes through new `forward_ehbp_request` (bearer auth) and
+  `forward_ehbp_x_cashu_request` (x-cashu auth) methods
+- Skips reactive 400 param correction (can't parse encrypted response body)
+- Still charges the user via `pay_for_request` (uses `max_cost_for_model` from
+  the model registry, not the body)
+
+### `routstr/upstream/base.py`
+
+Keeps EHBP as an explicit opt-in provider capability instead of making every
+upstream provider appear EHBP-capable:
+
+- `supports_ehbp = False` by default
+- `get_ehbp_forwarding_target(path, model_obj)` raises `NotImplementedError`
+  unless a provider opts in and returns a provider-specific EHBP target
+
+The actual EHBP forwarding logic does **not** live in `base.py`.
+
+### `routstr/upstream/ehbp.py`
+
+Contains the shared opaque EHBP transport helpers:
+
+- `EHBPForwardingTarget` — provider-specific target URL plus extra headers
+- `forward_ehbp_request()` — forwards the raw encrypted body to an EHBP-capable
+  provider, streams the encrypted response back untouched, and finalizes bearer
+  billing at max cost because usage is encrypted
+- `forward_ehbp_x_cashu_request()` — redeems the Cashu token, forwards raw,
+  refunds the full token on upstream failure, and refunds any value above
+  `max_cost_for_model` on success
+
+### `routstr/upstream/ppqai.py`
+
+- Sets `supports_ehbp = True`.
+- Implements `get_ehbp_forwarding_target()` to forward to
+  `https://api.ppq.ai/private/v1/...` — the PPQ.AI enclave endpoint that
+  understands EHBP and returns the `Ehbp-Response-Nonce` header.
+- Adds `X-Private-Model` with the model's `forwarded_model_id` (e.g.
+  `private/kimi-k2-6`). PPQ.AI's billing layer needs this since it can't
+  decrypt the body.
+
+## Why it's done this way
+
+The proxy is a **blind relay** for EHBP requests. It cannot decrypt the body
+(only the attested enclave can), so it must:
+
+1. Get the model ID from a header, not the body
+2. Forward the raw bytes without parsing or transformation
+3. Stream the response back without SSE/cost parsing
+4. Pass through EHBP protocol headers (`Ehbp-Encapsulated-Key` on request,
+   `Ehbp-Response-Nonce` on response)
+
+Cost tracking happens at the proxy level using `max_cost_for_model` from the
+model registry. Because EHBP responses are encrypted, Routstr cannot reconcile
+against token usage. Bearer requests reserve and then finalize max-cost billing;
+X-Cashu requests redeem the token and refund any amount above max cost.
+
+## End-to-end flow
+
+```
+SDK                                  Routstr Proxy                    PPQ.AI /private/
+  │                                       │                                │
+  │── X-Routstr-Model: tinfoil-kimi-k2-6 ─│                                │
+  │── Ehbp-Encapsulated-Key: <hex> ───────│                                │
+  │── Authorization: Bearer <cashu> ──────│                                │
+  │── body = HPKE-encrypted(kimi-k2-6) ───│                                │
+  │                                       │                                │
+  │                                  detects Ehbp-Encapsulated-Key         │
+  │                                  reads model from X-Routstr-Model      │
+  │                                  does billing/routing                  │
+  │                                       │                                │
+  │                                  adds X-Private-Model: private/kimi-k2-6
+  │                                  forwards raw body to /private/v1/...  │
+  │                                       │──────────────────────────────▶│
+  │                                       │                          enclave decrypts
+  │                                       │                          runs inference
+  │                                       │◀── Ehbp-Response-Nonce ──────│
+  │                                       │◀── encrypted response ────────│
+  │                                       │                                │
+  │                                  streams response back untouched       │
+  │◀── encrypted response ────────────────│                                │
+  │                                        │                                │
+  SecureClient reads nonce, decrypts       │                                │
+  SDK SSE processing sees plaintext        │                                │
+```
+
+## Model ID mapping
+
+Three parties see three different model IDs:
+
+| Party | Header/Body | Value | Source |
+|---|---|---|---|
+| Routstr proxy | `X-Routstr-Model` header | `tinfoil-kimi-k2-6` | SDK sends full caller-facing id |
+| PPQ.AI billing | `X-Private-Model` header | `private/kimi-k2-6` | Proxy sends `forwarded_model_id` |
+| Tinfoil enclave | `body.model` (encrypted) | `kimi-k2-6` | SDK strips `tinfoil-` prefix before encryption |
+
+## Implementation status
+
+A dedicated `TinfoilUpstreamProvider` (`routstr/upstream/tinfoil.py`) now
+implements the direct blind-upstream pattern described above. The shared EHBP
+helpers in `routstr/upstream/ehbp.py` were extended to:
+
+- Request usage metrics via `X-Tinfoil-Request-Usage-Metrics: true`.
+- Parse `X-Tinfoil-Usage-Metrics` from the response header (non-streaming).
+- Override the forwarding URL with `X-Tinfoil-Enclave-Url` when the SDK sends it.
+- Finalize bearer billing with actual token cost via `adjust_payment_for_tokens`.
+- Compute X-Cashu refunds from actual cost instead of max cost.
+
+See `docs/tinfoil-direct-integration.md` for the full implementation notes.
+
+## Not yet tested
+
+These changes were written without integration testing due to the complexity
+of the full stack (SDK + proxy + PPQ.AI enclave + Cashu mint). Needs end-to-end
+verification with a real `tinfoil-*` model request.
+
+Important assumptions to verify:
+
+- PPQ.AI accepts `/private/v1/...` with `X-Private-Model`.
+- PPQ.AI enforces consistency between `X-Private-Model` and the encrypted
+  `body.model`, otherwise a malicious client could understate
+  `X-Routstr-Model` for billing.
+- SDK behavior on non-2xx proxy-generated errors that do not carry
+  `Ehbp-Response-Nonce`.

docs/provider/configuration.md

@@ -78,9 +78,15 @@ Which mints to accept payments from:
 
 Automatic profit withdrawal:
 
-| Setting               | Description                     |
-| --------------------- | ------------------------------- |
-| **Lightning Address** | Your LN address for withdrawals |
+| Setting                              | Description                                                                                                | Default |
+| ------------------------------------ | ---------------------------------------------------------------------------------------------------------- | ------- |
+| **Lightning Address**                | Your LN address for withdrawals                                                                            | —       |
+| **Minimum Payout (sat)**             | Min available balance (in sats) before profit is paid out. Applies to both `sat` and `msat` mints (auto-converted). | `210`   |
+| **Payout Interval (seconds)**        | How often the payout loop wakes up and checks balances                                                     | `900`   |
+
+All payout amounts must be positive. Set the minimums above your wallet's
+minimum-invoice constraint (typically 1 sat) and high enough to amortise
+routing fees.
 
 ### Security
 
@@ -126,6 +132,8 @@ Use environment variables for:
 | `ENABLE_ANALYTICS_SHARING` | Enable usage analytics sharing to Nostr | `true`                         |
 | `CASHU_MINTS`        | Comma-separated mint URLs         | `https://mint.minibits.cash/Bitcoin` |
 | `RECEIVE_LN_ADDRESS` | Lightning address for withdrawals | —                                    |
+| `MIN_PAYOUT_SAT`     | Min payout balance in sats (applies to all mints) | `210`                |
+| `PAYOUT_INTERVAL_SECONDS` | Payout loop interval (seconds) | `900`                            |
 | `TOR_PROXY_URL`      | SOCKS5 proxy for Tor              | `socks5://127.0.0.1:9050`            |
 | `CORS_ORIGINS`       | Allowed CORS origins              | `*`                                  |
 | `RELAYS`             | Nostr relays (comma-separated)    | (default set)                        |

docs/provider/deployment.md

@@ -2,6 +2,57 @@
 
 Production deployment guide for Routstr Provider nodes.
 
+## All-in-One Docker Image (Preferred)
+
+The easiest way to deploy Routstr is using the all-in-one Docker image from Docker Hub, which includes both the FastAPI backend and the Next.js admin dashboard in a single container.
+
+### Quick Start
+
+```bash
+docker run -d \
+  --name routstr \
+  -p 8000:8000 \
+  -v routstr-data:/app/data \
+  -e DATABASE_URL="sqlite:////app/data/routstr.db" \
+  9qeklajc/routstr:latest
+```
+
+Access your node:
+- **API & Admin Dashboard**: http://localhost:8000
+
+### Docker Compose Setup
+
+Create `docker-compose.yml`:
+
+```yaml
+version: '3.8'
+
+services:
+  routstr:
+    image: 9qeklajc/routstr:latest
+    container_name: routstr
+    restart: unless-stopped
+    ports:
+      - "8000:8000"
+    volumes:
+      - routstr-data:/app/data
+    environment:
+      DATABASE_URL: "sqlite:////app/data/routstr.db"
+      ADMIN_KEY: "your-secure-admin-key"
+      LOG_LEVEL: "info"
+
+volumes:
+  routstr-data:
+```
+
+Start it:
+
+```bash
+docker compose up -d
+```
+
+---
+
 ## Docker Compose (Recommended)
 
 For production, use Docker Compose with persistent storage and optional Tor support.

docs/tinfoil-direct-integration.md

@@ -0,0 +1,460 @@
+# Tinfoil / PPQ Private-Mode Integration Notes
+
+This document summarizes the current options for integrating Tinfoil/PPQ private models with Routstr, based on the EHBP work in this branch and local testing against `ppq-private-mode-proxy`.
+
+## Background
+
+PPQ private models run behind a Tinfoil/EHBP flow:
+
+- Request bodies are HPKE-encrypted by a Tinfoil client.
+- The PPQ `/private/` endpoint routes ciphertext to the attested enclave.
+- The enclave decrypts, runs inference, and returns an encrypted response.
+- The caller's Tinfoil client decrypts the response locally.
+
+The PPQ private-mode proxy (`~/projects/ppq-private-mode-proxy`) uses this pattern with the JavaScript `tinfoil` SDK:
+
+```ts
+import { SecureClient } from "tinfoil";
+
+const apiBase = "https://api.ppq.ai";
+
+const client = new SecureClient({
+  baseURL: `${apiBase}/private/`,
+  attestationBundleURL: `${apiBase}/private`,
+  transport: "ehbp",
+});
+
+await client.ready();
+
+const response = await client.fetch(`${apiBase}/private/v1/chat/completions`, {
+  method: "POST",
+  headers: {
+    "Content-Type": "application/json",
+    Authorization: `Bearer ${process.env.PPQ_API_KEY}`,
+    "X-Private-Model": "private/gpt-oss-120b",
+    "x-query-source": "api",
+  },
+  body: JSON.stringify({
+    model: "gpt-oss-120b", // enclave-internal model id
+    messages: [{ role: "user", content: "Hello" }],
+  }),
+});
+
+const json = await response.json();
+console.log(json.usage);
+```
+
+`X-Private-Model` carries the PPQ-facing private model id, while the encrypted JSON body uses the enclave-internal model id without the `private/` prefix.
+
+## PPQ private model pricing
+
+PPQ exposes private model pricing through:
+
+```text
+GET https://api.ppq.ai/v1/models?type=all
+```
+
+Filter models whose IDs start with `private/`.
+
+Example private pricing observed:
+
+| Model | Input USD / 1M tokens | Output USD / 1M tokens |
+|---|---:|---:|
+| `private/gpt-oss-120b` | `0.79125` | `1.31875` |
+| `private/llama3-3-70b` | `1.84625` | `2.90125` |
+| `private/qwen3-vl-30b` | `1.31875` | `4.22` |
+| `private/glm-5-2` | `1.5825` | `5.53875` |
+| `private/gemma4-31b` | `0.47475` | `1.055` |
+| `private/kimi-k2-6` | `1.5825` | `5.53875` |
+
+The `ppq-private-mode-proxy` commit `ba984214793d3bca0f7d046b6955d42abc1c6843` changed OpenClaw display metadata to align with a 5% API margin. The live PPQ model endpoint returns the more precise rates above, which already include that margin.
+
+Actual PPQ private billing is:
+
+```text
+price_usd =
+  input_tokens  * input_per_1M_tokens  / 1_000_000
++ output_tokens * output_per_1M_tokens / 1_000_000
+```
+
+A test request to `private/gpt-oss-120b` produced:
+
+```json
+{
+  "input_count": 74,
+  "output_count": 3,
+  "price_in_usd": 0.00006250875
+}
+```
+
+which exactly matches:
+
+```text
+74 * 0.79125 / 1_000_000 + 3 * 1.31875 / 1_000_000
+= 0.00006250875 USD
+```
+
+## Integration architectures
+
+There are three materially different ways Routstr could integrate Tinfoil/PPQ private inference.
+
+## Option A: User integrates Tinfoil directly
+
+```text
+User app / SDK
+  -> Tinfoil SecureClient / TinfoilAI
+    -> EHBP-encrypted request
+      -> Tinfoil/PPQ private enclave
+```
+
+Properties:
+
+- Best privacy for the user.
+- Routstr is not in the request path.
+- User's client encrypts requests and decrypts responses.
+- Usage is visible to the user's app after decryption.
+- Billing is handled directly by Tinfoil/PPQ.
+
+Example with Tinfoil's OpenAI-compatible client:
+
+```ts
+import { TinfoilAI } from "tinfoil";
+
+const client = new TinfoilAI({
+  apiKey: process.env.TINFOIL_API_KEY,
+  transport: "ehbp",
+});
+
+const res = await client.chat.completions.create({
+  model: "llama3-3-70b",
+  messages: [{ role: "user", content: "Hello" }],
+});
+
+console.log(res.choices[0].message.content);
+console.log(res.usage);
+```
+
+This is not a Routstr marketplace flow unless Routstr only acts as discovery/UI around direct Tinfoil/PPQ usage.
+
+## Option B: Routstr integrates Tinfoil as an upstream client
+
+```text
+User -> Routstr plaintext request
+  -> Routstr Tinfoil SecureClient encrypts to PPQ/Tinfoil
+    -> PPQ private enclave
+  -> Routstr receives decrypted response
+  -> Routstr bills from decrypted usage
+  -> Routstr returns plaintext response to user
+```
+
+Properties:
+
+- Easier exact billing.
+- Routstr can read the decrypted OpenAI response and `usage` object.
+- Routstr can charge exact PPQ token pricing.
+- Privacy is different: the user sends plaintext to Routstr, and Routstr sees prompts/responses.
+- End-to-end encryption is only Routstr-to-enclave, not user-to-enclave.
+
+This should be considered a separate product/provider mode, not the same as an end-to-end private relay.
+
+Practical implementation options:
+
+1. Run a Node sidecar that uses the `tinfoil` npm package and expose it as a local HTTP upstream to Routstr.
+2. Port EHBP client behavior to Python.
+3. Reuse or adapt `ppq-private-mode-proxy` as a local upstream.
+
+A Node sidecar is probably the quickest implementation path because `ppq-private-mode-proxy` already demonstrates the full flow.
+
+## Option C: Routstr uses Tinfoil as a direct blind upstream
+
+This is the current branch's design intent and is likely the best fit if Tinfoil/PPQ exposes usage metadata headers:
+
+```text
+User Tinfoil SecureClient
+  -> encrypted request body
+    -> Routstr proxy
+      -> Tinfoil/PPQ private API
+         with X-Tinfoil-Request-Usage-Metrics: true
+        -> PPQ private enclave
+      <- encrypted response
+         plus X-Tinfoil-Usage-Metrics / cost headers
+    <- Routstr proxy
+  -> user decrypts response
+```
+
+In this mode Routstr is still a normal upstream proxy from the user's point of view, but the upstream is Tinfoil/PPQ private inference and the body remains opaque to Routstr.
+
+Properties:
+
+- Strongest privacy with Routstr in the path.
+- Routstr never sees plaintext prompt or plaintext response.
+- Routstr can authenticate and route based on plaintext headers.
+- Routstr should request Tinfoil usage metadata by adding `X-Tinfoil-Request-Usage-Metrics: true` to the upstream request.
+- Tinfoil documents `X-Tinfoil-Usage-Metrics` as an upstream response header for non-streaming requests.
+- For streaming requests, Tinfoil documents `X-Tinfoil-Usage-Metrics` as an HTTP trailer available only after the response body completes.
+- If Tinfoil/PPQ returns `X-Tinfoil-Usage-Metrics` or a cost header on the encrypted response, Routstr can bill exactly without decrypting the body.
+- If usage is only present inside the encrypted response body, Routstr still cannot read it and exact billing is not possible without a separate metadata path.
+
+This is the only architecture that preserves end-to-end encryption from the user to the PPQ/Tinfoil enclave while still letting Routstr mediate payment. The key requirement is that usage/cost metadata must be returned outside the encrypted body, ideally as a response header available before body streaming begins.
+
+## Current Routstr problem
+
+The current EHBP implementation charges successful EHBP requests at `max_cost_for_model` because Routstr cannot decrypt the response body:
+
+```text
+successful EHBP request -> charge full reserved max cost
+```
+
+That is incorrect for PPQ private models. Max cost should be only a reservation/solvency ceiling. Final charge should use actual PPQ private token pricing.
+
+Desired behavior:
+
+```text
+reserve max cost
+forward encrypted request
+obtain actual usage/cost metadata
+finalize actual cost
+refund/release the difference
+```
+
+## Usage/cost metadata requirement
+
+For blind-relay exact billing, PPQ should return one of the following outside the encrypted response body:
+
+```http
+X-PPQ-Cost-USD: 0.00006250875
+```
+
+or:
+
+```http
+X-Private-Usage-Metrics: input=74,output=3
+```
+
+or:
+
+```http
+X-Tinfoil-Usage-Metrics: prompt=74,completion=3,total=77
+```
+
+Tinfoil proxy documentation references a usage-metrics flow where a proxy can request usage via:
+
+```http
+X-Tinfoil-Request-Usage-Metrics: true
+```
+
+and read usage from:
+
+```http
+X-Tinfoil-Usage-Metrics
+```
+
+Docs/example references:
+
+- https://docs.tinfoil.sh/guides/proxy-server
+- https://github.com/tinfoilsh/encrypted-request-proxy-example
+
+During local PPQ testing, PPQ responses included this CORS exposure header:
+
+```http
+Access-Control-Expose-Headers: Ehbp-Response-Nonce, X-Private-Usage-Metrics, X-Encrypted-Usage-Metrics, X-Tinfoil-Usage-Metrics
+```
+
+However, the actual tested non-streaming response did not include any of these usage headers, even when `X-Tinfoil-Request-Usage-Metrics: true` was sent.
+
+The decrypted body did include normal OpenAI usage, but only the decrypting Tinfoil client can see that body.
+
+## Query-history fallback
+
+PPQ's query history endpoint exposes actual usage and cost:
+
+```text
+GET https://api.ppq.ai/queries/history?page=1&page_count=...
+```
+
+A record includes:
+
+```json
+{
+  "timestamp": "...",
+  "model": "private/gpt-oss-120b",
+  "input_count": 74,
+  "output_count": 3,
+  "price_in_usd": 0.00006250875,
+  "query_type": "chat_completion",
+  "query_source": "api"
+}
+```
+
+This could be used as a fallback, but it is less robust than response headers/trailers because matching a request to a history row can be race-prone under concurrency. It would need a reliable request identifier or metadata field that PPQ stores in history.
+
+## Recommended Routstr direction
+
+Use Tinfoil/PPQ as a direct blind upstream and have Routstr explicitly request usage metadata:
+
+```text
+User encrypts body
+Routstr reserves max cost
+Routstr forwards encrypted body to Tinfoil/PPQ /private/
+Routstr includes X-Tinfoil-Request-Usage-Metrics: true
+Tinfoil/PPQ returns X-Tinfoil-Usage-Metrics as a response header for non-streaming,
+or as an HTTP trailer after the body completes for streaming
+Routstr finalizes exact charge
+User decrypts encrypted response
+```
+
+This preserves both:
+
+- privacy: Routstr cannot read prompts/responses;
+- exact billing: Routstr can charge actual PPQ private model cost, assuming Tinfoil/PPQ returns usage/cost metadata outside the encrypted body.
+
+Implementation steps:
+
+1. Update PPQ model fetching to include private models:
+
+   ```text
+   GET https://api.ppq.ai/v1/models?type=all
+   ```
+
+2. Register `private/*` models and any Routstr-facing aliases with correct `forwarded_model_id`.
+
+3. Keep max-cost reservation for bearer keys and X-Cashu solvency checks.
+
+4. Add parsing support for possible usage/cost headers:
+
+   ```http
+   X-PPQ-Cost-USD
+   X-Private-Usage-Metrics
+   X-Tinfoil-Usage-Metrics
+   X-Encrypted-Usage-Metrics
+   ```
+
+5. Finalize by actual cost instead of max cost:
+
+   ```text
+   actual_msats = ceil((actual_usd / sats_usd_price()) * 1000)
+   actual_msats = max(actual_msats, settings.min_request_msat)
+   actual_msats = min(actual_msats, reserved_msats)
+   ```
+
+   or, if only token counts are available:
+
+   ```text
+   actual_usd =
+     input_tokens  * input_per_1M_tokens  / 1_000_000
+   + output_tokens * output_per_1M_tokens / 1_000_000
+   ```
+
+6. If usage/cost metadata is missing, choose an explicit policy:
+
+   - fail closed and refund/revert;
+   - query PPQ history as a fallback;
+   - fallback to max-cost billing only if explicitly configured and clearly disclosed.
+
+Silent max-cost billing should not be the default for PPQ private requests.
+
+## X-Cashu consideration
+
+For bearer-auth requests, finalization can happen after the response stream completes if usage is delivered as a trailer.
+
+For `X-Cashu`, Routstr needs to return the refund token in the response headers. If usage/cost is only available after consuming the encrypted response stream, Routstr may need to buffer EHBP responses before sending them to the client so it can compute the refund amount first.
+
+Possible approaches:
+
+1. Prefer a non-trailer response header with actual cost, available before streaming body starts.
+2. Buffer EHBP X-Cashu responses and then return `X-Cashu` refund.
+3. Introduce a later/refund-claim mechanism, which would be a larger protocol change.
+
+## Summary
+
+- PPQ private models are billed per actual input/output tokens.
+- Private model rates are available from `GET /v1/models?type=all`.
+- Current Routstr EHBP billing at max cost is wrong for PPQ private models.
+- Direct Tinfoil integration inside Routstr would enable exact usage billing but would make Routstr see plaintext.
+- A blind EHBP relay preserves privacy but requires PPQ/Tinfoil to expose usage/cost in plaintext headers/trailers.
+- The preferred solution is to keep Routstr blind and have PPQ return billing metadata outside the encrypted body.
+
+## Implementation status
+
+Direct Tinfoil upstream integration is implemented in `routstr/upstream/tinfoil.py`
+and `routstr/upstream/ehbp.py`.
+
+### What was built
+
+- `TinfoilUpstreamProvider` (`provider_type = "tinfoil"`):
+  - Base URL: `https://inference.tinfoil.sh`
+  - Fetches models from the public `GET /v1/models` endpoint (no auth needed).
+  - Parses Tinfoil's pricing (`inputTokenPricePer1M`, `outputTokenPricePer1M`,
+    `requestPrice`) into the standard `Model`/`Pricing` schema.
+  - `supports_ehbp = True` — acts as a blind EHBP relay.
+  - `get_ehbp_forwarding_target()` returns a target that includes
+    `X-Tinfoil-Request-Usage-Metrics: true`.
+  - `forward_get_request()` proxies `/attestation` to `https://atc.tinfoil.sh/attestation`
+    so the SDK can fetch attestation bundles through Routstr.
+  - Registered in `routstr/upstream/__init__.py` and seeded from
+    `TINFOIL_API_KEY` env var.
+
+- `routstr/upstream/ehbp.py`:
+  - `parse_tinfoil_usage_metrics()` parses `prompt=N,completion=N[,total=N]`
+    into an OpenAI-style usage dict.
+  - `_resolve_ehbp_target_url()` overrides the forwarding URL with
+    `X-Tinfoil-Enclave-Url` when the SDK sends it.
+  - `_strip_proxy_headers()` removes `X-Routstr-Model`,
+    `X-Tinfoil-Enclave-Url`, and `X-Tinfoil-Request-Usage-Metrics` before
+    forwarding to the enclave.
+  - `_compute_ehbp_actual_cost()` converts the usage header into msats via
+    `calculate_cost()`, clamped to `[min_request_msat, max_cost_for_model]`.
+  - `forward_ehbp_request()` (bearer auth): if `X-Tinfoil-Usage-Metrics` is
+    present in the response header, finalizes with `adjust_payment_for_tokens()`
+    for exact billing; otherwise falls back to max-cost.
+  - `forward_ehbp_x_cashu_request()`: if usage is available, computes the
+    refund from actual cost instead of max cost.
+
+- `routstr/proxy.py`: `/attestation` and `/.well-known/` paths are forwarded
+  to all enabled upstreams without model/cost/auth lookups.
+
+### Billing behavior
+
+| Request shape | Usage source | Billing |
+|---|---|---|
+| Bearer, non-streaming | `X-Tinfoil-Usage-Metrics` response header | Exact token cost via `adjust_payment_for_tokens` |
+| Bearer, streaming | `X-Tinfoil-Usage-Metrics` HTTP trailer | Exact token cost (h11 captures trailers) |
+| Bearer, no usage header/trailer | N/A | Max-cost fallback |
+| X-Cashu, non-streaming | `X-Tinfoil-Usage-Metrics` response header | Refund = `redeemed - actual_cost` |
+| X-Cashu, streaming | `X-Tinfoil-Usage-Metrics` HTTP trailer | Refund = `redeemed - actual_cost` (h11 captures trailers) |
+| X-Cashu, no usage header/trailer | N/A | Refund = `redeemed - max_cost` |
+
+### Cost response headers
+
+Since EHBP response bodies are opaque encrypted blobs, per-request cost cannot
+be injected into the JSON body (as done in the normal proxy flow). Instead,
+Routstr returns cost info as response headers:
+
+| Header | Auth | Description |
+|---|---|---|
+| `X-Routstr-Cost-Msats` | Bearer, X-Cashu | Total msats charged for this request |
+| `X-Routstr-Cost-Usd` | Bearer | USD equivalent of the charge |
+| `X-Routstr-Input-Cost-Msats` | Bearer, X-Cashu | msats attributed to input tokens |
+| `X-Routstr-Output-Cost-Msats` | Bearer, X-Cashu | msats attributed to output tokens |
+
+The client/Tinfoil SDK can read these headers from the HTTP response without
+needing to decrypt the body.
+
+### Setup
+
+```bash
+TINFOIL_API_KEY=your-tinfoil-api-key
+```
+
+The provider is auto-seeded on first startup.
+
+### What still needs verification
+
+- End-to-end test with a real Tinfoil SDK client against a Routstr node with
+  `TINFOIL_API_KEY` set.
+- Streaming requests: usage is delivered as an HTTP trailer. Currently the
+  bearer path finalizes max-cost before streaming begins. Supporting streaming
+  usage would require buffering the response (for X-Cashu) or a deferred
+  finalization (for bearer).
+- Whether Tinfoil's `/v1/responses` endpoint also returns usage metrics
+  headers.

migrations/versions/a2b3c4d5e6f7_add_constraints_to_lightning_invoices.py

@@ -0,0 +1,26 @@
+"""Add balance_limit, balance_limit_reset, validity_date to lightning_invoices
+
+Revision ID: a2b3c4d5e6f7
+Revises: f1a2b3c4d5e6
+Create Date: 2026-06-03 00:00:00.000000
+"""
+
+import sqlalchemy as sa
+from alembic import op
+
+revision = "a2b3c4d5e6f7"
+down_revision = "f1a2b3c4d5e6"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.add_column("lightning_invoices", sa.Column("balance_limit", sa.Integer(), nullable=True))
+    op.add_column("lightning_invoices", sa.Column("balance_limit_reset", sa.String(), nullable=True))
+    op.add_column("lightning_invoices", sa.Column("validity_date", sa.Integer(), nullable=True))
+
+
+def downgrade() -> None:
+    op.drop_column("lightning_invoices", "validity_date")
+    op.drop_column("lightning_invoices", "balance_limit_reset")
+    op.drop_column("lightning_invoices", "balance_limit")

migrations/versions/b5e7c9d1f3a2_add_reserved_at_to_api_keys.py

@@ -0,0 +1,25 @@
+"""add reserved_at to api_keys
+
+Revision ID: b5e7c9d1f3a2
+Revises: a2b3c4d5e6f7
+Create Date: 2026-06-12 00:00:00.000000
+"""
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "b5e7c9d1f3a2"
+down_revision = "a2b3c4d5e6f7"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    # existing keys keep NULL
+    # New reservations populate it via pay_for_request.
+    op.add_column("api_keys", sa.Column("reserved_at", sa.Integer(), nullable=True))
+
+
+def downgrade() -> None:
+    op.drop_column("api_keys", "reserved_at")

migrations/versions/f1a2b3c4d5e6_add_created_at_to_api_keys.py

@@ -0,0 +1,25 @@
+"""add created_at to api_keys
+
+Revision ID: f1a2b3c4d5e6
+Revises: cli_tokens_001
+Create Date: 2026-06-01 00:00:00.000000
+"""
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "f1a2b3c4d5e6"
+down_revision = "cli_tokens_001"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    # Nullable on purpose: existing keys keep NULL (unknown creation time) and
+    # sort last; new keys get populated by the model's default_factory.
+    op.add_column("api_keys", sa.Column("created_at", sa.Integer(), nullable=True))
+
+
+def downgrade() -> None:
+    op.drop_column("api_keys", "created_at")

pyproject.toml

@@ -10,17 +10,18 @@ dependencies = [
     "aiosqlite>=0.20",
     "sqlmodel>=0.0.24",
     "httpx[socks]>=0.25.2",
+    "h11>=0.14",
     "greenlet>=3.2.1",
     "alembic>=1.13",
     "python-json-logger>=2.0.0",
-    "cashu",
-    "secp256k1",
+    "cashu>=0.20",
     "marshmallow>=3.13,<4.0",
     "websockets>=12.0",
     "nostr>=0.0.2",
     "mdurl==0.1.2",
     "pillow>=10",
     "openai>=1.98.0",
+    "litellm>=1.55.0",
 ]
 
 [dependency-groups]
@@ -86,4 +87,3 @@ disallow_untyped_decorators = true
 
 [tool.uv.sources]
 routstr = { workspace = true }
-secp256k1 = { git = "https://github.com/saschanaz/secp256k1-py", branch = "upgrade060" }

routstr/algorithm.py

@@ -195,14 +195,15 @@ def create_model_mappings(
 
             # Add to unique models
             base_id = get_base_model_id(model_to_use.id)
-            if not is_openrouter or base_id not in unique_models:
+            unique_key = model_to_use.forwarded_model_id or base_id
+            if not is_openrouter or unique_key not in unique_models:
                 unique_model = model_to_use.copy(
                     update={
                         "id": base_id,
                         "upstream_provider_id": upstream.provider_type,
                     }
                 )
-                unique_models[base_id] = unique_model
+                unique_models[unique_key] = unique_model
 
             # Get all aliases for this model
             aliases = resolve_model_alias(
@@ -272,18 +273,19 @@ def create_model_mappings(
             continue
 
         base_id = get_base_model_id(model_to_use.id)
+        unique_key = model_to_use.forwarded_model_id or base_id
         is_openrouter = (
             getattr(upstream_for_override, "base_url", "")
             == "https://openrouter.ai/api/v1"
         )
-        if not is_openrouter or base_id not in unique_models:
+        if not is_openrouter or unique_key not in unique_models:
             unique_model = model_to_use.copy(
                 update={
                     "id": base_id,
                     "upstream_provider_id": upstream_for_override.provider_type,
                 }
             )
-            unique_models[base_id] = unique_model
+            unique_models[unique_key] = unique_model
 
         try:
             aliases = resolve_model_alias(
@@ -322,7 +324,26 @@ def create_model_mappings(
     provider_map: dict[str, list["BaseUpstreamProvider"]] = {}
 
     def alias_priority(model: "Model", alias: str) -> int:
-        """Rank how strong the mapping of alias->model is."""
+        """Rank how strong the mapping of alias->model is.
+
+        forwarded_model_id is the most specific identifier (set per-provider
+        instance), so a match there should beat a model_id match. This way,
+        when multiple providers have the same model_id but different
+        forwarded_model_ids, the one whose forwarded_model_id equals the
+        requested alias wins.
+        """
+        if (
+            model.forwarded_model_id
+            and model.forwarded_model_id.lower() == alias
+        ):
+            return 5
+
+        if (
+            model.id
+            and model.id.lower() == alias
+        ):
+            return 4
+
         model_base = get_base_model_id(model.id)
         if model_base == alias:
             return 3

routstr/auth.py

@@ -317,13 +317,13 @@ async def validate_bearer_key(
                 extra={"key_hash": hashed_key[:8] + "..."},
             )
 
-            logger.info(
+            logger.debug(
                 "AUTH: About to call credit_balance",
                 extra={"token_preview": bearer_key[:50]},
             )
             try:
                 msats = await credit_balance(bearer_key, new_key, session)
-                logger.info(
+                logger.debug(
                     "AUTH: credit_balance returned successfully", extra={"msats": msats}
                 )
             except Exception as credit_error:
@@ -341,6 +341,11 @@ async def validate_bearer_key(
                     "Token redemption returned zero or negative amount",
                     extra={"msats": msats, "key_hash": hashed_key[:8] + "..."},
                 )
+                # Defense-in-depth: credit_balance now refuses to commit on a
+                # zero/negative redemption, but if a row was nonetheless
+                # persisted, drop it so we never leave an orphan zero-balance key.
+                await session.delete(new_key)
+                await session.commit()
                 raise Exception("Token redemption failed")
 
             await session.refresh(new_key)
@@ -359,6 +364,7 @@ async def validate_bearer_key(
         except HTTPException:
             raise
         except Exception as e:
+            await session.rollback()
             logger.error(
                 "Cashu token redemption failed",
                 extra={
@@ -529,31 +535,19 @@ async def pay_for_request(
     )
 
     # Charge the base cost for the request atomically to avoid race conditions
+    reserved_at_now = int(time.time())
     stmt = (
         update(ApiKey)
         .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,
+            reserved_at=reserved_at_now,
             total_requests=col(ApiKey.total_requests) + 1,
         )
     )
     result = await session.exec(stmt)  # type: ignore[call-overload]
 
-    # Also increment total_requests and reserved_balance 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,
-                reserved_balance=col(ApiKey.reserved_balance) + cost_per_request,
-            )
-        )
-        await session.exec(child_stmt)  # type: ignore[call-overload]
-
-    await session.commit()
-
     if result.rowcount == 0:
         logger.error(
             "Concurrent request depleted balance",
@@ -565,7 +559,6 @@ async def pay_for_request(
             },
         )
 
-        # Another concurrent request spent the balance first
         raise HTTPException(
             status_code=402,
             detail={
@@ -577,6 +570,44 @@ async def pay_for_request(
             },
         )
 
+    # Also increment total_requests and reserved_balance on the child key if it's different.
+    # The balance_limit guard is enforced atomically here — the Python pre-check above
+    # is a fast-path rejection only and provides no concurrency guarantee.
+    if billing_key.hashed_key != key.hashed_key:
+        child_stmt = (
+            update(ApiKey)
+            .where(col(ApiKey.hashed_key) == key.hashed_key)
+            .where(
+                (col(ApiKey.balance_limit).is_(None))
+                | (
+                    col(ApiKey.total_spent)
+                    + col(ApiKey.reserved_balance)
+                    + cost_per_request
+                    <= col(ApiKey.balance_limit)
+                )
+            )
+            .values(
+                total_requests=col(ApiKey.total_requests) + 1,
+                reserved_balance=col(ApiKey.reserved_balance) + cost_per_request,
+                reserved_at=reserved_at_now,
+            )
+        )
+        child_result = await session.exec(child_stmt)  # type: ignore[call-overload]
+
+        if child_result.rowcount == 0:
+            raise HTTPException(
+                status_code=402,
+                detail={
+                    "error": {
+                        "message": f"Balance limit exceeded: {key.balance_limit} mSats limit. {key.total_spent} already spent ({key.reserved_balance} reserved), {cost_per_request} required for this request.",
+                        "type": "insufficient_quota",
+                        "code": "balance_limit_exceeded",
+                    }
+                },
+            )
+
+    await session.commit()
+
     await session.refresh(billing_key)
     if billing_key.hashed_key != key.hashed_key:
         await session.refresh(key)
@@ -615,12 +646,19 @@ async def revert_pay_for_request(
     False if the reservation was already released (prevents negative reserved_balance)."""
     billing_key = await get_billing_key(key, session)
 
+    # Keep reserved_at while other reservations remain
+    cleared_reserved_at = case(
+        (col(ApiKey.reserved_balance) - cost_per_request > 0, col(ApiKey.reserved_at)),
+        else_=None,
+    )
+
     stmt = (
         update(ApiKey)
         .where(col(ApiKey.hashed_key) == billing_key.hashed_key)
         .where(col(ApiKey.reserved_balance) >= cost_per_request)
         .values(
             reserved_balance=col(ApiKey.reserved_balance) - cost_per_request,
+            reserved_at=cleared_reserved_at,
             total_requests=col(ApiKey.total_requests) - 1,
         )
     )
@@ -636,6 +674,7 @@ async def revert_pay_for_request(
             .values(
                 total_requests=col(ApiKey.total_requests) - 1,
                 reserved_balance=col(ApiKey.reserved_balance) - cost_per_request,
+                reserved_at=cleared_reserved_at,
             )
         )
         await session.exec(child_stmt)  # type: ignore[call-overload]
@@ -670,12 +709,18 @@ async def revert_pay_for_request(
 
 
 async def adjust_payment_for_tokens(
-    key: ApiKey, response_data: dict, session: AsyncSession, deducted_max_cost: int
+    key: ApiKey,
+    response_data: dict,
+    session: AsyncSession,
+    deducted_max_cost: int,
 ) -> dict:
     """
     Adjusts the payment based on token usage in the response.
     This is called after the initial payment and the upstream request is complete.
     Returns cost data to be included in the response.
+
+    The response's usage object is normalized with the default union parser in
+    ``calculate_cost``.
     """
     billing_key = await get_billing_key(key, session)
     model = response_data.get("model", "unknown")
@@ -758,7 +803,7 @@ async def adjust_payment_for_tokens(
                     extra={"error": str(e), "fee_msats": fee_msats},
                 )
 
-    match await calculate_cost(response_data, deducted_max_cost, session):
+    match await calculate_cost(response_data, deducted_max_cost):
         case MaxCostData() as cost:
             logger.debug(
                 "Using max cost data (no token adjustment)",
@@ -976,32 +1021,37 @@ async def adjust_payment_for_tokens(
             # actual cost exceeded discounted reservation (due to tolerance_percentage)
             if cost_difference > 0:
                 # Always release the reservation and charge min(actual_cost, balance).
-                # Using a CASE expression makes this a single atomic UPDATE — no
-                # multi-level fallback needed and balance can never go negative.
+                # CASE expressions keep this atomic and safe even when the
+                # stale-reservation sweeper has already released the reservation.
                 chargeable = case(
                     (col(ApiKey.balance) >= total_cost_msats, total_cost_msats),
                     else_=col(ApiKey.balance),
                 )
+                overrun_safe_reserved = case(
+                    (
+                        col(ApiKey.reserved_balance) >= deducted_max_cost,
+                        col(ApiKey.reserved_balance) - deducted_max_cost,
+                    ),
+                    else_=0,
+                )
 
                 finalize_stmt = (
                     update(ApiKey)
                     .where(col(ApiKey.hashed_key) == billing_key.hashed_key)
-                    .where(col(ApiKey.reserved_balance) >= deducted_max_cost)
                     .values(
-                        reserved_balance=col(ApiKey.reserved_balance) - deducted_max_cost,
+                        reserved_balance=overrun_safe_reserved,
                         balance=col(ApiKey.balance) - chargeable,
                         total_spent=col(ApiKey.total_spent) + chargeable,
                     )
                 )
-                result = await session.exec(finalize_stmt)  # type: ignore[call-overload]
+                await session.exec(finalize_stmt)  # type: ignore[call-overload]
 
                 if billing_key.hashed_key != key.hashed_key:
                     child_stmt = (
                         update(ApiKey)
                         .where(col(ApiKey.hashed_key) == key.hashed_key)
-                        .where(col(ApiKey.reserved_balance) >= deducted_max_cost)
                         .values(
-                            reserved_balance=col(ApiKey.reserved_balance) - deducted_max_cost,
+                            reserved_balance=overrun_safe_reserved,
                             total_spent=col(ApiKey.total_spent) + min(billing_key.balance, total_cost_msats),
                         )
                     )
@@ -1009,51 +1059,38 @@ async def adjust_payment_for_tokens(
 
                 await session.commit()
 
-                if result.rowcount:
-                    await session.refresh(billing_key)
-                    if billing_key.hashed_key != key.hashed_key:
-                        await session.refresh(key)
-                    cost.total_msats = total_cost_msats
-                    logger.info(
-                        "Finalized payment with additional charge",
-                        extra={
-                            "key_hash": key.hashed_key[:8] + "...",
-                            "billing_key_hash": billing_key.hashed_key[:8] + "...",
-                            "charged_amount": total_cost_msats,
-                            "new_balance": billing_key.balance,
-                            "model": model,
-                        },
-                    )
-                    await _accumulate_fee(total_cost_msats)
-                    payments_logger.info(
-                        "FINALIZE",
-                        extra={
-                            "event": "finalize",
-                            "key_hash": key.hashed_key[:8] + "...",
-                            "billing_key_hash": billing_key.hashed_key[:8] + "...",
-                            "model": model,
-                            "cost_reserved": deducted_max_cost,
-                            "cost_charged": total_cost_msats,
-                            "input_tokens": cost.input_tokens,
-                            "output_tokens": cost.output_tokens,
-                            "balance": billing_key.balance,
-                            "reserved_balance": billing_key.reserved_balance,
-                            "total_spent": billing_key.total_spent,
-                            "finalize_type": "overrun",
-                        },
-                    )
-                else:
-                    # Guard fired: reservation was already released by a concurrent
-                    # finalization for this key. Nothing left to do.
-                    logger.warning(
-                        "Finalization skipped - reservation already released",
-                        extra={
-                            "key_hash": key.hashed_key[:8] + "...",
-                            "billing_key_hash": billing_key.hashed_key[:8] + "...",
-                            "attempted_charge": total_cost_msats,
-                            "model": model,
-                        },
-                    )
+                await session.refresh(billing_key)
+                if billing_key.hashed_key != key.hashed_key:
+                    await session.refresh(key)
+                cost.total_msats = total_cost_msats
+                logger.info(
+                    "Finalized payment with additional charge",
+                    extra={
+                        "key_hash": key.hashed_key[:8] + "...",
+                        "billing_key_hash": billing_key.hashed_key[:8] + "...",
+                        "charged_amount": total_cost_msats,
+                        "new_balance": billing_key.balance,
+                        "model": model,
+                    },
+                )
+                await _accumulate_fee(total_cost_msats)
+                payments_logger.info(
+                    "FINALIZE",
+                    extra={
+                        "event": "finalize",
+                        "key_hash": key.hashed_key[:8] + "...",
+                        "billing_key_hash": billing_key.hashed_key[:8] + "...",
+                        "model": model,
+                        "cost_reserved": deducted_max_cost,
+                        "cost_charged": total_cost_msats,
+                        "input_tokens": cost.input_tokens,
+                        "output_tokens": cost.output_tokens,
+                        "balance": billing_key.balance,
+                        "reserved_balance": billing_key.reserved_balance,
+                        "total_spent": billing_key.total_spent,
+                        "finalize_type": "overrun",
+                    },
+                )
             else:
                 # Refund some of the base cost
                 refund = abs(cost_difference)
@@ -1258,3 +1295,53 @@ async def periodic_key_reset() -> None:
             break
         except Exception as e:
             logger.error(f"Error in periodic_key_reset: {e}")
+
+
+async def periodic_dead_key_prune() -> None:
+    """Periodically prune dead API keys. Interval <= 0 disables it.
+
+    See ``prune_dead_api_keys`` for eligibility.
+    """
+    from .core.db import create_session, prune_dead_api_keys
+
+    interval = settings.dead_key_prune_interval_seconds
+    if interval <= 0:
+        logger.info("Dead-key pruning disabled (interval <= 0)")
+        return
+
+    while True:
+        try:
+            await asyncio.sleep(interval)
+        except asyncio.CancelledError:
+            break
+
+        try:
+            async with create_session() as session:
+                await prune_dead_api_keys(
+                    session, settings.dead_key_min_age_seconds
+                )
+        except asyncio.CancelledError:
+            break
+        except Exception as e:
+            logger.error(f"Error in periodic_dead_key_prune: {e}")
+
+
+STALE_RESERVATION_SWEEP_INTERVAL_SECONDS: int = 60
+
+
+async def periodic_stale_reservation_sweep() -> None:
+    """Background task that releases reservations leaked by client disconnects,
+    crashes or abandoned streams.
+    """
+    from .core.db import create_session, release_stale_reservations
+
+    while True:
+        try:
+            async with create_session() as session:
+                await release_stale_reservations(
+                    session, settings.stale_reservation_timeout_seconds
+                )
+        except Exception:
+            logger.exception("Error in periodic_stale_reservation_sweep")
+
+        await asyncio.sleep(STALE_RESERVATION_SWEEP_INTERVAL_SECONDS)

routstr/balance.py

@@ -7,7 +7,7 @@ from typing import Annotated, NoReturn
 from fastapi import APIRouter, Depends, Header, HTTPException
 from fastapi.responses import JSONResponse
 from pydantic import BaseModel
-from sqlmodel import col, select, update
+from sqlmodel import col, or_, select, update
 
 from .auth import get_billing_key, validate_bearer_key
 from .core.db import (
@@ -45,10 +45,9 @@ async def get_balance_info(key: ApiKey, session: AsyncSession) -> dict:
     billing_key = await get_billing_key(key, session)
     info = {
         "api_key": "sk-" + key.hashed_key,
-        "balance": billing_key.balance,
+        "balance": billing_key.total_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,
         "balance_limit": key.balance_limit,
@@ -56,7 +55,9 @@ async def get_balance_info(key: ApiKey, session: AsyncSession) -> dict:
         "validity_date": key.validity_date,
     }
 
-    if not key.parent_key_hash:
+    if key.parent_key_hash:
+        info["parent_key_preview"] = key.parent_key_hash[:8] + "..."
+    else:
         # Fetch child keys if this is a parent key
         statement = select(ApiKey).where(ApiKey.parent_key_hash == key.hashed_key)
         results = await session.exec(statement)
@@ -211,8 +212,20 @@ async def _refund_cache_set(authorization: str, value: dict[str, str]) -> None:
         _refund_cache[key] = (expiry, value)
 
 
+async def _lookup_key_no_create(
+    bearer_value: str, session: AsyncSession
+) -> ApiKey | None:
+    """Look up an existing API key without creating one Used by the refund endpoint"""
+    if bearer_value.startswith("sk-"):
+        return await session.get(ApiKey, bearer_value[3:])
+    if bearer_value.startswith("cashu"):
+        hashed = hashlib.sha256(bearer_value.encode()).hexdigest()
+        return await session.get(ApiKey, hashed)
+    return None
+
+
 async def _restore_balance(
-    session: AsyncSession, hashed_key: str, balance: int, reserved_balance: int
+    session: AsyncSession, hashed_key: str, balance: int, reserved_balance: int, mint_url: str
 ) -> None:
     """Restore balance after a failed refund mint attempt."""
     restore_stmt = (
@@ -227,7 +240,7 @@ async def _restore_balance(
     await session.commit()
     logger.info(
         "refund_wallet_endpoint: balance restored after mint failure",
-        extra={"hashed_key": hashed_key, "restored_balance": balance},
+        extra={"hashed_key": hashed_key, "restored_balance": balance, "mint_url": mint_url},
     )
 
 
@@ -282,7 +295,12 @@ async def refund_wallet_endpoint(
         )
 
     bearer_value: str = authorization[7:]
-    key: ApiKey = await validate_bearer_key(bearer_value, session)
+    key: ApiKey | None = await _lookup_key_no_create(bearer_value, session)
+    if key is None:
+        raise HTTPException(
+            status_code=401,
+            detail="Key not found. Deposit first via /v1/wallet/create before requesting a refund.",
+        )
 
     if key.total_balance <= 0:
         if cached := await _refund_cache_get(bearer_value):
@@ -295,9 +313,36 @@ async def refund_wallet_endpoint(
         )
 
     if key.reserved_balance > 0:
-        raise HTTPException(
-            status_code=400,
-            detail="Cannot refund key. There are ongoing requests for this api key.",
+        # Release the reservation if it is stale
+        cutoff = int(time.time()) - settings.stale_reservation_timeout_seconds
+        stale_release_stmt = (
+            update(ApiKey)
+            .where(col(ApiKey.hashed_key) == key.hashed_key)
+            .where(col(ApiKey.reserved_balance) > 0)
+            .where(
+                or_(
+                    col(ApiKey.reserved_at).is_(None),
+                    col(ApiKey.reserved_at) < cutoff,
+                )
+            )
+            .values(reserved_balance=0, reserved_at=None)
+        )
+        stale_result = await session.exec(stale_release_stmt)  # type: ignore[call-overload]
+        await session.commit()
+
+        if stale_result.rowcount == 0:
+            raise HTTPException(
+                status_code=400,
+                detail="Cannot refund key. There are ongoing requests for this api key.",
+            )
+
+        await session.refresh(key)
+        logger.warning(
+            "refund_wallet_endpoint: released stale reservation before refund",
+            extra={
+                "hashed_key": key.hashed_key,
+                "stale_timeout_seconds": settings.stale_reservation_timeout_seconds,
+            },
         )
 
     remaining_balance_msats: int = key.total_balance
@@ -324,7 +369,7 @@ async def refund_wallet_endpoint(
         .where(col(ApiKey.hashed_key) == key.hashed_key)
         .where(col(ApiKey.balance) == pre_debit_balance)
         .where(col(ApiKey.reserved_balance) == pre_debit_reserved)
-        .values(balance=0, reserved_balance=0)
+        .values(balance=0, reserved_balance=0, reserved_at=None)
     )
     debit_result = await session.exec(debit_stmt)  # type: ignore[call-overload]
     await session.commit()
@@ -337,21 +382,26 @@ async def refund_wallet_endpoint(
         )
 
     # --- MINT: balance is locked at zero, safe to create the refund token ---
+    # Proofs from untrusted mints are swapped to primary_mint on receive.
+    # Use primary_mint unless key.refund_mint_url is an explicitly trusted mint.
+    effective_refund_mint = (
+        key.refund_mint_url
+        if key.refund_mint_url and key.refund_mint_url in settings.cashu_mints
+        else settings.primary_mint
+    )
     try:
         if key.refund_address:
-            from .core.settings import settings as global_settings
-
             await send_to_lnurl(
                 remaining_balance,
                 key.refund_currency or "sat",
-                key.refund_mint_url or global_settings.primary_mint,
+                effective_refund_mint,
                 key.refund_address,
             )
             result = {"recipient": key.refund_address}
         else:
             refund_currency = key.refund_currency or "sat"
             token = await send_token(
-                remaining_balance, refund_currency, key.refund_mint_url
+                remaining_balance, refund_currency, effective_refund_mint
             )
             result = {"token": token}
 
@@ -373,21 +423,32 @@ async def refund_wallet_endpoint(
 
     except HTTPException:
         # Minting failed — restore the debited balance
-        await _restore_balance(session, key.hashed_key, pre_debit_balance, pre_debit_reserved)
+        await _restore_balance(session, key.hashed_key, pre_debit_balance, pre_debit_reserved, key.refund_mint_url or "")
         raise
     except Exception as e:
         # Minting failed — restore the debited balance
-        await _restore_balance(session, key.hashed_key, pre_debit_balance, pre_debit_reserved)
+        await _restore_balance(session, key.hashed_key, pre_debit_balance, pre_debit_reserved, key.refund_mint_url or "")
         error_msg = str(e)
+        logger.error(
+            "refund_wallet_endpoint: mint/send failed",
+            extra={
+                "error": error_msg,
+                "error_type": type(e).__name__,
+                "hashed_key": key.hashed_key,
+                "remaining_balance": remaining_balance,
+                "refund_currency": key.refund_currency,
+                "refund_mint_url": key.refund_mint_url,
+                "has_refund_address": bool(key.refund_address),
+            },
+        )
         if (
             "mint" in error_msg.lower()
             or "connection" in error_msg.lower()
-            or isinstance(e, Exception)
-            and "ConnectError" in str(type(e))
+            or "ConnectError" in str(type(e))
         ):
-            raise HTTPException(status_code=503, detail="Mint service unavailable")
+            raise HTTPException(status_code=503, detail=f"Mint service unavailable: {error_msg}")
         else:
-            raise HTTPException(status_code=500, detail="Refund failed")
+            raise HTTPException(status_code=500, detail=f"Refund failed: {error_msg}")
 
     await _refund_cache_set(bearer_value, result)
 
@@ -502,10 +563,24 @@ async def create_child_key(
             detail=f"Insufficient balance to create {count} child keys. {total_cost} mSats required.",
         )
 
-    # Deduct cost from parent
-    key.balance -= total_cost
-    key.total_spent += total_cost
-    session.add(key)
+    # Deduct cost from parent atomically — guards against concurrent requests
+    # that both pass the balance check above on stale in-memory state.
+    deduct_stmt = (
+        update(ApiKey)
+        .where(col(ApiKey.hashed_key) == key.hashed_key)
+        .where(col(ApiKey.balance) - col(ApiKey.reserved_balance) >= total_cost)
+        .values(
+            balance=col(ApiKey.balance) - total_cost,
+            total_spent=col(ApiKey.total_spent) + total_cost,
+        )
+    )
+    result = await session.exec(deduct_stmt)  # type: ignore[call-overload]
+
+    if result.rowcount == 0:
+        raise HTTPException(
+            status_code=402,
+            detail=f"Insufficient balance to create {count} child keys. {total_cost} mSats required.",
+        )
 
     # Generate new keys
     import secrets
@@ -530,6 +605,7 @@ async def create_child_key(
         new_keys.append("sk-" + new_key_hash)
 
     await session.commit()
+    await session.refresh(key)
 
     response_data = {
         "api_keys": new_keys,
@@ -576,26 +652,6 @@ async def reset_child_key_spent(
     return {"success": True, "message": "Child key balance reset successfully."}
 
 
-@router.get("/cashu-refund/{payment_token_hash}")
-async def get_cashu_refund(
-    payment_token_hash: str,
-    session: AsyncSession = Depends(get_session),
-) -> dict:
-    """Retrieve a stored Cashu refund token by the hash of the original payment token."""
-    result = await session.get(CashuTransaction, payment_token_hash)
-    if result is None:
-        raise HTTPException(status_code=404, detail="Refund not found")
-    if result.swept:
-        raise HTTPException(status_code=410, detail="Refund has been swept")
-    result.collected = True
-    session.add(result)
-    await session.commit()
-    return {
-        "refund_token": result.token,
-        "amount": result.amount,
-        "unit": result.unit,
-    }
-
 
 @router.api_route(
     "/{path:path}",
@@ -609,7 +665,7 @@ async def wallet_catch_all(path: str) -> NoReturn:
     )
 
 
-balance_router.include_router(lightning_router)
+balance_router.include_router(lightning_router, include_in_schema=False)
 balance_router.include_router(router)
 
 deprecated_wallet_router = APIRouter(prefix="/v1/wallet", include_in_schema=False)

routstr/core/admin.py

@@ -5,7 +5,8 @@ from datetime import datetime, timezone
 from pathlib import Path
 
 from fastapi import APIRouter, Depends, HTTPException, Query, Request
-from pydantic import BaseModel
+from pydantic import BaseModel, RootModel
+from pydantic.v1 import ValidationError as PydanticValidationError
 from sqlmodel import select
 
 from ..payment.models import _row_to_model, list_models
@@ -21,6 +22,7 @@ from .db import (
     ApiKey,
     CashuTransaction,
     CliToken,
+    LightningInvoice,
     ModelRow,
     UpstreamProviderRow,
     create_session,
@@ -66,26 +68,89 @@ async def require_admin_api(request: Request) -> None:
 
 
 @admin_router.get("/api/temporary-balances", dependencies=[Depends(require_admin_api)])
-async def get_temporary_balances_api(request: Request) -> list[dict[str, object]]:
+async def get_temporary_balances_api(
+    request: Request,
+    search: str | None = None,
+    limit: int = 50,
+    offset: int = 0,
+) -> dict[str, object]:
+    from sqlalchemy import case
+    from sqlmodel import col, func
+
+    filters = []
+    if search:
+        pattern = f"%{search}%"
+        filters.append(
+            col(ApiKey.hashed_key).like(pattern)
+            | col(ApiKey.refund_address).like(pattern)
+        )
+
     async with create_session() as session:
-        result = await session.exec(select(ApiKey))
+        base = select(ApiKey).where(*filters)
+
+        count_result = await session.exec(
+            select(func.count()).select_from(base.subquery())
+        )
+        total = count_result.one()
+
+        # Aggregate totals across the whole (search-filtered) set, not just the
+        # current page. Balance counts only parent (non-child) keys to avoid
+        # double-counting, since child keys draw from their parent's balance.
+        totals_result = await session.exec(
+            select(
+                func.coalesce(
+                    func.sum(
+                        case(
+                            (col(ApiKey.parent_key_hash).is_(None), ApiKey.balance),
+                            else_=0,
+                        )
+                    ),
+                    0,
+                ),
+                func.coalesce(func.sum(ApiKey.total_spent), 0),
+                func.coalesce(func.sum(ApiKey.total_requests), 0),
+            ).where(*filters)
+        )
+        total_balance, total_spent, total_requests = totals_result.one()
+
+        # Latest created first; keys with no created_at (legacy rows) sort last.
+        # Use an explicit CASE rather than relying on dialect NULL-ordering so
+        # the behaviour is identical on SQLite and Postgres.
+        stmt = (
+            base.order_by(
+                case((col(ApiKey.created_at).is_(None), 1), else_=0),
+                col(ApiKey.created_at).desc(),
+            )
+            .offset(offset)
+            .limit(limit)
+        )
+        result = await session.exec(stmt)
         api_keys = result.all()
 
-    return [
-        {
-            "hashed_key": key.hashed_key,
-            "balance": key.balance,
-            "total_spent": key.total_spent,
-            "total_requests": key.total_requests,
-            "refund_address": key.refund_address,
-            "key_expiry_time": key.key_expiry_time,
-            "parent_key_hash": key.parent_key_hash,
-            "balance_limit": key.balance_limit,
-            "balance_limit_reset": key.balance_limit_reset,
-            "validity_date": key.validity_date,
-        }
-        for key in api_keys
-    ]
+    return {
+        "balances": [
+            {
+                "hashed_key": key.hashed_key,
+                "balance": key.balance,
+                "total_spent": key.total_spent,
+                "total_requests": key.total_requests,
+                "refund_address": key.refund_address,
+                "key_expiry_time": key.key_expiry_time,
+                "parent_key_hash": key.parent_key_hash,
+                "balance_limit": key.balance_limit,
+                "balance_limit_reset": key.balance_limit_reset,
+                "validity_date": key.validity_date,
+                "created_at": key.created_at,
+            }
+            for key in api_keys
+        ],
+        "total": total,
+        "totals": {
+            "total_balance": total_balance,
+            "total_spent": total_spent,
+            "total_requests": total_requests,
+        },
+    }
 
 
 class ApiKeyUpdate(BaseModel):
@@ -142,8 +207,8 @@ async def get_settings(request: Request) -> dict:
     return data
 
 
-class SettingsUpdate(BaseModel):
-    __root__: dict[str, object]
+class SettingsUpdate(RootModel[dict[str, object]]):
+    pass
 
 
 class PasswordUpdate(BaseModel):
@@ -154,14 +219,19 @@ class PasswordUpdate(BaseModel):
 @admin_router.patch("/api/settings", dependencies=[Depends(require_admin_api)])
 async def update_settings(request: Request, update: SettingsUpdate) -> dict:
     # Remove sensitive fields from general settings update
-    settings_data = update.__root__.copy()
+    settings_data = update.root.copy()
     sensitive_fields = ["admin_password", "upstream_api_key", "nsec"]
     for field in sensitive_fields:
         if field in settings_data:
             del settings_data[field]
 
-    async with create_session() as session:
-        new_settings = await SettingsService.update(settings_data, session)
+    try:
+        async with create_session() as session:
+            new_settings = await SettingsService.update(settings_data, session)
+    except PydanticValidationError as e:
+        # Surface validation issues (e.g. non-positive payout amounts)
+        # as a clean 400 instead of a 500.
+        raise HTTPException(status_code=400, detail=e.errors()) from e
     data = new_settings.dict()
     if "upstream_api_key" in data:
         data["upstream_api_key"] = "[REDACTED]" if data["upstream_api_key"] else ""
@@ -1471,6 +1541,52 @@ async def get_transactions_api(
         }
 
 
+@admin_router.get(
+    "/api/lightning-invoices", dependencies=[Depends(require_admin_api)]
+)
+async def get_lightning_invoices_api(
+    status: str | None = None,
+    purpose: str | None = None,
+    search: str | None = None,
+    limit: int = 50,
+    offset: int = 0,
+) -> dict:
+    async with create_session() as session:
+        from sqlmodel import col, func
+
+        base = select(LightningInvoice)
+        if status:
+            base = base.where(LightningInvoice.status == status)
+        if purpose:
+            base = base.where(LightningInvoice.purpose == purpose)
+        if search:
+            pattern = f"%{search}%"
+            base = base.where(
+                (col(LightningInvoice.id).like(pattern))
+                | (col(LightningInvoice.bolt11).like(pattern))
+                | (col(LightningInvoice.payment_hash).like(pattern))
+                | (col(LightningInvoice.api_key_hash).like(pattern))
+            )
+
+        count_result = await session.exec(
+            select(func.count()).select_from(base.subquery())
+        )
+        total = count_result.one()
+
+        stmt = (
+            base.order_by(col(LightningInvoice.created_at).desc())
+            .offset(offset)
+            .limit(limit)
+        )
+        results = await session.exec(stmt)
+        invoices = results.all()
+
+        return {
+            "invoices": [inv.dict() for inv in invoices],
+            "total": total,
+        }
+
+
 @admin_router.post(
     "/api/upstream-providers/{provider_id}/routstr/refund",
     dependencies=[Depends(require_admin_api)],

routstr/core/db.py

@@ -9,9 +9,10 @@ from typing import AsyncGenerator
 from alembic import command
 from alembic.config import Config
 from alembic.util.exc import CommandError
-from sqlalchemy import UniqueConstraint
+from sqlalchemy import UniqueConstraint, delete
 from sqlalchemy.exc import OperationalError
 from sqlalchemy.ext.asyncio.engine import create_async_engine
+from sqlalchemy.orm import aliased
 from sqlmodel import Field, Relationship, SQLModel, col, func, select, update
 from sqlmodel.ext.asyncio.session import AsyncSession
 
@@ -33,6 +34,14 @@ class ApiKey(SQLModel, table=True):  # type: ignore
     reserved_balance: int = Field(
         default=0, description="Reserved balance in millisatoshis (msats)"
     )
+    reserved_at: int | None = Field(
+        default=None,
+        description=(
+            "Unix timestamp of the most recent balance reservation. Used to "
+            "detect and release stale reservations (e.g. after client "
+            "disconnects). NULL when no reservation has been made yet."
+        ),
+    )
     refund_address: str | None = Field(
         default=None,
         description="Lightning address to refund remaining balance after key expires",
@@ -45,6 +54,14 @@ class ApiKey(SQLModel, table=True):  # type: ignore
         default=0, description="Total spent in millisatoshis (msats)"
     )
     total_requests: int = Field(default=0)
+    created_at: int | None = Field(
+        default_factory=lambda: int(time.time()),
+        nullable=True,
+        description=(
+            "Unix timestamp when the key was created. Nullable: keys created "
+            "before this column existed have no value and sort last."
+        ),
+    )
     refund_mint_url: str | None = Field(
         default=None,
         description="URL of the mint used to create the cashu-token",
@@ -79,11 +96,91 @@ class ApiKey(SQLModel, table=True):  # type: ignore
 
 
 async def reset_all_reserved_balances(session: AsyncSession) -> None:
-    logger.info("Resetting all reserved balances to 0")
-    stmt = update(ApiKey).values(reserved_balance=0)
+    stmt = update(ApiKey).values(reserved_balance=0, reserved_at=None)
     await session.exec(stmt)  # type: ignore[call-overload]
     await session.commit()
-    logger.info("Reserved balances reset successfully")
+    logger.info("Reset reserved balances on startup")
+
+
+async def release_stale_reservations(
+    session: AsyncSession, max_age_seconds: int
+) -> int:
+    """Release reservations whose last reserve is older than max_age_seconds.
+    """
+    cutoff = int(time.time()) - max_age_seconds
+    stmt = (
+        update(ApiKey)
+        .where(col(ApiKey.reserved_balance) > 0)
+        .where(col(ApiKey.reserved_at).is_not(None))
+        .where(col(ApiKey.reserved_at) < cutoff)
+        .values(reserved_balance=0, reserved_at=None)
+    )
+    result = await session.exec(stmt)  # type: ignore[call-overload]
+    await session.commit()
+    released = int(result.rowcount or 0)
+    if released:
+        logger.warning(
+            "Released stale balance reservations",
+            extra={"released_keys": released, "max_age_seconds": max_age_seconds},
+        )
+    return released
+
+
+async def prune_dead_api_keys(session: AsyncSession, min_age_seconds: int) -> int:
+    """Delete dead parentless API keys; return the count removed.
+
+    Dead = 0 balance/reservation/spend/requests, older than the grace period,
+    no parent, no children, no pending invoice. Cashu rows are unlinked (not
+    deleted) first to keep the audit trail.
+    """
+    cutoff = int(time.time()) - min_age_seconds
+
+    child = aliased(ApiKey)
+    has_children = (
+        select(child.hashed_key).where(
+            col(child.parent_key_hash) == col(ApiKey.hashed_key)
+        )
+    ).exists()
+    pending_invoice = (
+        select(LightningInvoice.id)
+        .where(col(LightningInvoice.api_key_hash) == col(ApiKey.hashed_key))
+        .where(col(LightningInvoice.status) == "pending")
+    ).exists()
+
+    eligible_hashes = (
+        select(ApiKey.hashed_key)
+        .where(col(ApiKey.balance) == 0)
+        .where(col(ApiKey.reserved_balance) == 0)
+        .where(col(ApiKey.total_spent) == 0)
+        .where(col(ApiKey.total_requests) == 0)
+        .where(col(ApiKey.parent_key_hash).is_(None))
+        .where(
+            (col(ApiKey.created_at).is_(None)) | (col(ApiKey.created_at) < cutoff)
+        )
+        .where(~pending_invoice)
+        .where(~has_children)
+    )
+
+    # Unlink transactions rather than cascade-deleting them, so the financial
+    # audit trail survives. The eligibility predicate is re-evaluated inside both
+    # statements so a key that gained balance mid-run is left untouched.
+    await session.exec(  # type: ignore[call-overload]
+        update(CashuTransaction)
+        .where(col(CashuTransaction.api_key_hashed_key).in_(eligible_hashes))
+        .values(api_key_hashed_key=None)
+    )
+
+    result = await session.exec(  # type: ignore[call-overload]
+        delete(ApiKey).where(col(ApiKey.hashed_key).in_(eligible_hashes))
+    )
+    await session.commit()
+
+    pruned = int(result.rowcount or 0)
+    logger.info(
+        "Pruned dead API keys",
+        extra={"pruned_keys": pruned, "min_age_seconds": min_age_seconds},
+    )
+    return pruned
 
 
 class ModelRow(SQLModel, table=True):  # type: ignore
@@ -133,6 +230,18 @@ class LightningInvoice(SQLModel, table=True):  # type: ignore
     )
     expires_at: int = Field(description="Unix timestamp when invoice expires")
     paid_at: int | None = Field(default=None, description="Unix timestamp when paid")
+    balance_limit: int | None = Field(
+        default=None,
+        description="Max spendable msats for the created key",
+    )
+    balance_limit_reset: str | None = Field(
+        default=None,
+        description="Reset policy for balance limit (daily, weekly, monthly)",
+    )
+    validity_date: int | None = Field(
+        default=None,
+        description="Unix timestamp after which the created key expires",
+    )
 
 
 class CashuTransaction(SQLModel, table=True):  # type: ignore

routstr/core/exceptions.py

@@ -7,11 +7,26 @@ logger = get_logger(__name__)
 
 
 class UpstreamError(Exception):
-    """Exception raised when an upstream provider fails."""
+    """Exception raised when an upstream provider fails.
 
-    def __init__(self, message: str, status_code: int = 502):
+    ``code`` carries a stable, machine-readable classification (e.g.
+    ``UPSTREAM_RATE_LIMIT``) so callers can distinguish failure kinds without
+    string-matching the message. ``details`` holds optional structured,
+    redaction-safe context. Both default to ``None`` for backwards
+    compatibility.
+    """
+
+    def __init__(
+        self,
+        message: str,
+        status_code: int = 502,
+        code: str | None = None,
+        details: dict[str, object] | None = None,
+    ):
         self.message = message
         self.status_code = status_code
+        self.code = code
+        self.details = details
         super().__init__(message)
 
 
@@ -22,16 +37,20 @@ async def http_exception_handler(request: Request, exc: Exception) -> JSONRespon
     # Get status code and detail - works for both FastAPI and Starlette HTTPException
     status_code = getattr(exc, "status_code", 500)
     detail = getattr(exc, "detail", str(exc))
+    path = request.url.path
 
-    logger.warning(
-        "HTTP exception",
-        extra={
-            "request_id": request_id,
-            "status_code": status_code,
-            "detail": detail,
-            "path": request.url.path,
-        },
-    )
+    # 4xx is client behaviour; the uvicorn access log already records it.
+    # Only 5xx warrants a server-side warning/error log here.
+    if status_code >= 500:
+        logger.error(
+            f"HTTP {status_code} on {path}: {detail}",
+            extra={
+                "request_id": request_id,
+                "status_code": status_code,
+                "detail": detail,
+                "path": path,
+            },
+        )
 
     return JSONResponse(
         status_code=status_code,

routstr/core/logging.py

@@ -41,14 +41,25 @@ import logging.config
 import logging.handlers
 import os
 import re
+import sys
 import tomllib
 from datetime import datetime
 from pathlib import Path
 from typing import Any
 
 from pythonjsonlogger import jsonlogger
+from rich.console import Console
 from rich.logging import RichHandler
 
+from .redaction import redact_obj, redact_org_ids
+
+# Only use RichHandler when stdout is a real TTY. In non-TTY contexts
+# (docker logs, pipes, CI) Rich pads every line to width and wraps long
+# records, producing visually-empty trailing whitespace and split records.
+# A plain StreamHandler avoids both problems.
+_stdout_is_tty = sys.stdout.isatty()
+_console = Console(soft_wrap=True) if _stdout_is_tty else None
+
 # Define custom TRACE level
 TRACE_LEVEL = 5
 logging.addLevelName(TRACE_LEVEL, "TRACE")
@@ -171,6 +182,37 @@ class RequestIdFilter(logging.Filter):
         return True
 
 
+# Standard ``LogRecord`` attributes that are never user-supplied ``extra``
+# fields; skipped when redacting structured extras (``msg``/``message`` are
+# handled separately above).
+_NON_EXTRA_RECORD_ATTRS = frozenset(
+    {
+        "name",
+        "msg",
+        "args",
+        "levelname",
+        "levelno",
+        "pathname",
+        "filename",
+        "module",
+        "exc_info",
+        "exc_text",
+        "stack_info",
+        "lineno",
+        "funcName",
+        "created",
+        "msecs",
+        "relativeCreated",
+        "thread",
+        "threadName",
+        "processName",
+        "process",
+        "taskName",
+        "message",
+    }
+)
+
+
 class SecurityFilter(logging.Filter):
     """Filter to remove sensitive information from logs."""
 
@@ -194,6 +236,7 @@ class SecurityFilter(logging.Filter):
         """Filter out sensitive information from log records."""
         try:
             message = record.getMessage()
+            message = redact_org_ids(message)
             standalone_patterns = [
                 r"Bearer\s+([a-zA-Z0-9_\-\.]{10,})",  # Bearer token (must be 10 characters or more to reduce false-positives)
                 r"cashu[A-Z]+([a-zA-Z0-9_\-\.=/+]+)",  # Cashu tokens
@@ -215,6 +258,16 @@ class SecurityFilter(logging.Filter):
             record.msg = message
             record.args = ()
 
+            # Structured `extra={...}` fields are emitted by the JSON formatter
+            # straight from the record dict and never pass through the message
+            # formatting above. Redact organization IDs from any string-valued
+            # extra so they cannot leak via structured logs.
+            for attr, value in list(record.__dict__.items()):
+                if attr in _NON_EXTRA_RECORD_ATTRS:
+                    continue
+                if isinstance(value, (str, dict, list, tuple)):
+                    record.__dict__[attr] = redact_obj(value)
+
         except Exception:
             pass
 
@@ -261,6 +314,26 @@ def setup_logging() -> None:
     if console_enabled:
         handlers.append("console")
 
+    if _stdout_is_tty:
+        console_handler: dict[str, Any] = {
+            "()": RichHandler,
+            "level": log_level,
+            "show_time": False,
+            "show_path": False,
+            "rich_tracebacks": True,
+            "markup": True,
+            "console": _console,
+            "filters": ["request_id_filter", "security_filter"],
+        }
+    else:
+        console_handler = {
+            "class": "logging.StreamHandler",
+            "level": log_level,
+            "formatter": "plain",
+            "stream": "ext://sys.stdout",
+            "filters": ["request_id_filter", "security_filter"],
+        }
+
     LOGGING_CONFIG = {
         "version": 1,
         "disable_existing_loggers": False,
@@ -270,6 +343,10 @@ def setup_logging() -> None:
                 "format": "%(asctime)s %(name)s %(levelname)s %(message)s %(pathname)s %(lineno)d %(version)s %(request_id)s",
                 "datefmt": "%Y-%m-%d %H:%M:%S",
             },
+            "plain": {
+                "format": "%(asctime)s %(levelname)-7s %(name)s %(message)s",
+                "datefmt": "%Y-%m-%d %H:%M:%S",
+            },
         },
         "filters": {
             "version_filter": {"()": VersionFilter},
@@ -277,15 +354,7 @@ def setup_logging() -> None:
             "security_filter": {"()": SecurityFilter},
         },
         "handlers": {
-            "console": {
-                "()": RichHandler,
-                "level": log_level,
-                "show_time": False,
-                "show_path": False,
-                "rich_tracebacks": True,
-                "markup": True,
-                "filters": ["request_id_filter", "security_filter"],
-            },
+            "console": console_handler,
             "file": {
                 "()": DailyRotatingFileHandler,
                 "level": log_level,

routstr/core/main.py

@@ -1,5 +1,4 @@
 import asyncio
-import os
 from contextlib import asynccontextmanager
 from pathlib import Path
 from typing import AsyncGenerator
@@ -9,9 +8,16 @@ from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import FileResponse, RedirectResponse
 from fastapi.staticfiles import StaticFiles
 from starlette.exceptions import HTTPException
+from starlette.responses import Response as StarletteResponse
+from starlette.types import Scope
 
-from ..auth import periodic_key_reset
+from ..auth import (
+    periodic_dead_key_prune,
+    periodic_key_reset,
+    periodic_stale_reservation_sweep,
+)
 from ..balance import balance_router, deprecated_wallet_router
+from ..lightning import lightning_router, periodic_invoice_watcher
 from ..nostr import (
     announce_provider,
     providers_cache_refresher,
@@ -22,24 +28,23 @@ from ..payment.models import models_router, update_sats_pricing
 from ..payment.price import update_prices_periodically
 from ..proxy import initialize_upstreams, proxy_router, refresh_model_maps_periodically
 from ..upstream.auto_topup import periodic_auto_topup
+from ..upstream.deepseek_v4_pricing_shim import register_deepseek_v4_pricing
+from ..upstream.litellm_routing import configure_litellm
 from ..wallet import periodic_payout, periodic_refund_sweep, periodic_routstr_fee_payout
 from .admin import admin_router
 from .db import create_session, init_db, run_migrations
 from .exceptions import general_exception_handler, http_exception_handler
 from .logging import get_logger, setup_logging
 from .middleware import LoggingMiddleware
+from .not_found import _NOT_FOUND_HTML, not_found_catch_all  # noqa: F401
 from .settings import SettingsService
 from .settings import settings as global_settings
+from .version import __version__
 
 # Initialize logging first
 setup_logging()
 logger = get_logger(__name__)
 
-if os.getenv("VERSION_SUFFIX") is not None:
-    __version__ = f"0.4.3-{os.getenv('VERSION_SUFFIX')}"
-else:
-    __version__ = "0.4.3"
-
 
 @asynccontextmanager
 async def lifespan(_: FastAPI) -> AsyncGenerator[None, None]:
@@ -54,11 +59,23 @@ async def lifespan(_: FastAPI) -> AsyncGenerator[None, None]:
     models_refresh_task = None
     model_maps_refresh_task = None
     key_reset_task = None
+    stale_reservation_task = None
+    dead_key_prune_task = None
     auto_topup_task = None
     refund_sweep_task = None
     routstr_fee_task = None
+    invoice_watcher_task = None
 
     try:
+        # Apply litellm-wide settings (drop_params, chat-completions URL,
+        # debug logging) before any upstream provider dispatches a request.
+        configure_litellm()
+
+        # TEMPORARY: backfill DeepSeek V4 pricing missing from litellm's cost
+        # map (BerriAI/litellm#30430). Remove this call and
+        # deepseek_v4_pricing_shim.py once litellm ships these models.
+        register_deepseek_v4_pricing()
+
         # Run database migrations on startup
         run_migrations()
 
@@ -103,8 +120,11 @@ async def lifespan(_: FastAPI) -> AsyncGenerator[None, None]:
         btc_price_task = asyncio.create_task(update_prices_periodically())
         pricing_task = asyncio.create_task(update_sats_pricing())
         if global_settings.models_refresh_interval_seconds > 0:
+            # Pass the accessor (not its current value) so the loop sees providers
+            # added/changed via reinitialize_upstreams() instead of staying pinned
+            # to the startup snapshot.
             models_refresh_task = asyncio.create_task(
-                refresh_upstreams_models_periodically(get_upstreams())
+                refresh_upstreams_models_periodically(get_upstreams)
             )
         model_maps_refresh_task = asyncio.create_task(refresh_model_maps_periodically())
         payout_task = asyncio.create_task(periodic_payout())
@@ -114,9 +134,14 @@ async def lifespan(_: FastAPI) -> AsyncGenerator[None, None]:
         if global_settings.providers_refresh_interval_seconds > 0:
             providers_task = asyncio.create_task(providers_cache_refresher())
         key_reset_task = asyncio.create_task(periodic_key_reset())
+        stale_reservation_task = asyncio.create_task(
+            periodic_stale_reservation_sweep()
+        )
+        dead_key_prune_task = asyncio.create_task(periodic_dead_key_prune())
         auto_topup_task = asyncio.create_task(periodic_auto_topup())
         refund_sweep_task = asyncio.create_task(periodic_refund_sweep())
         routstr_fee_task = asyncio.create_task(periodic_routstr_fee_payout())
+        invoice_watcher_task = asyncio.create_task(periodic_invoice_watcher())
 
         yield
 
@@ -150,12 +175,18 @@ async def lifespan(_: FastAPI) -> AsyncGenerator[None, None]:
             model_maps_refresh_task.cancel()
         if key_reset_task is not None:
             key_reset_task.cancel()
+        if stale_reservation_task is not None:
+            stale_reservation_task.cancel()
+        if dead_key_prune_task is not None:
+            dead_key_prune_task.cancel()
         if auto_topup_task is not None:
             auto_topup_task.cancel()
         if refund_sweep_task is not None:
             refund_sweep_task.cancel()
         if routstr_fee_task is not None:
             routstr_fee_task.cancel()
+        if invoice_watcher_task is not None:
+            invoice_watcher_task.cancel()
 
         try:
             tasks_to_wait = []
@@ -177,12 +208,18 @@ async def lifespan(_: FastAPI) -> AsyncGenerator[None, None]:
                 tasks_to_wait.append(model_maps_refresh_task)
             if key_reset_task is not None:
                 tasks_to_wait.append(key_reset_task)
+            if stale_reservation_task is not None:
+                tasks_to_wait.append(stale_reservation_task)
+            if dead_key_prune_task is not None:
+                tasks_to_wait.append(dead_key_prune_task)
             if auto_topup_task is not None:
                 tasks_to_wait.append(auto_topup_task)
             if refund_sweep_task is not None:
                 tasks_to_wait.append(refund_sweep_task)
             if routstr_fee_task is not None:
                 tasks_to_wait.append(routstr_fee_task)
+            if invoice_watcher_task is not None:
+                tasks_to_wait.append(invoice_watcher_task)
 
             if tasks_to_wait:
                 await asyncio.gather(*tasks_to_wait, return_exceptions=True)
@@ -194,6 +231,23 @@ async def lifespan(_: FastAPI) -> AsyncGenerator[None, None]:
             )
 
 
+class _ImmutableStaticFiles(StaticFiles):
+    """Static files with long Cache-Control for content-hashed Next.js assets.
+
+    Files under `/_next/static/` are emitted with content hashes in their
+    filenames and never mutate, so we serve them with a one-year immutable
+    cache header so browsers and CDNs stop revalidating on every reload.
+    """
+
+    async def get_response(self, path: str, scope: Scope) -> StarletteResponse:
+        response = await super().get_response(path, scope)
+        if response.status_code == 200:
+            response.headers["Cache-Control"] = (
+                "public, max-age=31536000, immutable"
+            )
+        return response
+
+
 app = FastAPI(version=__version__, lifespan=lifespan)
 
 
@@ -203,7 +257,20 @@ app.add_middleware(
     allow_credentials=True,
     allow_methods=["*"],
     allow_headers=["*"],
-    expose_headers=["x-routstr-request-id", "x-cashu"],
+    expose_headers=[
+        "x-routstr-request-id",
+        "x-cashu",
+        "x-routstr-cost-msats",
+        "x-routstr-cost-usd",
+        "x-routstr-input-cost-msats",
+        "x-routstr-output-cost-msats",
+        # EHBP (Tinfoil) protocol headers must be exposed so browser clients
+        # can detect and decrypt encrypted responses. Without these, the
+        # browser hides them via CORS and the SDK treats the response as a
+        # plaintext proxy error, returning raw ciphertext.
+        "Ehbp-Response-Nonce",
+        "Ehbp-Encapsulated-Key",
+    ],
 )
 
 # Add logging middleware
@@ -240,7 +307,7 @@ if UI_DIST_PATH.exists() and UI_DIST_PATH.is_dir():
 
     app.mount(
         "/_next",
-        StaticFiles(directory=UI_DIST_PATH / "_next", check_dir=True),
+        _ImmutableStaticFiles(directory=UI_DIST_PATH / "_next", check_dir=True),
         name="next-static",
     )
 
@@ -248,100 +315,70 @@ if UI_DIST_PATH.exists() and UI_DIST_PATH.is_dir():
     async def serve_root_ui() -> FileResponse:
         return FileResponse(UI_DIST_PATH / "index.html")
 
-    # Add explicit route for /index.txt to redirect to /
+    # Serve the App Router RSC payload for the home page.
     @app.get("/index.txt", include_in_schema=False)
-    async def redirect_index_txt() -> RedirectResponse:
-        return RedirectResponse("/")
+    async def serve_root_rsc() -> FileResponse:
+        return FileResponse(
+            UI_DIST_PATH / "index.txt", media_type="text/x-component"
+        )
+
+    # Next.js is built with `trailingSlash: true`, so all UI page URLs end
+    # with a slash (e.g. `/login/`). The proxy router catches `/{path:path}`
+    # before FastAPI's `redirect_slashes` logic can normalize the URL, so we
+    # must register both the with-slash and without-slash variants here.
+    UI_PAGES = (
+        "dashboard",
+        "login",
+        "model",
+        "providers",
+        "settings",
+        "transactions",
+        "balances",
+        "logs",
+        "usage",
+        "unauthorized",
+    )
+
+    def _register_ui_page(name: str) -> None:
+        page_dir = UI_DIST_PATH / name
+        index_html = page_dir / "index.html"
+        index_txt = page_dir / "index.txt"
+
+        async def serve_page() -> FileResponse:
+            return FileResponse(index_html)
+
+        async def serve_page_rsc() -> FileResponse:
+            return FileResponse(index_txt, media_type="text/x-component")
+
+        app.add_api_route(
+            f"/{name}",
+            serve_page,
+            methods=["GET"],
+            include_in_schema=False,
+            name=f"serve_{name}_ui",
+        )
+        app.add_api_route(
+            f"/{name}/",
+            serve_page,
+            methods=["GET"],
+            include_in_schema=False,
+            name=f"serve_{name}_ui_slash",
+        )
+        app.add_api_route(
+            f"/{name}/index.txt",
+            serve_page_rsc,
+            methods=["GET"],
+            include_in_schema=False,
+            name=f"serve_{name}_rsc",
+        )
+
+    for _page in UI_PAGES:
+        _register_ui_page(_page)
 
     @app.get("/admin")
     async def admin_redirect() -> FileResponse:
         return FileResponse(UI_DIST_PATH / "index.html")
 
-    @app.get("/dashboard", include_in_schema=False)
-    async def serve_dashboard_ui() -> FileResponse:
-        return FileResponse(UI_DIST_PATH / "index.html")
-
-    @app.get("/login", include_in_schema=False)
-    async def serve_login_ui() -> FileResponse:
-        return FileResponse(UI_DIST_PATH / "login" / "index.html")
-
-    # Add explicit route for /login/index.txt to redirect to /login
-    @app.get("/login/index.txt", include_in_schema=False)
-    async def redirect_login_index_txt() -> RedirectResponse:
-        return RedirectResponse("/login")
-
-    @app.get("/model", include_in_schema=False)
-    async def serve_models_ui() -> FileResponse:
-        return FileResponse(UI_DIST_PATH / "model" / "index.html")
-
-    # Add explicit route for /model/index.txt to redirect to /model
-    @app.get("/model/index.txt", include_in_schema=False)
-    async def redirect_model_index_txt() -> RedirectResponse:
-        return RedirectResponse("/model")
-
-    @app.get("/providers", include_in_schema=False)
-    async def serve_providers_ui() -> FileResponse:
-        return FileResponse(UI_DIST_PATH / "providers" / "index.html")
-
-    # Add explicit route for /providers/index.txt to redirect to /providers
-    @app.get("/providers/index.txt", include_in_schema=False)
-    async def redirect_providers_index_txt() -> RedirectResponse:
-        return RedirectResponse("/providers")
-
-    @app.get("/settings", include_in_schema=False)
-    async def serve_settings_ui() -> FileResponse:
-        return FileResponse(UI_DIST_PATH / "settings" / "index.html")
-
-    # Add explicit route for /settings/index.txt to redirect to /settings
-    @app.get("/settings/index.txt", include_in_schema=False)
-    async def redirect_settings_index_txt() -> RedirectResponse:
-        return RedirectResponse("/settings")
-
-    @app.get("/transactions", include_in_schema=False)
-    async def serve_transactions_ui() -> FileResponse:
-        return FileResponse(UI_DIST_PATH / "transactions" / "index.html")
-
-    # Add explicit route for /transactions/index.txt to redirect to /transactions
-    @app.get("/transactions/index.txt", include_in_schema=False)
-    async def redirect_transactions_index_txt() -> RedirectResponse:
-        return RedirectResponse("/transactions")
-
-    @app.get("/balances", include_in_schema=False)
-    async def serve_balances_ui() -> FileResponse:
-        return FileResponse(UI_DIST_PATH / "balances" / "index.html")
-
-    # Add explicit route for /balances/index.txt to redirect to /balances
-    @app.get("/balances/index.txt", include_in_schema=False)
-    async def redirect_balances_index_txt() -> RedirectResponse:
-        return RedirectResponse("/balances")
-
-    @app.get("/logs", include_in_schema=False)
-    async def serve_logs_ui() -> FileResponse:
-        return FileResponse(UI_DIST_PATH / "logs" / "index.html")
-
-    # Add explicit route for /logs/index.txt to redirect to /logs
-    @app.get("/logs/index.txt", include_in_schema=False)
-    async def redirect_logs_index_txt() -> RedirectResponse:
-        return RedirectResponse("/logs")
-
-    @app.get("/usage", include_in_schema=False)
-    async def serve_usage_ui() -> FileResponse:
-        return FileResponse(UI_DIST_PATH / "usage" / "index.html")
-
-    # Add explicit route for /usage/index.txt to redirect to /usage
-    @app.get("/usage/index.txt", include_in_schema=False)
-    async def redirect_usage_index_txt() -> RedirectResponse:
-        return RedirectResponse("/usage")
-
-    @app.get("/unauthorized", include_in_schema=False)
-    async def serve_unauthorized_ui() -> FileResponse:
-        return FileResponse(UI_DIST_PATH / "unauthorized" / "index.html")
-
-    # Add explicit route for /unauthorized/index.txt to redirect to /unauthorized
-    @app.get("/unauthorized/index.txt", include_in_schema=False)
-    async def redirect_unauthorized_index_txt() -> RedirectResponse:
-        return RedirectResponse("/unauthorized")
-
     @app.get("/favicon.ico", include_in_schema=False)
     async def serve_favicon() -> FileResponse:
         icon_path = UI_DIST_PATH / "icon.ico"
@@ -353,12 +390,12 @@ if UI_DIST_PATH.exists() and UI_DIST_PATH.is_dir():
     async def serve_icon() -> FileResponse:
         return FileResponse(UI_DIST_PATH / "icon.ico")
 
-    app.mount(
-        "/static", StaticFiles(directory=UI_DIST_PATH, check_dir=True), name="ui-static"
-    )
 else:
     logger.warning(
-        f"UI dist directory not found at {UI_DIST_PATH}, skipping static file serving"
+        "UI dist directory not found at %s; serving API only. Run `make ui-build` "
+        "to build the static UI served from here, or `make ui-dev` for the Next.js "
+        "dev server with hot reload on :3000 (it targets this backend on :8000).",
+        UI_DIST_PATH,
     )
 
     @app.get("/", include_in_schema=False)
@@ -375,6 +412,7 @@ else:
 app.include_router(models_router)
 app.include_router(admin_router)
 app.include_router(balance_router)
+app.include_router(lightning_router)
 app.include_router(deprecated_wallet_router)
 app.include_router(providers_router)
 app.include_router(proxy_router)

routstr/core/middleware.py

@@ -14,8 +14,54 @@ logger = get_logger(__name__)
 request_id_context: ContextVar[str | None] = ContextVar("request_id")
 
 
+# Methods that are never logged: HEAD requests are health probes from
+# monitoring/load balancers, OPTIONS are CORS preflights — both are framework
+# chatter, not user-meaningful events.
+_SKIP_LOG_METHODS: frozenset[str] = frozenset({"HEAD", "OPTIONS"})
+
+# Path prefixes to skip. Includes Next.js static chunks and the admin
+# dashboard's internal polling API (/admin/api/*) which the UI hits on a timer
+# to refresh balances, logs, providers, etc. — high volume, low diagnostic
+# value. Mutating admin actions are recorded separately in the audit log.
+_SKIP_LOG_PREFIXES: tuple[str, ...] = (
+    "/_next/",
+    "/admin/api/",
+)
+
+# Exact paths to skip. RSC payload prefetches (`*/index.txt`) fire automatically
+# as the user hovers near `<Link>`s, and `/v1/wallet/info` is polled by the UI.
+_SKIP_LOG_EXACT: frozenset[str] = frozenset(
+    {
+        "/favicon.ico",
+        "/icon.ico",
+        "/v1/wallet/info",
+        "/index.txt",
+        "/login/index.txt",
+        "/model/index.txt",
+        "/providers/index.txt",
+        "/settings/index.txt",
+        "/transactions/index.txt",
+        "/balances/index.txt",
+        "/logs/index.txt",
+        "/usage/index.txt",
+        "/unauthorized/index.txt",
+    }
+)
+
+
+def _should_log(method: str, path: str) -> bool:
+    if method in _SKIP_LOG_METHODS:
+        return False
+    if path in _SKIP_LOG_EXACT:
+        return False
+    return not any(path.startswith(prefix) for prefix in _SKIP_LOG_PREFIXES)
+
+
 class LoggingMiddleware(BaseHTTPMiddleware):
-    """Middleware to log detailed request and response information."""
+    """Middleware to log proxy interactions and page navigation.
+
+    Skips logging for static assets and Next.js chunks to avoid noise.
+    """
 
     async def dispatch(self, request: Request, call_next: Callable) -> Response:
         # Generate request ID
@@ -25,56 +71,20 @@ class LoggingMiddleware(BaseHTTPMiddleware):
         # Set request ID in context for logging
         token = request_id_context.set(request_id)
 
+        path = request.url.path
+        should_log = _should_log(request.method, path)
+
         # Start timing
         start_time = time.time()
 
-        # Log request details
-        request_body = None
-        if request.method in ["POST", "PUT", "PATCH"]:
-            try:
-                # Only read body for non-streaming requests
-                if hasattr(request, "_body"):
-                    request_body = await request.body()
-            except Exception:
-                pass
-
-        # Log incoming request
-        logger.info(
-            "Incoming request",
-            extra={
-                "request_id": request_id,
-                "method": request.method,
-                "path": request.url.path,
-                "query_params": dict(request.query_params),
-                "headers": {
-                    k: v
-                    for k, v in request.headers.items()
-                    if k.lower()
-                    not in [
-                        "authorization",
-                        "x-cashu",
-                        "cookie",
-                        "cf-connecting-ip",
-                        "cf-ipcountry",
-                        "x-forwarded-for",
-                        "x-real-ip",
-                    ]
-                },
-                "body_size": len(request_body) if request_body else 0,
-            },
-        )
-
-        # Log at TRACE level for full body (security filter will redact sensitive data)
-        if request_body and hasattr(logger, "exception"):
-            logger.exception(
-                "Request body",
+        if should_log:
+            logger.info(
+                "Incoming request",
                 extra={
                     "request_id": request_id,
                     "method": request.method,
-                    "path": request.url.path,
-                    "body": request_body.decode("utf-8", errors="ignore")[
-                        :1000
-                    ],  # Limit size
+                    "path": path,
+                    "query_params": dict(request.query_params),
                 },
             )
 
@@ -82,36 +92,32 @@ class LoggingMiddleware(BaseHTTPMiddleware):
         try:
             response = await call_next(request)
 
-            # Calculate duration
-            duration = time.time() - start_time
-
-            # Log response
-            logger.info(
-                "Request completed",
-                extra={
-                    "request_id": request_id,
-                    "method": request.method,
-                    "path": request.url.path,
-                    "status_code": response.status_code,
-                    "duration_ms": round(duration * 1000, 2),
-                },
-            )
+            if should_log:
+                duration = time.time() - start_time
+                logger.info(
+                    "Request completed",
+                    extra={
+                        "request_id": request_id,
+                        "method": request.method,
+                        "path": path,
+                        "status_code": response.status_code,
+                        "duration_ms": round(duration * 1000, 2),
+                    },
+                )
             if hasattr(response, "headers"):
                 response.headers["x-routstr-request-id"] = request_id
 
             return response
 
         except Exception as e:
-            # Calculate duration
+            # Always log failures, even for skipped paths, so we don't lose errors.
             duration = time.time() - start_time
-
-            # Log error
             logger.error(
                 "Request failed",
                 extra={
                     "request_id": request_id,
                     "method": request.method,
-                    "path": request.url.path,
+                    "path": path,
                     "duration_ms": round(duration * 1000, 2),
                     "error": str(e),
                     "error_type": type(e).__name__,

routstr/core/not_found.py

@@ -0,0 +1,55 @@
+"""Shared 404 handler used by the proxy catch-all and tests."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from fastapi import Request
+from fastapi.responses import HTMLResponse, JSONResponse, Response
+
+_NOT_FOUND_HTML_FILE = Path(__file__).parent.parent.parent / "ui_out" / "404.html"
+
+
+def _read_not_found_html() -> str | None:
+    try:
+        return _NOT_FOUND_HTML_FILE.read_text(encoding="utf-8")
+    except OSError:
+        return None
+
+
+_NOT_FOUND_HTML: str | None = _read_not_found_html()
+
+
+def build_not_found_response(request: Request, path: str) -> Response:
+    """Return a 404 response.
+
+    HTML 404 page only for GET requests from browsers (Accept: text/html).
+    All POST requests and API clients receive a JSON 404.
+    """
+    accept = request.headers.get("accept", "").lower()
+    prefers_html = (
+        request.method == "GET"
+        and "text/html" in accept
+        and "application/json" not in accept
+    )
+    request_id = getattr(request.state, "request_id", "unknown")
+
+    if prefers_html and _NOT_FOUND_HTML is not None:
+        return HTMLResponse(content=_NOT_FOUND_HTML, status_code=404)
+
+    return JSONResponse(
+        status_code=404,
+        content={
+            "error": {
+                "message": f"Path '/{path}' not found",
+                "type": "not_found",
+                "code": 404,
+            },
+            "request_id": request_id,
+        },
+    )
+
+
+async def not_found_catch_all(request: Request, path: str) -> Response:
+    """ASGI handler form of :func:`build_not_found_response`."""
+    return build_not_found_response(request, path)

routstr/core/redaction.py

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

routstr/core/settings.py

@@ -41,6 +41,15 @@ class Settings(BaseSettings):
     primary_mint: str = Field(default="", env="PRIMARY_MINT_URL")
     primary_mint_unit: str = Field(default="sat", env="PRIMARY_MINT_UNIT")
 
+    # Lightning payout configuration
+    # Minimum available balance (in satoshis) before profit is paid out over
+    # Lightning
+    min_payout_sat: int = Field(default=210, gt=0, env="MIN_PAYOUT_SAT")
+    # Interval (seconds) between periodic payout attempts. Must be positive.
+    payout_interval_seconds: int = Field(
+        default=900, gt=0, env="PAYOUT_INTERVAL_SECONDS"
+    )
+
     # Pricing
     # Default behavior: derive pricing from MODELS
     # If fixed_pricing is True -> use fixed_cost_per_request and ignore tokens
@@ -58,6 +67,20 @@ class Settings(BaseSettings):
     reset_reserved_balance_on_startup: bool = Field(
         default=True, env="RESET_RESERVED_BALANCE_ON_STARTUP"
     )  # deactivate in horizontal scaling setups
+    # Reservations older than this are considered leaked (client disconnect,
+    # crash, abandoned stream) and released by the background sweeper and the
+    # refund endpoint.
+    stale_reservation_timeout_seconds: int = Field(
+        default=300, env="STALE_RESERVATION_TIMEOUT_SECONDS"
+    )
+    # Background prune of dead (zero balance, never used) API keys.
+    # Interval 0 disables it; min-age is a grace period (default 1 week).
+    dead_key_prune_interval_seconds: int = Field(
+        default=3600, env="DEAD_KEY_PRUNE_INTERVAL_SECONDS"
+    )
+    dead_key_min_age_seconds: int = Field(
+        default=604_800, env="DEAD_KEY_MIN_AGE_SECONDS"
+    )
 
     # Network
     cors_origins: list[str] = Field(default_factory=lambda: ["*"], env="CORS_ORIGINS")

routstr/core/version.py

@@ -0,0 +1,81 @@
+"""Application version resolution.
+
+Priority order:
+
+1. ``VERSION_SUFFIX`` env var (manual override; preserves prior behaviour).
+2. Bare base version when HEAD is on the matching release tag (detected via
+   ``GIT_TAG`` env or ``git describe --tags --exact-match HEAD``).
+3. ``GIT_COMMIT`` env var (build-time injection) -> ``<base>+g<sha>``.
+4. Local ``.git`` lookup (source checkouts) -> ``<base>+g<sha>``.
+5. Fallback: bare base version.
+
+The ``+g<sha>`` form is PEP 440 local-version syntax so the result remains a
+valid package version.
+"""
+
+from __future__ import annotations
+
+import os
+import subprocess
+from functools import lru_cache
+from pathlib import Path
+
+BASE_VERSION = "0.4.3"
+
+_REPO_ROOT = Path(__file__).resolve().parents[2]
+_GIT_TIMEOUT_SECONDS = 2.0
+
+
+def _run_git(*args: str) -> str | None:
+    try:
+        result = subprocess.run(  # noqa: S603 - fixed argv, no shell
+            ["git", *args],
+            cwd=_REPO_ROOT,
+            check=False,
+            capture_output=True,
+            text=True,
+            timeout=_GIT_TIMEOUT_SECONDS,
+        )
+    except (FileNotFoundError, subprocess.SubprocessError, OSError):
+        return None
+    if result.returncode != 0:
+        return None
+    return result.stdout.strip() or None
+
+
+def _git_short_sha() -> str | None:
+    sha = os.getenv("GIT_COMMIT", "").strip()
+    if sha:
+        return sha[:7]
+    return _run_git("rev-parse", "--short=7", "HEAD")
+
+
+def _on_tagged_release() -> bool:
+    tag = os.getenv("GIT_TAG", "").strip()
+    if tag:
+        return tag.lstrip("v") == BASE_VERSION
+    described = _run_git("describe", "--tags", "--exact-match", "HEAD")
+    if described and described.lstrip("v") == BASE_VERSION:
+        return True
+    return False
+
+
+@lru_cache(maxsize=1)
+def get_version() -> str:
+    suffix = os.getenv("VERSION_SUFFIX")
+    if suffix is not None:
+        return f"{BASE_VERSION}-{suffix}"
+
+    if _on_tagged_release():
+        return BASE_VERSION
+
+    sha = _git_short_sha()
+    if not sha:
+        return BASE_VERSION
+
+    return f"{BASE_VERSION}+g{sha}"
+
+
+__version__ = get_version()
+
+__all__ = ["BASE_VERSION", "__version__", "get_version"]

routstr/lightning.py

@@ -1,13 +1,14 @@
+import asyncio
 import hashlib
 import secrets
 import time
 
-from fastapi import APIRouter, Depends, HTTPException
+from fastapi import APIRouter, Depends, Header, HTTPException
 from pydantic import BaseModel, Field
-from sqlmodel import select
+from sqlmodel import col, select
 from sqlmodel.ext.asyncio.session import AsyncSession
 
-from .core.db import ApiKey, LightningInvoice, get_session
+from .core.db import ApiKey, LightningInvoice, create_session, get_session
 from .core.logging import get_logger
 from .core.settings import settings
 from .wallet import get_wallet
@@ -19,15 +20,29 @@ lightning_router = APIRouter(prefix="/lightning")
 
 class InvoiceCreateRequest(BaseModel):
     amount_sats: int = Field(gt=0, le=1_000_000, description="Amount in satoshis")
-    purpose: str = Field(description="create or topup", pattern="^(create|topup)$")
+    purpose: str = Field(
+        default="create",
+        description="create or topup",
+        pattern="^(create|topup)$",
+    )
     api_key: str | None = Field(
-        default=None, description="Required for topup operations"
+        default=None,
+        description="Deprecated: legacy field for topup. Prefer Authorization header.",
     )
     balance_limit: int | None = Field(default=None)
     balance_limit_reset: str | None = Field(default=None)
     validity_date: int | None = Field(default=None)
 
 
+def _extract_bearer_api_key(authorization: str | None) -> str | None:
+    if not authorization:
+        return None
+    token = authorization.strip()
+    if token.lower().startswith("bearer "):
+        token = token[7:].strip()
+    return token or None
+
+
 class InvoiceCreateResponse(BaseModel):
     invoice_id: str
     bolt11: str
@@ -64,18 +79,21 @@ def generate_invoice_id() -> str:
 @lightning_router.post("/invoice", response_model=InvoiceCreateResponse)
 async def create_invoice(
     request: InvoiceCreateRequest,
+    authorization: str | None = Header(default=None),
     session: AsyncSession = Depends(get_session),
 ) -> InvoiceCreateResponse:
-    if request.purpose == "topup" and not request.api_key:
-        raise HTTPException(
-            status_code=400, detail="api_key is required for topup operations"
-        )
+    api_key_token = _extract_bearer_api_key(authorization) or request.api_key
 
-    if request.purpose == "topup" and request.api_key:
-        if not request.api_key.startswith("sk-"):
+    if request.purpose == "topup":
+        if not api_key_token:
+            raise HTTPException(
+                status_code=401,
+                detail="Authorization bearer api key is required for topup",
+            )
+        if not api_key_token.startswith("sk-"):
             raise HTTPException(status_code=400, detail="Invalid API key format")
 
-        api_key = await session.get(ApiKey, request.api_key[3:])
+        api_key = await session.get(ApiKey, api_key_token[3:])
         if not api_key:
             raise HTTPException(status_code=404, detail="API key not found")
 
@@ -95,7 +113,7 @@ async def create_invoice(
             description=description,
             payment_hash=payment_hash,
             status="pending",
-            api_key_hash=request.api_key[3:] if request.api_key else None,
+            api_key_hash=api_key_token[3:] if api_key_token else None,
             purpose=request.purpose,
             balance_limit=request.balance_limit,
             balance_limit_reset=request.balance_limit_reset,
@@ -142,13 +160,13 @@ async def get_invoice_status(
     if not invoice:
         raise HTTPException(status_code=404, detail="Invoice not found")
 
+    if invoice.status == "pending":
+        await check_invoice_payment(invoice, session)
+
     if invoice.status == "pending" and int(time.time()) > invoice.expires_at:
         invoice.status = "expired"
         await session.commit()
 
-    if invoice.status == "pending":
-        await check_invoice_payment(invoice, session)
-
     api_key = None
     if invoice.status == "paid" and invoice.purpose == "create":
         if invoice.api_key_hash:
@@ -251,6 +269,9 @@ async def create_api_key_from_invoice(
         balance=invoice.amount_sats * 1000,  # Convert to msats
         refund_currency="sat",
         refund_mint_url=settings.primary_mint,
+        balance_limit=invoice.balance_limit,
+        balance_limit_reset=invoice.balance_limit_reset,
+        validity_date=invoice.validity_date,
     )
 
     session.add(api_key)
@@ -274,3 +295,41 @@ async def topup_api_key_from_invoice(
 
     api_key.balance += invoice.amount_sats * 1000  # Convert to msats
     await session.flush()
+
+
+INVOICE_WATCH_INTERVAL_SECONDS = 5
+INVOICE_WATCH_BATCH_LIMIT = 100
+
+
+async def periodic_invoice_watcher() -> None:
+    """Background task: detect paid Lightning invoices and credit balances.
+
+    Removes the need for clients to poll the status endpoint after paying.
+    """
+    while True:
+        try:
+            async with create_session() as session:
+                now = int(time.time())
+                result = await session.exec(
+                    select(LightningInvoice)
+                    .where(
+                        LightningInvoice.status == "pending",
+                        col(LightningInvoice.expires_at) > now,
+                    )
+                    .limit(INVOICE_WATCH_BATCH_LIMIT)
+                )
+                pending = result.all()
+                for invoice in pending:
+                    try:
+                        await check_invoice_payment(invoice, session)
+                    except Exception as e:
+                        logger.error(
+                            "Invoice watcher failed for invoice",
+                            extra={"invoice_id": invoice.id, "error": str(e)},
+                        )
+        except asyncio.CancelledError:
+            raise
+        except Exception as e:
+            logger.error(f"Invoice watcher loop error: {e}")
+
+        await asyncio.sleep(INVOICE_WATCH_INTERVAL_SECONDS)

routstr/nostr/listing.py

@@ -26,7 +26,7 @@ logger = get_logger(__name__)
 
 def get_app_version() -> str | None:
     try:
-        from ..core.main import __version__ as imported_version
+        from ..core.version import __version__ as imported_version
 
         return imported_version
     except Exception:

routstr/payment/cost_calculation.py

@@ -3,9 +3,17 @@ import math
 from pydantic.v1 import BaseModel
 
 from ..core import get_logger
-from ..core.db import AsyncSession
 from ..core.settings import settings
 from .price import sats_usd_price
+from .usage import normalize_usage, parse_token_count
+
+__all__ = [
+    "CostData",
+    "CostDataError",
+    "MaxCostData",
+    "calculate_cost",
+    "parse_token_count",
+]
 
 logger = get_logger(__name__)
 
@@ -18,6 +26,10 @@ class CostData(BaseModel):
     total_usd: float = 0.0
     input_tokens: int = 0
     output_tokens: int = 0
+    cache_read_input_tokens: int = 0
+    cache_creation_input_tokens: int = 0
+    cache_read_msats: int = 0
+    cache_creation_msats: int = 0
 
 
 class MaxCostData(CostData):
@@ -29,11 +41,11 @@ class CostDataError(BaseModel):
     code: str
 
 
-async def calculate_cost(  # todo: can be sync
-    response_data: dict, max_cost: int, session: AsyncSession
+async def calculate_cost(
+    response_data: dict,
+    max_cost: int,
 ) -> CostData | MaxCostData | CostDataError:
-    """
-    Calculate the cost of an API request based on token usage.
+    """Calculate the cost of an API request based on token usage.
 
     Args:
         response_data: Response data containing usage information
@@ -41,6 +53,9 @@ async def calculate_cost(  # todo: can be sync
 
     Returns:
         Cost data or error information
+
+    The response's usage object is normalized with the default union parser;
+    this function holds no vendor-dialect knowledge of its own.
     """
     logger.debug(
         "Starting cost calculation",
@@ -51,12 +66,21 @@ async def calculate_cost(  # todo: can be sync
         },
     )
 
-    if "usage" not in response_data or response_data["usage"] is None:
+    usage = normalize_usage(response_data.get("usage"))
+
+    if usage is None:
         logger.warning(
-            "No usage data in response, using base cost only",
+            "No usage data in response — billing at MaxCostData with zero "
+            "tokens. Dashboard will show this request as `(0+0)`. Most "
+            "common cause: upstream stream did not include a final usage "
+            "chunk (OpenAI-compat backends require "
+            "`stream_options.include_usage=true`).",
             extra={
                 "max_cost_msats": max_cost,
                 "model": response_data.get("model", "unknown"),
+                "response_keys": sorted(response_data.keys())
+                if isinstance(response_data, dict)
+                else None,
             },
         )
         return MaxCostData(
@@ -67,116 +91,56 @@ async def calculate_cost(  # todo: can be sync
             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,
         )
 
-    usage_data = response_data["usage"]
+    usage_data = response_data.get("usage") or {}
+    if not isinstance(usage_data, dict):
+        usage_data = {}
 
-    def parse_token_count(value: object) -> int:
-        if isinstance(value, bool):
-            return 0
-        if isinstance(value, int):
-            return max(0, value)
-        if isinstance(value, float):
-            return max(0, int(value))
-        if isinstance(value, str):
-            try:
-                return max(0, int(float(value)))
-            except ValueError:
-                return 0
-        return 0
-
-    input_tokens = parse_token_count(usage_data.get("prompt_tokens", 0))
-    output_tokens = parse_token_count(usage_data.get("completion_tokens", 0))
-    input_tokens = (
-        input_tokens
-        if input_tokens != 0
-        else parse_token_count(usage_data.get("input_tokens", 0))
-    )
-    output_tokens = (
-        output_tokens
-        if output_tokens != 0
-        else parse_token_count(usage_data.get("output_tokens", 0))
-    )
-    input_tokens = (
-        input_tokens
-        if input_tokens != 0
-        else parse_token_count(response_data.get("usage", {}).get("input_tokens", 0))
-    )
-    output_tokens = (
-        output_tokens
-        if output_tokens != 0
-        else parse_token_count(response_data.get("usage", {}).get("output_tokens", 0))
-    )
-
-    usd_cost = 0.0
-    input_usd = 0.0
-    output_usd = 0.0
-
-    if "cost_details" in usage_data:
-        usd_cost = float(
-            usage_data["cost_details"].get("upstream_inference_cost", 0) or 0
-        )
-        input_usd = float(
-            usage_data["cost_details"].get("upstream_inference_prompt_cost", 0) or 0
-        )
-        output_usd = float(
-            usage_data["cost_details"].get("upstream_inference_completions_cost", 0)
-            or 0
-        )
-
-    # Fallback to cost field if upstream_inference_cost is 0
-    if usd_cost == 0 and "cost" in usage_data:
-        try:
-            usd_cost = float(usage_data.get("cost", 0) or 0)
-        except Exception:
-            pass
-
-    MSATS_PER_1K_INPUT_TOKENS: float = (
-        float(settings.fixed_per_1k_input_tokens) * 1000.0
-    )
-    MSATS_PER_1K_OUTPUT_TOKENS: float = (
-        float(settings.fixed_per_1k_output_tokens) * 1000.0
-    )
+    input_tokens = usage.input_tokens
+    output_tokens = usage.output_tokens
+    cache_read_tokens = usage.cache_read_tokens
+    cache_creation_tokens = usage.cache_write_tokens
 
+    # Try USD cost first
+    usd_cost = _resolve_usd_cost(usage_data, response_data)
     if usd_cost > 0:
-        try:
-            sats_per_usd = 1.0 / sats_usd_price()
-            cost_in_sats = usd_cost * sats_per_usd
-            cost_in_msats = math.ceil(cost_in_sats * 1000)
-
-            input_msats = 0
-            output_msats = 0
-
-            if input_usd > 0 or output_usd > 0:
-                input_msats = int((input_usd * sats_per_usd) * 1000)
-                output_msats = int((output_usd * sats_per_usd) * 1000)
-            else:
-                total_tokens = input_tokens + output_tokens
-                if total_tokens > 0:
-                    input_ratio = input_tokens / total_tokens
-                    input_msats = int(cost_in_msats * input_ratio)
-                    output_msats = cost_in_msats - input_msats
-                else:
-                    output_msats = cost_in_msats
-
-            logger.info(
-                "Using cost from usage data/details",
+        if input_tokens == 0 and output_tokens == 0:
+            logger.warning(
+                "Upstream reported a USD cost but no token counts — "
+                "billing the USD-derived cost while the dashboard will "
+                "show this request as `(0+0)` tokens. Check that the "
+                "upstream actually emits `usage.input_tokens` and "
+                "`usage.output_tokens` (OpenAI-compat streams require "
+                "`stream_options.include_usage=true`).",
                 extra={
-                    "usd_cost": usd_cost,
-                    "cost_in_sats": cost_in_sats,
-                    "cost_in_msats": cost_in_msats,
                     "model": response_data.get("model", "unknown"),
+                    "usd_cost": usd_cost,
+                    "usage_keys": sorted(usage_data.keys())
+                    if isinstance(usage_data, dict)
+                    else None,
                 },
             )
-
-            return CostData(
-                base_msats=0,
-                input_msats=input_msats,
-                output_msats=output_msats,
-                total_msats=cost_in_msats,
-                total_usd=usd_cost,
-                input_tokens=input_tokens,
-                output_tokens=output_tokens,
+        try:
+            input_usd = _coerce_usd(
+                usage_data.get("cost_details", {}).get("input_cost", 0)
+            )
+            output_usd = _coerce_usd(
+                usage_data.get("cost_details", {}).get("output_cost", 0)
+            )
+            return _calculate_from_usd_cost(
+                usd_cost,
+                input_usd,
+                output_usd,
+                input_tokens,
+                cache_read_tokens,
+                cache_creation_tokens,
+                output_tokens,
+                response_data,
             )
         except Exception as e:
             logger.warning(
@@ -187,62 +151,35 @@ async def calculate_cost(  # todo: can be sync
                     "model": response_data.get("model", "unknown"),
                 },
             )
-            # Fall through to token-based calculation
 
-    if not settings.fixed_pricing:
-        response_model = response_data.get("model", "")
-        logger.debug(
-            "Using model-based pricing",
-            extra={"model": response_model},
-        )
+    # Fall back to token-based pricing
+    try:
+        pricing_rates = _get_pricing_rates(response_data)
+    except ValueError as e:
+        return CostDataError(message=str(e), code="pricing_error")
 
-        from ..proxy import get_model_instance
+    if pricing_rates is None:
+        input_rate = float(settings.fixed_per_1k_input_tokens) * 1000.0
+        output_rate = float(settings.fixed_per_1k_output_tokens) * 1000.0
+        cache_read_rate = input_rate
+        cache_creation_rate = input_rate
+    else:
+        input_rate, output_rate, cache_read_rate, cache_creation_rate = pricing_rates
 
-        model_obj = get_model_instance(response_model)
-
-        if not model_obj:
-            logger.error(
-                "Invalid model in response",
-                extra={"response_model": response_model},
-            )
-            return CostDataError(
-                message=f"Invalid model in response: {response_model}",
-                code="model_not_found",
-            )
-
-        if not model_obj.sats_pricing:
-            logger.error(
-                "Model pricing not defined",
-                extra={"model": response_model, "model_id": response_model},
-            )
-            return CostDataError(
-                message="Model pricing not defined", code="pricing_not_found"
-            )
-
-        try:
-            mspp = float(model_obj.sats_pricing.prompt)
-            mspc = float(model_obj.sats_pricing.completion)
-        except Exception:
-            return CostDataError(message="Invalid pricing data", code="pricing_invalid")
-
-        MSATS_PER_1K_INPUT_TOKENS = mspp * 1_000_000.0
-        MSATS_PER_1K_OUTPUT_TOKENS = mspc * 1_000_000.0
-
-        logger.info(
-            "Applied model-specific pricing",
-            extra={
-                "model": response_model,
-                "input_price_msats_per_1k": MSATS_PER_1K_INPUT_TOKENS,
-                "output_price_msats_per_1k": MSATS_PER_1K_OUTPUT_TOKENS,
-            },
-        )
-
-    if not (MSATS_PER_1K_OUTPUT_TOKENS and MSATS_PER_1K_INPUT_TOKENS):
+    if not (input_rate and output_rate):
         logger.warning(
-            "No token pricing configured, using base cost",
+            "No token pricing configured — billing at flat MaxCostData. "
+            "Token counts %s in the upstream response but cannot be "
+            "priced; the request will appear in dashboards with the "
+            "raw counts and a fixed max-cost charge.",
+            "are present"
+            if (input_tokens > 0 or output_tokens > 0)
+            else "are zero",
             extra={
                 "base_cost_msats": max_cost,
                 "model": response_data.get("model", "unknown"),
+                "input_tokens": input_tokens,
+                "output_tokens": output_tokens,
             },
         )
         return MaxCostData(
@@ -252,12 +189,216 @@ async def calculate_cost(  # todo: can be sync
             total_msats=max_cost,
             input_tokens=input_tokens,
             output_tokens=output_tokens,
+            cache_read_input_tokens=cache_read_tokens,
+            cache_creation_input_tokens=cache_creation_tokens,
+            cache_read_msats=0,
+            cache_creation_msats=0,
         )
 
-    calc_input_msats = round(input_tokens / 1000 * MSATS_PER_1K_INPUT_TOKENS, 3)
+    return _calculate_from_tokens(
+        input_tokens,
+        output_tokens,
+        cache_read_tokens,
+        cache_creation_tokens,
+        input_rate,
+        output_rate,
+        cache_read_rate,
+        cache_creation_rate,
+        response_data,
+    )
 
-    calc_output_msats = round(output_tokens / 1000 * MSATS_PER_1K_OUTPUT_TOKENS, 3)
-    token_based_cost = math.ceil(calc_input_msats + calc_output_msats)
+
+# ============================================================================
+# Helper Functions (ordered by call sequence in calculate_cost)
+# ============================================================================
+
+
+def _coerce_usd(value: object) -> float:
+    """Coerce a value to USD float, handling various formats safely."""
+    if value is None or isinstance(value, bool):
+        return 0.0
+    if not isinstance(value, (int, float, str)):
+        return 0.0
+    try:
+        return max(0.0, float(value))
+    except (TypeError, ValueError):
+        return 0.0
+
+
+def _resolve_usd_cost(usage_data: dict, response_data: dict) -> float:
+    """Resolve USD cost with clear priority order.
+
+    Priority: cost_details.total_cost → total_cost → cost (in both usage and response).
+    """
+    cost_details = usage_data.get("cost_details")
+    if isinstance(cost_details, dict):
+        cost = _coerce_usd(cost_details.get("total_cost"))
+        if cost > 0:
+            return cost
+
+    for source in [usage_data, response_data]:
+        if not isinstance(source, dict):
+            continue
+        for field in ("total_cost", "cost"):
+            cost = _coerce_usd(source.get(field))
+            if cost > 0:
+                return cost
+
+    return 0.0
+
+
+def _get_pricing_rates(
+    response_data: dict,
+) -> tuple[float, float, float, float] | None:
+    """Get model-based pricing rates or None if using fixed pricing.
+
+    Returns: (input_rate, output_rate, cache_read_rate, cache_write_rate)
+    """
+    if settings.fixed_pricing:
+        return None
+
+    from ..proxy import get_model_instance
+
+    response_model = response_data.get("model", "")
+    model_obj = get_model_instance(response_model)
+
+    if not model_obj:
+        logger.error("Invalid model in response", extra={"response_model": response_model})
+        raise ValueError(f"Invalid model: {response_model}")
+
+    if not model_obj.sats_pricing:
+        logger.error(
+            "Model pricing not defined",
+            extra={"model": response_model, "model_id": response_model},
+        )
+        raise ValueError("Model pricing not defined")
+
+    try:
+        mspp = float(model_obj.sats_pricing.prompt)
+        mspc = float(model_obj.sats_pricing.completion)
+        mscr = float(model_obj.sats_pricing.input_cache_read or 0)
+        mscw = float(model_obj.sats_pricing.input_cache_write or 0)
+
+        mspp_1k = mspp * 1_000_000.0
+        mspc_1k = mspc * 1_000_000.0
+        mscr_1k = mscr * 1_000_000.0 if mscr > 0 else mspp_1k
+        mscw_1k = mscw * 1_000_000.0 if mscw > 0 else mspp_1k
+
+        logger.info(
+            "Applied model-specific pricing",
+            extra={
+                "model": response_model,
+                "input_price_msats_per_1k": mspp_1k,
+                "output_price_msats_per_1k": mspc_1k,
+                "cache_read_price_msats_per_1k": mscr_1k,
+                "cache_write_price_msats_per_1k": mscw_1k,
+            },
+        )
+        return mspp_1k, mspc_1k, mscr_1k, mscw_1k
+    except Exception as e:
+        logger.error("Invalid pricing data", extra={"error": str(e)})
+        raise ValueError("Invalid pricing data") from e
+
+
+def _resolve_provider_fee(model_id: str) -> float:
+    """Resolve the provider fee multiplier for the given model id.
+
+    Falls back to 1.0 (no markup) when the provider cannot be resolved so
+    the USD cost path never silently double-applies or omits the fee.
+    """
+    from ..proxy import get_provider_for_model
+
+    if not model_id:
+        return 1.0
+    providers = get_provider_for_model(model_id)
+    if not providers:
+        return 1.0
+    return float(providers[0].provider_fee)
+
+
+def _calculate_from_usd_cost(
+    usd_cost: float,
+    input_usd: float,
+    output_usd: float,
+    input_tokens: int,
+    cache_read_tokens: int,
+    cache_creation_tokens: int,
+    output_tokens: int,
+    response_data: dict,
+) -> CostData:
+    """Calculate cost from USD figures, deriving input/output split from tokens."""
+    provider_fee = _resolve_provider_fee(response_data.get("model", ""))
+    usd_cost = usd_cost * provider_fee
+    input_usd = input_usd * provider_fee
+    output_usd = output_usd * provider_fee
+    sats_per_usd = 1.0 / sats_usd_price()
+    cost_in_sats = usd_cost * sats_per_usd
+    cost_in_msats = math.ceil(cost_in_sats * 1000)
+
+    if input_usd > 0 or output_usd > 0:
+        input_msats = int((input_usd * sats_per_usd) * 1000)
+        output_msats = int((output_usd * sats_per_usd) * 1000)
+    else:
+        effective_input_tokens = (
+            input_tokens + cache_read_tokens + cache_creation_tokens
+        )
+        total_tokens = effective_input_tokens + output_tokens
+        input_msats = (
+            int(cost_in_msats * effective_input_tokens / total_tokens)
+            if total_tokens > 0
+            else 0
+        )
+        output_msats = cost_in_msats - input_msats
+
+    logger.info(
+        "Using cost from usage data/details",
+        extra={
+            "usd_cost": usd_cost,
+            "cost_in_sats": cost_in_sats,
+            "cost_in_msats": cost_in_msats,
+            "model": response_data.get("model", "unknown"),
+        },
+    )
+
+    return CostData(
+        base_msats=0,
+        input_msats=input_msats,
+        output_msats=output_msats,
+        total_msats=cost_in_msats,
+        total_usd=usd_cost,
+        input_tokens=input_tokens,
+        output_tokens=output_tokens,
+        cache_read_input_tokens=cache_read_tokens,
+        cache_creation_input_tokens=cache_creation_tokens,
+        cache_read_msats=0,
+        cache_creation_msats=0,
+    )
+
+
+def _calculate_from_tokens(
+    input_tokens: int,
+    output_tokens: int,
+    cache_read_tokens: int,
+    cache_creation_tokens: int,
+    input_rate: float,
+    output_rate: float,
+    cache_read_rate: float,
+    cache_creation_rate: float,
+    response_data: dict,
+) -> CostData:
+    """Calculate cost from token counts using pricing rates."""
+    calc_input_msats = round(input_tokens / 1000 * input_rate, 3)
+    calc_output_msats = round(output_tokens / 1000 * output_rate, 3)
+    calc_cache_read_msats = round(cache_read_tokens / 1000 * cache_read_rate, 3)
+    calc_cache_write_msats = round(
+        cache_creation_tokens / 1000 * cache_creation_rate, 3
+    )
+    token_based_cost = math.ceil(
+        calc_input_msats
+        + calc_output_msats
+        + calc_cache_read_msats
+        + calc_cache_write_msats
+    )
     total_usd = (token_based_cost / 1000.0) * sats_usd_price()
 
     logger.info(
@@ -265,20 +406,39 @@ async def calculate_cost(  # todo: can be sync
         extra={
             "input_tokens": input_tokens,
             "output_tokens": output_tokens,
+            "cache_read_input_tokens": cache_read_tokens,
+            "cache_creation_input_tokens": cache_creation_tokens,
             "input_cost_msats": calc_input_msats,
             "output_cost_msats": calc_output_msats,
+            "cache_read_cost_msats": calc_cache_read_msats,
+            "cache_creation_cost_msats": calc_cache_write_msats,
             "total_cost_msats": token_based_cost,
             "total_usd": total_usd,
             "model": response_data.get("model", "unknown"),
         },
     )
 
+    # Fold the cache-read/write cost into the visible ``input_msats`` so a
+    # dashboard that renders I / O / T sees ``input + output == total``
+    # exactly. This mirrors ``_fold_cache_into_input_tokens`` (which rolls the
+    # cache token counts into the visible prompt total). The standalone
+    # ``cache_read_msats`` / ``cache_creation_msats`` fields stay populated for
+    # clients that want the breakdown; nothing sums the components to derive
+    # ``total_msats`` (it is computed independently above), so this is
+    # display-only and does not change what is billed.
+    visible_output_msats = int(calc_output_msats)
+    visible_input_msats = token_based_cost - visible_output_msats
+
     return CostData(
         base_msats=0,
-        input_msats=int(calc_input_msats),
-        output_msats=int(calc_output_msats),
+        input_msats=visible_input_msats,
+        output_msats=visible_output_msats,
         total_msats=token_based_cost,
         total_usd=total_usd,
         input_tokens=input_tokens,
         output_tokens=output_tokens,
+        cache_read_input_tokens=cache_read_tokens,
+        cache_creation_input_tokens=cache_creation_tokens,
+        cache_read_msats=int(calc_cache_read_msats),
+        cache_creation_msats=int(calc_cache_write_msats),
     )

routstr/payment/helpers.py

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

routstr/payment/models.py

@@ -3,7 +3,8 @@ import json
 import random
 
 import httpx
-from fastapi import APIRouter, Depends
+from fastapi import APIRouter, Depends, HTTPException, Request
+from pydantic import BaseModel as V2BaseModel
 from pydantic.v1 import BaseModel
 from sqlmodel.ext.asyncio.session import AsyncSession
 
@@ -16,6 +17,24 @@ logger = get_logger(__name__)
 
 models_router = APIRouter()
 
+_MODEL_TEST_ENDPOINT_PATHS = {
+    "chat-completions": "chat/completions",
+    "completions": "completions",
+    "embeddings": "embeddings",
+    "responses": "responses",
+}
+
+# Cap the caller-supplied test payload to avoid forwarding oversized bodies
+# upstream on the operator's credentials.
+_MODEL_TEST_MAX_REQUEST_BYTES = 64 * 1024
+
+
+async def _require_admin_api(request: Request) -> None:
+    """Require admin auth without creating an import-time cycle with core.admin."""
+    from ..core.admin import require_admin_api
+
+    await require_admin_api(request)
+
 
 class Architecture(BaseModel):
     modality: str
@@ -66,6 +85,48 @@ class Model(BaseModel):
         return hash(self.id)
 
 
+def backfill_cache_pricing(model_id: str, pricing: Pricing) -> Pricing:
+    """Fill missing cache rates from litellm's bundled cost map.
+
+    The OpenRouter model feed omits ``input_cache_read``/``input_cache_write``
+    for many models (most DeepSeek entries, openai/gpt-4o, ...). Without a
+    cache rate, billing falls back to the full input rate, which overcharges
+    cache reads (DeepSeek hits are 10x cheaper) and undercharges Anthropic
+    cache writes (1.25x). litellm ships per-model USD rates keyed by the exact
+    OpenRouter id (deepseek/deepseek-chat) or by the bare model name
+    (gpt-4o, claude-sonnet-4-5), so both spellings are tried.
+
+    Rates already present (e.g. provided by OpenRouter) are authoritative and
+    never overwritten. Unknown models are returned unchanged.
+    """
+    needs_read = (pricing.input_cache_read or 0.0) <= 0.0
+    needs_write = (pricing.input_cache_write or 0.0) <= 0.0
+    if not (needs_read or needs_write):
+        return pricing
+
+    import litellm
+
+    info: dict | None = None
+    for key in (model_id, model_id.split("/", 1)[-1]):
+        candidate = litellm.model_cost.get(key)
+        if isinstance(candidate, dict):
+            info = candidate
+            break
+    if info is None:
+        return pricing
+
+    updated = Pricing.parse_obj(pricing.dict())
+    if needs_read:
+        read_rate = info.get("cache_read_input_token_cost")
+        if isinstance(read_rate, (int, float)) and read_rate > 0:
+            updated.input_cache_read = float(read_rate)
+    if needs_write:
+        write_rate = info.get("cache_creation_input_token_cost")
+        if isinstance(write_rate, (int, float)) and write_rate > 0:
+            updated.input_cache_write = float(write_rate)
+    return updated
+
+
 def _has_valid_pricing(model: dict) -> bool:
     """Check if model has valid pricing (not free, no negative values)."""
     pricing = model.get("pricing", {})
@@ -350,6 +411,9 @@ async def _update_sats_pricing_once() -> None:
     from ..proxy import get_upstreams, refresh_model_maps
 
     upstreams = get_upstreams()
+    if not upstreams:
+        return
+
     sats_to_usd = sats_usd_price()
 
     updated_count = 0
@@ -359,11 +423,14 @@ async def _update_sats_pricing_once() -> None:
             for m in upstream.get_cached_models()
         ]
         upstream._models_cache = updated_models
-        upstream._models_by_id = {m.id: m for m in updated_models}
+        upstream._models_by_id = {m.forwarded_model_id or m.id: m for m in updated_models}
         updated_count += len(updated_models)
 
     if updated_count > 0:
-        logger.info("Updated sats pricing", extra={"models_updated": updated_count})
+        logger.info(
+            f"Updated sats pricing for {updated_count} models",
+            extra={"models_updated": updated_count},
+        )
         await refresh_model_maps()
 
 
@@ -405,13 +472,15 @@ async def update_sats_pricing() -> None:
             logger.error(f"Error updating sats pricing: {e}")
 
 
-class ModelTestRequest(BaseModel):
+class ModelTestRequest(V2BaseModel):
     model_id: str
     endpoint_type: str
     request_data: dict
 
 
-@models_router.post("/api/models/test")
+@models_router.post(
+    "/api/models/test", dependencies=[Depends(_require_admin_api)]
+)
 async def test_model(
     payload: ModelTestRequest,
     session: AsyncSession = Depends(get_session),
@@ -439,16 +508,35 @@ async def test_model(
             "status_code": 404,
         }
 
-    base_url = provider.base_url.rstrip("/")
-    if payload.endpoint_type == "chat-completions":
-        url = f"{base_url}/chat/completions"
-    else:
-        url = f"{base_url}/{payload.endpoint_type}"
+    endpoint_path = _MODEL_TEST_ENDPOINT_PATHS.get(payload.endpoint_type)
+    if endpoint_path is None:
+        raise HTTPException(status_code=400, detail="Unsupported endpoint_type")
 
     actual_model_id = model_row.forwarded_model_id or model_row.id
     request_data = dict(payload.request_data)
     request_data["model"] = actual_model_id
 
+    try:
+        request_size = len(json.dumps(request_data).encode("utf-8"))
+    except (TypeError, ValueError):
+        raise HTTPException(status_code=400, detail="Invalid request_data")
+    if request_size > _MODEL_TEST_MAX_REQUEST_BYTES:
+        raise HTTPException(status_code=413, detail="request_data too large")
+
+    base_url = provider.base_url.rstrip("/")
+    url = f"{base_url}/{endpoint_path}"
+
+    logger.info(
+        "admin model test",
+        extra={
+            "model_id": payload.model_id,
+            "forwarded_model_id": actual_model_id,
+            "endpoint_type": payload.endpoint_type,
+            "upstream_provider_id": model_row.upstream_provider_id,
+            "request_bytes": request_size,
+        },
+    )
+
     headers = {
         "Content-Type": "application/json",
         "Authorization": f"Bearer {provider.api_key}",

routstr/payment/usage.py

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

routstr/proxy.py

@@ -1,3 +1,4 @@
+import asyncio
 import json
 from typing import Any
 
@@ -17,16 +18,20 @@ from .core.db import (
     get_session,
 )
 from .core.exceptions import UpstreamError
+from .core.not_found import build_not_found_response
 from .core.settings import settings
 from .payment.helpers import (
     calculate_discounted_max_cost,
     check_token_balance,
     create_error_response,
+    create_upstream_error_response,
     get_max_cost_for_model,
 )
 from .payment.models import Model
 from .upstream import BaseUpstreamProvider
+from .upstream.ehbp import forward_ehbp_request, forward_ehbp_x_cashu_request
 from .upstream.helpers import init_upstreams
+from .upstream.request_correction import correct_request, extract_error_message
 
 logger = get_logger(__name__)
 proxy_router = APIRouter()
@@ -100,6 +105,29 @@ def get_unique_models() -> list[Model]:
     return list(_unique_models.values())
 
 
+def _is_tinfoil_attestation_path(path: str) -> bool:
+    """Return True for Tinfoil attestation-bundle proxy paths."""
+    return path in {"attestation", "tee/attestation"}
+
+
+def _select_unauthenticated_get_upstreams(
+    path: str, upstreams: list[BaseUpstreamProvider]
+) -> list[BaseUpstreamProvider]:
+    """Select upstream candidates for unauthenticated GET bypass paths.
+
+    Tinfoil attestation endpoints are provider-specific. Trying every enabled
+    upstream can return an unrelated provider's 404 before Tinfoil is reached,
+    so route those paths only to Tinfoil providers.
+    """
+    if _is_tinfoil_attestation_path(path):
+        return [
+            upstream
+            for upstream in upstreams
+            if getattr(upstream, "provider_type", None) == "tinfoil"
+        ]
+    return upstreams
+
+
 async def refresh_model_maps() -> None:
     """Refresh global model and provider maps using the cost-based algorithm."""
     from sqlalchemy.orm import selectinload
@@ -150,20 +178,107 @@ async def refresh_model_maps_periodically() -> None:
             )
 
 
+_API_PATH_PREFIXES = (
+    "v1/",
+    "responses",
+    "chat/",
+    "completions",
+    "models",
+    "embeddings",
+    "audio/",
+    "images/",
+    "moderations",
+    "providers",
+    "tee/",
+    "attestation",
+    ".well-known/",
+)
+
+
 @proxy_router.api_route("/{path:path}", methods=["GET", "POST"], response_model=None)
 async def proxy(
     request: Request, path: str, session: AsyncSession = Depends(get_session)
 ) -> Response | StreamingResponse:
+    # GET requests must hit a known API prefix; otherwise return a 404 (HTML
+    # for browsers, JSON for API clients). POST requests are always forwarded
+    # so that OpenAI-style endpoints work with or without the `v1/` prefix
+    # (e.g. `/chat/completions` as well as `/v1/chat/completions`).
+    if request.method == "GET" and not path.startswith(_API_PATH_PREFIXES):
+        return build_not_found_response(request, path)
+
     headers = dict(request.headers)
 
     is_responses_api = path.startswith("v1/responses") or path.startswith("responses")
     request_body = await request.body()
-    request_body_dict = parse_request_body_json(request_body, path)
 
-    if is_responses_api:
-        model_id = extract_model_from_responses_request(request_body_dict)
+    # EHBP (Encrypted HTTP Body Protocol) requests carry an Ehbp-Encapsulated-Key
+    # header and a binary HPKE-sealed body. The proxy cannot parse the body to
+    # extract the model id, so the SDK sends it in X-Routstr-Model. Forward the
+    # raw encrypted body to the upstream's /private/ endpoint and stream the
+    # encrypted response back untouched — the SDK's SecureClient decrypts it.
+    is_ehbp = "ehbp-encapsulated-key" in headers
+    if is_ehbp:
+        request_body_dict = {}
+        model_id = headers.get("x-routstr-model", "")
+        if not model_id:
+            return create_error_response(
+                "invalid_request",
+                "EHBP request missing X-Routstr-Model header",
+                400,
+                request=request,
+            )
     else:
-        model_id = request_body_dict.get("model", "unknown")
+        request_body_dict = parse_request_body_json(request_body, path)
+        if is_responses_api:
+            model_id = extract_model_from_responses_request(request_body_dict)
+        else:
+            model_id = request_body_dict.get("model", "unknown")
+
+    # /tee/*, /attestation and /.well-known/* GET requests don't map to models
+    # — forward without model/cost/auth lookups. Tinfoil attestation paths are
+    # routed only to Tinfoil providers so an unrelated upstream's 404 cannot
+    # short-circuit before the attestation proxy is tried.
+    if request.method == "GET" and (
+        path.startswith("tee/")
+        or path.startswith("attestation")
+        or path.startswith(".well-known/")
+    ):
+        selected_upstreams = _select_unauthenticated_get_upstreams(path, _upstreams)
+        if not selected_upstreams:
+            return create_error_response(
+                "upstream_error",
+                "No upstream available for unauthenticated GET path",
+                502,
+                request=request,
+            )
+
+        last_error_response = None
+        for i, upstream in enumerate(selected_upstreams):
+            try:
+                headers = upstream.prepare_headers(dict(request.headers))
+                response = await upstream.forward_get_request(request, path, headers)
+                if response.status_code in [502, 429] and i < len(selected_upstreams) - 1:
+                    logger.warning(
+                        "Upstream %s returned %s for unauthenticated GET %s, trying next",
+                        upstream.provider_type,
+                        response.status_code,
+                        path,
+                    )
+                    continue
+                return response
+            except UpstreamError as e:
+                logger.warning(
+                    "Upstream %s failed for unauthenticated GET %s: %s",
+                    upstream.provider_type,
+                    path,
+                    e,
+                )
+                if i == len(selected_upstreams) - 1:
+                    last_error_response = create_upstream_error_response(e, request)
+                continue
+        return last_error_response or create_error_response(
+            "upstream_error", "All upstreams failed", 502, request=request
+        )
 
     model_obj = get_model_instance(model_id)
 
@@ -181,6 +296,16 @@ async def proxy(
             request=request,
         )
 
+    if is_ehbp:
+        upstreams = [upstream for upstream in upstreams if upstream.supports_ehbp]
+        if not upstreams:
+            return create_error_response(
+                "unsupported_request",
+                f"No EHBP-capable provider found for model '{model_id}'",
+                400,
+                request=request,
+            )
+
     # todo figure out cost calculation since fallback provider is usually not the same price
     # Use first provider for initial checks/cost calculation
     # primary_upstream = upstreams[0]
@@ -200,7 +325,23 @@ async def proxy(
         last_error = None
         for i, upstream in enumerate(upstreams):
             try:
-                if is_responses_api:
+                if is_ehbp:
+                    if not upstream.supports_ehbp:
+                        logger.warning(
+                            "Upstream %s does not support EHBP for model=%s",
+                            upstream.provider_type,
+                            model_id,
+                        )
+                        continue
+                    return await forward_ehbp_x_cashu_request(
+                        request=request,
+                        x_cashu_token=x_cashu,
+                        path=path,
+                        max_cost_for_model=max_cost_for_model,
+                        model_obj=model_obj,
+                        upstream=upstream,
+                    )
+                elif is_responses_api:
                     return await upstream.handle_x_cashu_responses(
                         request, x_cashu, path, max_cost_for_model, model_obj
                     )
@@ -210,17 +351,24 @@ async def proxy(
                     )
             except UpstreamError as e:
                 logger.warning(
-                    f"Upstream {upstream.provider_type} failed (x-cashu): {e}"
+                    "Upstream %s failed (x-cashu) for model=%s: %s",
+                    upstream.provider_type,
+                    model_id,
+                    e,
+                    extra={
+                        "provider": upstream.provider_type,
+                        "model": model_id,
+                        "status_code": e.status_code,
+                    },
                 )
                 if i == len(upstreams) - 1:
                     last_error = e
                 continue
 
+        if last_error is not None:
+            return create_upstream_error_response(last_error, request)
         return create_error_response(
-            "upstream_error",
-            str(last_error) if last_error else "All upstreams failed",
-            502,
-            request=request,
+            "upstream_error", "All upstreams failed", 502, request=request
         )
 
     elif auth := headers.get("authorization", None):
@@ -276,61 +424,117 @@ async def proxy(
             except UpstreamError as e:
                 logger.warning(f"Upstream {upstream.provider_type} failed (GET): {e}")
                 if i == len(upstreams) - 1:
-                    last_error_response = create_error_response(
-                        "upstream_error", str(e), 502, request=request
-                    )
+                    last_error_response = create_upstream_error_response(e, request)
                 continue
         return last_error_response or create_error_response(
             "upstream_error", "All upstreams failed", 502, request=request
         )
 
-    if request_body_dict:
+    if is_ehbp or request_body_dict:
         await pay_for_request(key, max_cost_for_model, session)
 
+    # Tracks request params already removed in response to upstream rejections,
+    # shared across providers so a stripped param stays stripped on failover and
+    # the reactive retry can never loop unboundedly.
+    already_stripped: set[str] = set()
+
     for i, upstream in enumerate(upstreams):
         headers = upstream.prepare_headers(dict(request.headers))
 
         try:
-            try:
-                if is_responses_api:
-                    response = await upstream.forward_responses_request(
-                        request,
-                        path,
-                        headers,
-                        request_body,
-                        key,
-                        max_cost_for_model,
-                        session,
-                        model_obj,
+            while True:
+                try:
+                    if is_ehbp:
+                        if not upstream.supports_ehbp:
+                            logger.warning(
+                                "Upstream %s does not support EHBP for model=%s",
+                                upstream.provider_type,
+                                model_id,
+                            )
+                            raise UpstreamError(
+                                f"Provider {upstream.provider_type} does not support EHBP",
+                                status_code=400,
+                            )
+                        response = await forward_ehbp_request(
+                            request=request,
+                            path=path,
+                            headers=headers,
+                            request_body=request_body,
+                            upstream=upstream,
+                            key=key,
+                            max_cost_for_model=max_cost_for_model,
+                            session=session,
+                            model_obj=model_obj,
+                        )
+                    elif is_responses_api:
+                        response = await upstream.forward_responses_request(
+                            request,
+                            path,
+                            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,
+                        )
+                except UpstreamError:
+                    # Let the outer UpstreamError handler manage retry/revert
+                    raise
+                except Exception as e:
+                    # Unexpected error (not an upstream failure) — revert and propagate
+                    logger.error(
+                        "Unexpected error in upstream request, reverting payment",
+                        extra={
+                            "error": str(e),
+                            "error_type": type(e).__name__,
+                            "path": path,
+                            "key_hash": key.hashed_key[:8] + "...",
+                            "max_cost_for_model": max_cost_for_model,
+                        },
                     )
-                else:
-                    response = await upstream.forward_request(
-                        request,
-                        path,
-                        headers,
+                    await revert_pay_for_request(key, session, max_cost_for_model)
+                    raise
+
+                # Reactive recovery: some models reject one specific request
+                # param (e.g. newer Anthropic models deprecating `temperature`).
+                # When the upstream 400s naming such a param, strip it from the
+                # body and retry the SAME upstream. ``already_stripped`` bounds
+                # this to one retry per distinct param so it always terminates.
+                if response.status_code == 400 and not is_ehbp:
+                    correction = correct_request(
                         request_body,
-                        key,
-                        max_cost_for_model,
-                        session,
-                        model_obj,
+                        extract_error_message(response),
+                        already_stripped,
                     )
-            except UpstreamError:
-                # Let the outer UpstreamError handler manage retry/revert
-                raise
-            except Exception as e:
-                # Unexpected error (not an upstream failure) — revert and propagate
-                logger.error(
-                    "Unexpected error in upstream request, reverting payment",
-                    extra={
-                        "error": str(e),
-                        "error_type": type(e).__name__,
-                        "path": path,
-                        "key_hash": key.hashed_key[:8] + "...",
-                        "max_cost_for_model": max_cost_for_model,
-                    },
-                )
-                await revert_pay_for_request(key, session, max_cost_for_model)
-                raise
+                    if correction is not None:
+                        request_body, bad_param = correction.body, correction.label
+                        already_stripped.add(bad_param)
+                        logger.warning(
+                            "Upstream %s rejected param '%s' for model=%s; "
+                            "stripping and retrying same upstream",
+                            upstream.provider_type,
+                            bad_param,
+                            model_id,
+                            extra={
+                                "provider": upstream.provider_type,
+                                "model": model_id,
+                                "stripped_param": bad_param,
+                                "path": path,
+                            },
+                        )
+                        continue
+                break
 
             if response.status_code != 200:
                 # Check if we should retry (502 Upstream Error or 429 Rate Limit)
@@ -355,10 +559,14 @@ async def proxy(
                     )
 
                     logger.warning(
-                        f"Upstream {upstream.provider_type} returned {response.status_code}, trying next provider",
+                        "Upstream %s returned %s for model=%s, trying next provider",
+                        upstream.provider_type,
+                        response.status_code,
+                        model_id,
                         extra={
                             "status_code": response.status_code,
-                            "upstream": upstream.provider_type,
+                            "provider": upstream.provider_type,
+                            "model": model_id,
                         },
                     )
                     continue
@@ -366,34 +574,59 @@ async def proxy(
                 # 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",
+                    "Upstream request failed, revert payment "
+                    "(provider=%s model=%s status=%s path=%s)",
+                    upstream.provider_type,
+                    model_id,
+                    response.status_code,
+                    path,
                     extra={
                         "status_code": response.status_code,
                         "path": path,
+                        "provider": upstream.provider_type,
+                        "model": model_id,
                         "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 asyncio.CancelledError:
+            logger.warning(
+                "Client disconnected mid-request, reverting reservation",
+                extra={
+                    "path": path,
+                    "model": model_id,
+                    "key_hash": key.hashed_key[:8] + "...",
+                    "max_cost_for_model": max_cost_for_model,
+                },
+            )
+            await asyncio.shield(
+                revert_pay_for_request(key, session, max_cost_for_model)
+            )
+            raise
+
         except UpstreamError as e:
             logger.warning(
-                f"Upstream {upstream.provider_type} failed: {e}",
-                extra={"retry": i < len(upstreams) - 1},
+                "Upstream %s failed for model=%s: %s",
+                upstream.provider_type,
+                model_id,
+                e,
+                extra={
+                    "provider": upstream.provider_type,
+                    "model": model_id,
+                    "status_code": e.status_code,
+                    "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
-                )
+                return create_upstream_error_response(e, request)
 
             # Otherwise loop continues to next provider
             continue

routstr/upstream/__init__.py

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

routstr/upstream/anthropic.py

@@ -13,6 +13,8 @@ class AnthropicUpstreamProvider(BaseUpstreamProvider):
     provider_type = "anthropic"
     default_base_url = "https://api.anthropic.com/v1"
     platform_url = "https://console.anthropic.com/settings/keys"
+    supports_anthropic_messages = True
+    litellm_provider_prefix = "anthropic/"
 
     def __init__(self, api_key: str, provider_fee: float = 1.01):
         super().__init__(

routstr/upstream/azure.py

@@ -13,6 +13,7 @@ class AzureUpstreamProvider(BaseUpstreamProvider):
     provider_type = "azure"
     default_base_url = None
     platform_url = "https://portal.azure.com/"
+    litellm_provider_prefix = "azure/"
 
     def __init__(
         self,

routstr/upstream/cache_breakpoints.py

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

routstr/upstream/clients/gemini.py

@@ -26,14 +26,14 @@ class GeminiClient(BaseAPIClient):
         max_tokens: int | None = None,
         **kwargs: Any,
     ) -> dict[str, Any]:
-        from openai import NOT_GIVEN
+        from openai import omit
 
         response = await self.client.chat.completions.create(
             model=model,
             messages=messages,  # type: ignore
-            temperature=temperature if temperature is not None else NOT_GIVEN,
-            max_tokens=max_tokens if max_tokens is not None else NOT_GIVEN,
-            top_p=kwargs.get("top_p", NOT_GIVEN),
+            temperature=temperature if temperature is not None else omit,
+            max_tokens=max_tokens if max_tokens is not None else omit,
+            top_p=kwargs.get("top_p", omit),
         )
         return response.model_dump()
 
@@ -45,7 +45,7 @@ class GeminiClient(BaseAPIClient):
         max_tokens: int | None = None,
         **kwargs: Any,
     ) -> AsyncGenerator[dict[str, Any], None]:
-        from openai import NOT_GIVEN
+        from openai import omit
 
         usage_callback = kwargs.get("usage_callback")
         completion_callback = kwargs.get("completion_callback")
@@ -55,9 +55,9 @@ class GeminiClient(BaseAPIClient):
             messages=messages,  # type: ignore
             stream=True,
             stream_options={"include_usage": True},
-            temperature=temperature if temperature is not None else NOT_GIVEN,
-            max_tokens=max_tokens if max_tokens is not None else NOT_GIVEN,
-            top_p=kwargs.get("top_p", NOT_GIVEN),
+            temperature=temperature if temperature is not None else omit,
+            max_tokens=max_tokens if max_tokens is not None else omit,
+            top_p=kwargs.get("top_p", omit),
         )
 
         final_usage = None

routstr/upstream/count_tokens.py

@@ -0,0 +1,108 @@
+"""Local handling of Anthropic ``/v1/messages/count_tokens`` for upstreams
+that do not natively expose the endpoint.
+
+Most non-Anthropic upstreams (OpenAI-compat, Gemini OpenAI-compat,
+OpenRouter chat-completions, generic providers) return 400/404 when asked
+to ``POST /messages/count_tokens``. Claude Code and other Anthropic SDK
+clients call this endpoint before each turn to size context windows and
+trigger compaction, so a failure breaks the whole chat.
+
+We answer locally. ``litellm.token_counter`` understands the Anthropic
+message shape and the per-model tokenizers, so we prefer it. If it raises
+(unknown model, encoding lookup failure, ...), we fall back to the
+project's own ``estimate_tokens`` heuristic, which is always defined and
+never raises.
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+import litellm
+from fastapi.responses import Response
+
+from ..core import get_logger
+from ..payment.helpers import estimate_tokens
+from ..payment.models import Model
+
+logger = get_logger(__name__)
+
+
+def _parse_request_body(request_body: bytes | None) -> dict[str, Any]:
+    if not request_body:
+        return {}
+    try:
+        parsed = json.loads(request_body)
+    except (ValueError, TypeError):
+        return {}
+    return parsed if isinstance(parsed, dict) else {}
+
+
+def _count_with_litellm(model: str, body: dict[str, Any]) -> int:
+    messages = body.get("messages")
+    if not isinstance(messages, list):
+        messages = []
+
+    system = body.get("system")
+    if isinstance(system, str) and system:
+        messages = [{"role": "system", "content": system}, *messages]
+    elif isinstance(system, list):
+        text = "".join(
+            block.get("text", "")
+            for block in system
+            if isinstance(block, dict) and block.get("type") == "text"
+        )
+        if text:
+            messages = [{"role": "system", "content": text}, *messages]
+
+    tools = body.get("tools") if isinstance(body.get("tools"), list) else None
+
+    return int(
+        litellm.token_counter(
+            model=model,
+            messages=messages,
+            tools=tools,
+        )
+    )
+
+
+def count_tokens_locally(
+    request_body: bytes | None,
+    model_obj: Model | None,
+) -> Response:
+    """Return an Anthropic-compatible count_tokens response without
+    touching the upstream. Always returns 200; never raises."""
+    body = _parse_request_body(request_body)
+
+    model_name = ""
+    if model_obj is not None:
+        model_name = model_obj.forwarded_model_id or model_obj.id or ""
+    if not model_name:
+        body_model = body.get("model")
+        if isinstance(body_model, str):
+            model_name = body_model
+
+    input_tokens: int
+    try:
+        input_tokens = _count_with_litellm(model_name, body)
+    except Exception as exc:
+        messages = body.get("messages")
+        fallback_messages = messages if isinstance(messages, list) else []
+        input_tokens = estimate_tokens(fallback_messages)
+        logger.debug(
+            "litellm token_counter failed; using local estimator",
+            extra={
+                "model": model_name,
+                "error": str(exc),
+                "error_type": type(exc).__name__,
+                "estimated_tokens": input_tokens,
+            },
+        )
+
+    payload = {"input_tokens": max(0, int(input_tokens))}
+    return Response(
+        content=json.dumps(payload).encode(),
+        status_code=200,
+        media_type="application/json",
+    )

routstr/upstream/deepseek_v4_pricing_shim.py

@@ -0,0 +1,66 @@
+"""TEMPORARY: local DeepSeek V4 pricing shim.
+
+litellm's bundled cost map does not yet ship ``deepseek-v4-flash`` /
+``deepseek-v4-pro``. Without an entry, ``backfill_cache_pricing`` cannot find a
+``cache_read_input_token_cost`` and cache reads fall back to the full input
+rate — a ~60% overcharge on cache hits (DeepSeek hits are ~0.2x input).
+
+This module injects the missing entries into ``litellm.model_cost`` at startup
+so the existing backfill path resolves them. Rates mirror the open upstream PR
+https://github.com/BerriAI/litellm/pull/26380 (issue
+https://github.com/BerriAI/litellm/issues/30430).
+
+=== REMOVAL (once litellm ships these models) ===
+Delete this file and the single ``register_deepseek_v4_pricing()`` call in
+``routstr/core/main.py``. Nothing else depends on it. Entries are only added
+when absent, so a stale shim is harmless after upstream lands — but remove it.
+"""
+
+import litellm
+
+from ..core import get_logger
+
+logger = get_logger(__name__)
+
+# USD per token. Source: BerriAI/litellm PR #26380.
+_DEEPSEEK_V4_RATES: dict[str, dict[str, float]] = {
+    "deepseek-v4-flash": {
+        "input_cost_per_token": 1.4e-07,
+        "output_cost_per_token": 2.8e-07,
+        "cache_read_input_token_cost": 2.8e-08,
+        "cache_creation_input_token_cost": 0.0,
+        "input_cost_per_token_cache_hit": 2.8e-08,
+    },
+    "deepseek-v4-pro": {
+        "input_cost_per_token": 1.74e-06,
+        "output_cost_per_token": 3.48e-06,
+        "cache_read_input_token_cost": 1.4e-07,
+        "cache_creation_input_token_cost": 0.0,
+        "input_cost_per_token_cache_hit": 1.4e-07,
+    },
+}
+
+
+def register_deepseek_v4_pricing() -> None:
+    """Inject DeepSeek V4 pricing into ``litellm.model_cost`` if absent.
+
+    Idempotent and non-destructive: a key already present in the cost map
+    (e.g. once litellm ships it) is left untouched. Registers both the bare
+    (``deepseek-v4-flash``) and prefixed (``deepseek/deepseek-v4-flash``)
+    spellings since ``backfill_cache_pricing`` tries both.
+    """
+    added = []
+    for bare, rates in _DEEPSEEK_V4_RATES.items():
+        for key in (bare, f"deepseek/{bare}"):
+            if key in litellm.model_cost:
+                continue
+            entry: dict[str, object] = dict(rates)
+            entry["litellm_provider"] = "deepseek"
+            entry["mode"] = "chat"
+            litellm.model_cost[key] = entry
+            added.append(key)
+    if added:
+        logger.info(
+            "Registered temporary DeepSeek V4 pricing shim",
+            extra={"models": added},
+        )

routstr/upstream/fireworks.py

@@ -12,6 +12,7 @@ class FireworksUpstreamProvider(BaseUpstreamProvider):
     provider_type = "fireworks"
     default_base_url = "https://api.fireworks.ai/inference/v1"
     platform_url = "https://app.fireworks.ai/settings/users/api-keys"
+    litellm_provider_prefix = "fireworks_ai/"
 
     def __init__(self, api_key: str, provider_fee: float = 1.01):
         super().__init__(

routstr/upstream/gemini.py

@@ -1,17 +1,13 @@
 from __future__ import annotations
 
-import json
-from collections.abc import AsyncGenerator
 from typing import TYPE_CHECKING, Any
 
-from fastapi import Request
-from fastapi.responses import Response, StreamingResponse
-
+from . import gemini_messages
 from .base import BaseUpstreamProvider
 from .clients.gemini import GeminiClient
 
 if TYPE_CHECKING:
-    from ..core.db import ApiKey, AsyncSession, UpstreamProviderRow
+    from ..core.db import UpstreamProviderRow
     from ..payment.models import Model
 
 from ..core.logging import get_logger
@@ -20,9 +16,19 @@ logger = get_logger(__name__)
 
 
 class GeminiUpstreamProvider(BaseUpstreamProvider):
+    """Gemini provider — proxies through Gemini's OpenAI-compat surface.
+
+    The chat-completions, embeddings, and models paths all flow through
+    :meth:`BaseUpstreamProvider.forward_request`; we only override
+    ``get_request_base_url`` to redirect to ``{base}/openai/...`` and
+    ``_dispatch_anthropic_messages`` to inject thought-signatures on the
+    /v1/messages path (see :mod:`gemini_messages` for that rationale).
+    """
+
     provider_type = "gemini"
     default_base_url = "https://generativelanguage.googleapis.com/v1beta"
     platform_url = "https://aistudio.google.com/app/apikey"
+    litellm_provider_prefix = "gemini/"
 
     def __init__(
         self,
@@ -39,7 +45,7 @@ class GeminiUpstreamProvider(BaseUpstreamProvider):
 
     @property
     def client(self) -> GeminiClient:
-        """Get or create the Gemini API client."""
+        """Get or create the Gemini API client (used for the models listing)."""
         if self._client is None:
             self._client = GeminiClient(api_key=self.api_key)
         return self._client
@@ -65,248 +71,66 @@ class GeminiUpstreamProvider(BaseUpstreamProvider):
         }
 
     def transform_model_name(self, model_id: str) -> str:
-        return model_id.removeprefix("gemini/")
+        """Reduce a routstr model id to the bare upstream Gemini name.
 
-    async def forward_request(
+        Gemini's OpenAI-compat surface expects the literal model id
+        (e.g. ``gemini-3.1-flash-lite-preview``) — no ``gemini/`` provider
+        prefix and no ``google/`` vendor sub-prefix. Take the last path
+        segment so we tolerate any of:
+
+            ``gemini-2.0-flash``
+            ``gemini/gemini-2.0-flash``
+            ``gemini/google/gemini-3.1-flash-lite-preview``
+        """
+        return model_id.rsplit("/", 1)[-1]
+
+    @property
+    def compat_base_url(self) -> str:
+        """Gemini's OpenAI-compat surface, regardless of what's stored.
+
+        Stored ``base_url`` may be ``.../v1beta`` (the native Gemini API
+        root) or ``.../v1beta/openai`` (already pointed at the compat
+        surface). Normalize to the latter.
+        """
+        return self.base_url.rstrip("/").removesuffix("/openai") + "/openai"
+
+    def get_request_base_url(
+        self, path: str, model_obj: "Model | None" = None
+    ) -> str:
+        """Route every proxied request to the OpenAI-compat surface.
+
+        Required because the stored ``base_url`` typically points at the
+        native Gemini API (``/v1beta``), but :meth:`forward_request`
+        forwards OpenAI-shaped paths (``/chat/completions``,
+        ``/embeddings``, ``/models``) which only exist under the
+        ``/openai`` subtree.
+        """
+        return self.compat_base_url
+
+    async def _dispatch_anthropic_messages(
         self,
-        request: Request,
-        path: str,
-        headers: dict,
         request_body: bytes | None,
-        key: ApiKey,
-        max_cost_for_model: int,
-        session: AsyncSession,
-        model_obj: Model,
-    ) -> Response | StreamingResponse:
-        # Remove provider prefix from model ID for Gemini API
-        if "/" in model_obj.id:
-            model_obj.id = model_obj.id.split("/", 1)[1]
+        model_obj: "Model",
+        *,
+        log_extra: dict[str, Any] | None = None,
+    ) -> tuple[bool, Any, str | None]:
+        """Dispatch /v1/messages via the gemini-specific httpx path.
 
-        if not path.startswith("chat/completions"):
-            return await super().forward_request(
-                request,
-                path,
-                headers,
-                request_body,
-                key,
-                max_cost_for_model,
-                session,
-                model_obj,
-            )
-
-        if not request_body:
-            return await super().forward_request(
-                request,
-                path,
-                headers,
-                request_body,
-                key,
-                max_cost_for_model,
-                session,
-                model_obj,
-            )
-
-        try:
-            openai_data = json.loads(request_body)
-            messages = openai_data.get("messages", [])
-            temperature = openai_data.get("temperature")
-            max_tokens = openai_data.get("max_tokens")
-            top_p = openai_data.get("top_p")
-            is_streaming = openai_data.get("stream", False)
-
-            logger.info(
-                "Processing Gemini request with client abstraction",
-                extra={
-                    "model": model_obj.id,
-                    "is_streaming": is_streaming,
-                    "message_count": len(messages),
-                    "key_hash": key.hashed_key[:8] + "...",
-                },
-            )
-
-            if is_streaming:
-                final_usage_data: dict | None = None
-
-                def usage_callback(usage_data: dict[str, Any]) -> None:
-                    """Callback to capture usage data during streaming"""
-                    nonlocal final_usage_data
-                    final_usage_data = usage_data
-
-                async def completion_callback(
-                    model: str, usage_data: dict[str, Any] | None
-                ) -> None:
-                    """Callback to handle payment when streaming completes"""
-                    nonlocal final_usage_data
-                    if usage_data:
-                        final_usage_data = usage_data
-
-                    payment_data = {
-                        "model": model,
-                        "usage": final_usage_data,
-                    }
-
-                    from ..auth import adjust_payment_for_tokens
-                    from ..core.db import create_session
-
-                    async with create_session() as new_session:
-                        fresh_key = await new_session.get(key.__class__, key.hashed_key)
-                        if fresh_key:
-                            try:
-                                cost_data = await adjust_payment_for_tokens(
-                                    fresh_key,
-                                    payment_data,
-                                    new_session,
-                                    max_cost_for_model,
-                                )
-
-                                logger.info(
-                                    "Gemini streaming payment finalized",
-                                    extra={
-                                        "cost_data": cost_data,
-                                        "usage_data": final_usage_data,
-                                        "key_hash": key.hashed_key[:8] + "...",
-                                    },
-                                )
-                            except Exception as cost_error:
-                                logger.error(
-                                    "Error finalizing Gemini streaming payment",
-                                    extra={
-                                        "error": str(cost_error),
-                                        "key_hash": key.hashed_key[:8] + "...",
-                                    },
-                                )
-
-                response_generator = self.client.generate_content_stream(
-                    model=model_obj.id,
-                    messages=messages,
-                    temperature=temperature,
-                    max_tokens=max_tokens,
-                    top_p=top_p,
-                    usage_callback=usage_callback,
-                    completion_callback=completion_callback,
-                )
-
-                async def stream_with_cost() -> AsyncGenerator[bytes, None]:
-                    payment_finalized = False
-
-                    async def finalize_payment() -> None:
-                        nonlocal payment_finalized
-                        if payment_finalized:
-                            return
-                        from ..auth import adjust_payment_for_tokens
-                        from ..core.db import create_session
-
-                        async with create_session() as new_session:
-                            fresh_key = await new_session.get(
-                                key.__class__, key.hashed_key
-                            )
-                            if fresh_key:
-                                try:
-                                    await adjust_payment_for_tokens(
-                                        fresh_key,
-                                        {
-                                            "model": model_obj.id,
-                                            "usage": final_usage_data,
-                                        },
-                                        new_session,
-                                        max_cost_for_model,
-                                    )
-                                    payment_finalized = True
-                                except Exception as cost_error:
-                                    logger.error(
-                                        "Error finalizing Gemini streaming payment in fallback",
-                                        extra={
-                                            "error": str(cost_error),
-                                            "key_hash": key.hashed_key[:8] + "...",
-                                        },
-                                    )
-
-                    try:
-                        async for chunk in response_generator:
-                            sse_data = f"data: {json.dumps(chunk)}\n\n"
-                            yield sse_data.encode()
-                    except Exception as e:
-                        logger.error(
-                            "Error in Gemini streaming response",
-                            extra={
-                                "error": str(e),
-                                "error_type": type(e).__name__,
-                                "key_hash": key.hashed_key[:8] + "...",
-                            },
-                        )
-                        raise
-                    finally:
-                        if not payment_finalized:
-                            await finalize_payment()
-
-                return StreamingResponse(
-                    stream_with_cost(),
-                    media_type="text/event-stream",
-                    headers={"Cache-Control": "no-cache", "Connection": "keep-alive"},
-                )
-
-            else:
-                openai_format_response = await self.client.generate_content(
-                    model=model_obj.id,
-                    messages=messages,
-                    temperature=temperature,
-                    max_tokens=max_tokens,
-                    top_p=top_p,
-                )
-
-                from ..auth import adjust_payment_for_tokens
-
-                cost_data = await adjust_payment_for_tokens(
-                    key, openai_format_response, session, max_cost_for_model
-                )
-                await session.refresh(key)
-                remaining_balance_msats = key.balance
-                openai_format_response["cost"] = cost_data
-                openai_format_response["cost"]["sats_cost"] = (
-                    cost_data.get("total_msats", 0) // 1000
-                )
-                openai_format_response["cost"]["remaining_balance_msats"] = (
-                    remaining_balance_msats
-                )
-
-                logger.info(
-                    "Gemini non-streaming payment completed",
-                    extra={
-                        "cost_data": cost_data,
-                        "model": model_obj.id,
-                        "key_hash": key.hashed_key[:8] + "...",
-                    },
-                )
-
-                return Response(
-                    content=json.dumps(openai_format_response),
-                    media_type="application/json",
-                    headers={"Cache-Control": "no-cache"},
-                )
-
-        except Exception as e:
-            logger.error(
-                "Error in Gemini forward_request",
-                extra={
-                    "error": str(e),
-                    "error_type": type(e).__name__,
-                    "path": path,
-                    "key_hash": key.hashed_key[:8] + "...",
-                },
-            )
-            return await super().forward_request(
-                request,
-                path,
-                headers,
-                request_body,
-                key,
-                max_cost_for_model,
-                session,
-                model_obj,
-            )
+        See :mod:`routstr.upstream.gemini_messages` for the full rationale
+        (thought-signature injection, why litellm + the openai SDK can't
+        carry the required ``extra_content`` field).
+        """
+        return await gemini_messages.dispatch_gemini_messages(
+            request_body=request_body,
+            model_obj=model_obj,
+            base_url=self.compat_base_url,
+            api_key=self.api_key,
+            transform_model_name=self.transform_model_name,
+            log_extra=log_extra,
+        )
 
     async def _fetch_provider_models(self) -> dict:
-        """Fetch models from Gemini API."""
+        """Fetch models from Gemini API via the OpenAI-compat client."""
         try:
             models_data = await self.client.list_models()
 

routstr/upstream/gemini_messages.py

@@ -0,0 +1,480 @@
+"""Custom /v1/messages dispatcher for Gemini's OpenAI-compat endpoint.
+
+Why this exists
+---------------
+
+Gemini 2.5 / 3 thinking models reject inbound ``functionCall`` parts that
+lack a ``thought_signature`` field once any prior turn in the conversation
+contains a function call. Anthropic-Messages clients (Claude Code etc.)
+have no concept of thought signatures, so multi-turn tool conversations
+fail with::
+
+    Function call is missing a thought_signature in functionCall parts.
+
+Google's published escape hatch (https://ai.google.dev/gemini-api/docs/
+thought-signatures, FAQ #1) is the dummy signature
+``"skip_thought_signature_validator"`` placed at
+``tool_calls[i].extra_content.google.thought_signature`` for every tool
+call in the request. The hatch is documented specifically for
+"transferring a trace from a different model that does not include thought
+signatures" — exactly our case.
+
+Why we can't reach the wire via litellm
+---------------------------------------
+
+``litellm.anthropic.messages.acreate`` flows through the openai SDK, whose
+pydantic ``ChatCompletionMessageToolCall`` model silently drops unknown
+fields like ``extra_content``. Litellm has no openai-compat translator
+that emits ``extra_content.google.thought_signature``. So we bypass both
+litellm and the openai SDK at the transport layer.
+
+Pipeline
+--------
+
+1. Translate Anthropic body → OpenAI body via litellm's
+   ``AnthropicAdapter`` (the same translator
+   ``litellm.anthropic.messages.acreate`` uses internally).
+2. Inject ``extra_content.google.thought_signature`` on every
+   ``tool_calls[]`` entry.
+3. Set ``reasoning_effort="none"`` to disable Gemini's thinking pass.
+4. POST directly to ``{base_url}/chat/completions`` with ``stream=true``
+   via ``httpx`` (preserves arbitrary fields verbatim).
+5. Translate OpenAI streaming chunks → Anthropic SSE events.
+"""
+
+from __future__ import annotations
+
+import json
+import uuid
+from collections.abc import AsyncGenerator, AsyncIterator
+from typing import Any, Callable
+
+import httpx
+
+from ..core import get_logger
+from ..core.exceptions import UpstreamError
+from ..payment.models import Model
+from .messages_dispatch import (
+    ANTHROPIC_ONLY_FIELDS,
+    aggregate_anthropic_events_to_message,
+)
+
+logger = get_logger(__name__)
+
+DUMMY_THOUGHT_SIGNATURE = "skip_thought_signature_validator"
+
+# Mapping: OpenAI finish_reason → Anthropic stop_reason
+_FINISH_TO_STOP = {
+    "stop": "end_turn",
+    "length": "max_tokens",
+    "tool_calls": "tool_use",
+    "function_call": "tool_use",
+    "content_filter": "refusal",
+}
+
+
+def inject_thought_signatures(messages: list[dict]) -> None:
+    """Add ``extra_content.google.thought_signature`` to every tool_call.
+
+    Mutates ``messages`` in place. Idempotent: existing signatures are not
+    overwritten.
+    """
+    for msg in messages:
+        tool_calls = msg.get("tool_calls")
+        if not isinstance(tool_calls, list):
+            continue
+        for tc in tool_calls:
+            if not isinstance(tc, dict):
+                continue
+            extra = tc.get("extra_content")
+            if not isinstance(extra, dict):
+                extra = tc["extra_content"] = {}
+            google_cfg = extra.get("google")
+            if not isinstance(google_cfg, dict):
+                google_cfg = extra["google"] = {}
+            google_cfg.setdefault("thought_signature", DUMMY_THOUGHT_SIGNATURE)
+
+
+def _translate_anthropic_to_openai(body: dict, model: str) -> dict:
+    """Use litellm's translator to convert an Anthropic /messages body to
+    OpenAI /chat/completions kwargs.
+
+    Imported lazily because the litellm internal path is heavy and not
+    needed for any other code path in routstr.
+    """
+    from litellm.llms.anthropic.experimental_pass_through.adapters.transformation import (  # noqa: E501
+        AnthropicAdapter,
+    )
+
+    kwargs = {"model": model, **body}
+    translated = AnthropicAdapter().translate_completion_input_params(kwargs)
+    if translated is None:
+        raise UpstreamError(
+            "Failed to translate Anthropic body to OpenAI format",
+            status_code=500,
+        )
+    return dict(translated)
+
+
+def _sse_event(event_type: str, payload: dict) -> bytes:
+    return f"event: {event_type}\ndata: {json.dumps(payload)}\n\n".encode()
+
+
+async def _openai_chunks_to_anthropic_events(
+    line_iter: AsyncIterator[str], requested_model: str | None
+) -> AsyncGenerator[bytes, None]:
+    """Translate an OpenAI chat-completions SSE byte stream into the
+    Anthropic-Messages SSE event sequence.
+
+    Maintains per-chunk state across:
+
+    * one optional text content block (lazy-opened on first text delta)
+    * any number of tool_use blocks indexed by openai's ``delta.tool_calls[].index``
+    * final ``stop_reason`` / ``usage`` carried out via ``message_delta`` /
+      ``message_stop``
+    """
+    msg_id = f"msg_{uuid.uuid4().hex[:24]}"
+    started = False
+    text_block_idx: int | None = None
+    tool_block_indices: dict[int, int] = {}
+    next_block_idx = 0
+    final_finish_reason: str | None = None
+    final_usage: dict[str, int] = {"input_tokens": 0, "output_tokens": 0}
+
+    def open_text_block() -> bytes:
+        nonlocal text_block_idx, next_block_idx
+        text_block_idx = next_block_idx
+        next_block_idx += 1
+        return _sse_event(
+            "content_block_start",
+            {
+                "type": "content_block_start",
+                "index": text_block_idx,
+                "content_block": {"type": "text", "text": ""},
+            },
+        )
+
+    def open_tool_block(delta_idx: int, tc: dict) -> bytes:
+        nonlocal next_block_idx
+        block_idx = next_block_idx
+        next_block_idx += 1
+        tool_block_indices[delta_idx] = block_idx
+        fn = tc.get("function") or {}
+        return _sse_event(
+            "content_block_start",
+            {
+                "type": "content_block_start",
+                "index": block_idx,
+                "content_block": {
+                    "type": "tool_use",
+                    "id": tc.get("id") or f"toolu_{uuid.uuid4().hex[:24]}",
+                    "name": fn.get("name") or "",
+                    "input": {},
+                },
+            },
+        )
+
+    def close_block(idx: int) -> bytes:
+        return _sse_event(
+            "content_block_stop",
+            {"type": "content_block_stop", "index": idx},
+        )
+
+    async for raw_line in line_iter:
+        line = raw_line.strip()
+        if not line:
+            continue
+        if not line.startswith("data:"):
+            continue
+        payload = line[5:].lstrip()
+        if not payload or payload == "[DONE]":
+            continue
+        try:
+            chunk = json.loads(payload)
+        except json.JSONDecodeError:
+            continue
+        if not isinstance(chunk, dict):
+            continue
+
+        if not started:
+            started = True
+            yield _sse_event(
+                "message_start",
+                {
+                    "type": "message_start",
+                    "message": {
+                        "id": chunk.get("id") or msg_id,
+                        "type": "message",
+                        "role": "assistant",
+                        "model": requested_model or chunk.get("model") or "",
+                        "content": [],
+                        "stop_reason": None,
+                        "stop_sequence": None,
+                        "usage": {
+                            "input_tokens": 0,
+                            "output_tokens": 0,
+                        },
+                    },
+                },
+            )
+
+        usage = chunk.get("usage")
+        if isinstance(usage, dict):
+            in_tok = usage.get("prompt_tokens") or usage.get("input_tokens") or 0
+            out_tok = usage.get("completion_tokens") or usage.get("output_tokens") or 0
+            if in_tok:
+                final_usage["input_tokens"] = int(in_tok)
+            if out_tok:
+                final_usage["output_tokens"] = int(out_tok)
+
+        choices = chunk.get("choices") or []
+        if not choices:
+            continue
+        choice = choices[0] if isinstance(choices[0], dict) else {}
+        delta = choice.get("delta") or {}
+        if not isinstance(delta, dict):
+            delta = {}
+
+        # Text delta
+        text = delta.get("content")
+        if isinstance(text, str) and text:
+            if text_block_idx is None:
+                yield open_text_block()
+            yield _sse_event(
+                "content_block_delta",
+                {
+                    "type": "content_block_delta",
+                    "index": text_block_idx,
+                    "delta": {"type": "text_delta", "text": text},
+                },
+            )
+
+        # Tool call deltas
+        tool_calls_delta = delta.get("tool_calls")
+        if isinstance(tool_calls_delta, list):
+            for tc in tool_calls_delta:
+                if not isinstance(tc, dict):
+                    continue
+                d_idx = int(tc.get("index") or 0)
+                if d_idx not in tool_block_indices:
+                    yield open_tool_block(d_idx, tc)
+                block_idx = tool_block_indices[d_idx]
+                fn = tc.get("function") or {}
+                args = fn.get("arguments")
+                if isinstance(args, str) and args:
+                    yield _sse_event(
+                        "content_block_delta",
+                        {
+                            "type": "content_block_delta",
+                            "index": block_idx,
+                            "delta": {
+                                "type": "input_json_delta",
+                                "partial_json": args,
+                            },
+                        },
+                    )
+
+        finish = choice.get("finish_reason")
+        if finish:
+            final_finish_reason = finish
+
+    # Close any open content blocks
+    if text_block_idx is not None:
+        yield close_block(text_block_idx)
+    for block_idx in tool_block_indices.values():
+        yield close_block(block_idx)
+
+    # message_delta with stop_reason and usage
+    stop_reason = _FINISH_TO_STOP.get(final_finish_reason or "", "end_turn")
+    yield _sse_event(
+        "message_delta",
+        {
+            "type": "message_delta",
+            "delta": {"stop_reason": stop_reason, "stop_sequence": None},
+            "usage": final_usage,
+        },
+    )
+    yield _sse_event("message_stop", {"type": "message_stop"})
+
+
+async def _post_and_stream(
+    base_url: str,
+    api_key: str,
+    payload: dict,
+    log_extra: dict[str, Any] | None,
+) -> tuple[httpx.AsyncClient, httpx.Response]:
+    """POST to upstream chat-completions and return (client, response) for
+    streaming. Caller is responsible for closing both."""
+    url = f"{base_url.rstrip('/')}/chat/completions"
+    client = httpx.AsyncClient(timeout=httpx.Timeout(120.0, read=120.0))
+    try:
+        request = client.build_request(
+            "POST",
+            url,
+            json=payload,
+            headers={
+                "Authorization": f"Bearer {api_key}",
+                "Content-Type": "application/json",
+                "Accept": "text/event-stream",
+            },
+        )
+        response = await client.send(request, stream=True)
+    except Exception as exc:
+        await client.aclose()
+        logger.error(
+            "Gemini messages dispatch HTTP error",
+            extra={"error": str(exc), "url": url, **(log_extra or {})},
+        )
+        raise UpstreamError(
+            f"Failed to reach Gemini upstream: {exc}", status_code=502
+        ) from exc
+
+    if response.status_code >= 400:
+        try:
+            body_bytes = await response.aread()
+        finally:
+            await response.aclose()
+            await client.aclose()
+        body_text = body_bytes.decode("utf-8", errors="replace")
+        logger.error(
+            "Gemini messages dispatch upstream error",
+            extra={
+                "status_code": response.status_code,
+                "body": body_text[:1000],
+                "url": url,
+                **(log_extra or {}),
+            },
+        )
+        raise UpstreamError(
+            f"Upstream error via gemini compat: {body_text}",
+            status_code=response.status_code,
+        )
+
+    return client, response
+
+
+async def dispatch_gemini_messages(
+    *,
+    request_body: bytes | None,
+    model_obj: Model,
+    base_url: str,
+    api_key: str,
+    transform_model_name: Callable[[str], str],
+    log_extra: dict[str, Any] | None = None,
+) -> tuple[bool, Any, str | None]:
+    """Dispatch a /v1/messages request to Gemini's OpenAI-compat endpoint
+    with thought-signature injection.
+
+    Returns ``(client_stream, result, requested_model)`` where ``result``
+    is either an ``AsyncIterator[bytes]`` of Anthropic-format SSE events
+    (for streaming clients) or an Anthropic Message dict (after the caller
+    aggregates).
+    """
+    if not request_body:
+        raise UpstreamError(
+            "Missing request body for /v1/messages", status_code=400
+        )
+
+    try:
+        body: dict = json.loads(request_body)
+    except json.JSONDecodeError as exc:
+        raise UpstreamError(
+            f"Invalid JSON in /v1/messages body: {exc}", status_code=400
+        ) from exc
+
+    body.pop("model", None)
+    client_stream = bool(body.pop("stream", False))
+
+    # Anthropic-Messages-only fields that don't translate to OpenAI
+    # Chat Completions. litellm's translator passes through unknown
+    # top-level fields verbatim and Gemini's compat surface 400s on
+    # unknown names like ``context_management`` / ``output_config``.
+    dropped: dict[str, Any] = {}
+    for field in ANTHROPIC_ONLY_FIELDS:
+        if field in body:
+            dropped[field] = body.pop(field)
+    if dropped:
+        logger.debug(
+            "Dropped anthropic-only fields before gemini compat dispatch",
+            extra={"dropped_keys": sorted(dropped.keys())},
+        )
+
+    requested_model = (
+        (model_obj.forwarded_model_id or model_obj.id) if model_obj else None
+    )
+    upstream_model = transform_model_name(model_obj.id)
+
+    openai_kwargs = _translate_anthropic_to_openai(body, upstream_model)
+
+    messages = openai_kwargs.get("messages") or []
+    if isinstance(messages, list):
+        inject_thought_signatures(messages)
+
+    # Disable Gemini's thinking pass; the dummy signature already lifts
+    # validation, but skipping thinking entirely avoids degraded model
+    # output and keeps tool-calling deterministic.
+    openai_kwargs.setdefault("reasoning_effort", "none")
+    openai_kwargs["stream"] = True
+    openai_kwargs["model"] = upstream_model
+    # OpenAI-compat backends (including Gemini's) only emit a final
+    # ``usage`` chunk when the request opts in via this flag. Without it
+    # the cost-calculation pipeline can't read real token counts and
+    # falls back to MaxCostData billing.
+    existing_stream_options = openai_kwargs.get("stream_options")
+    merged_stream_options = (
+        dict(existing_stream_options)
+        if isinstance(existing_stream_options, dict)
+        else {}
+    )
+    merged_stream_options.setdefault("include_usage", True)
+    openai_kwargs["stream_options"] = merged_stream_options
+
+    logger.info(
+        "Dispatching /v1/messages via gemini compat (httpx)",
+        extra={
+            "model": upstream_model,
+            "client_stream": client_stream,
+            "messages_with_tool_calls": sum(
+                1 for m in messages if isinstance(m, dict) and m.get("tool_calls")
+            ),
+            **(log_extra or {}),
+        },
+    )
+
+    http_client, response = await _post_and_stream(
+        base_url, api_key, openai_kwargs, log_extra
+    )
+
+    async def line_iter() -> AsyncGenerator[str, None]:
+        try:
+            async for line in response.aiter_lines():
+                yield line
+        finally:
+            await response.aclose()
+            await http_client.aclose()
+
+    anthropic_event_iter = _openai_chunks_to_anthropic_events(
+        line_iter(), requested_model
+    )
+
+    if not client_stream:
+        # Aggregate the Anthropic SSE byte stream into a single Message dict
+        # so the rest of the pipeline (cost calc, metadata injection,
+        # response building) can treat it identically to a non-streaming
+        # litellm response.
+        try:
+            aggregated = await aggregate_anthropic_events_to_message(
+                anthropic_event_iter
+            )
+        except Exception as exc:
+            logger.error(
+                "Failed to aggregate Gemini compat events into message",
+                extra={"error": str(exc), **(log_extra or {})},
+            )
+            raise UpstreamError(
+                f"Failed to aggregate upstream stream: {exc}",
+                status_code=502,
+            ) from exc
+        return client_stream, aggregated, requested_model
+
+    return client_stream, anthropic_event_iter, requested_model

routstr/upstream/groq.py

@@ -12,6 +12,7 @@ class GroqUpstreamProvider(BaseUpstreamProvider):
     provider_type = "groq"
     default_base_url = "https://api.groq.com/openai/v1"
     platform_url = "https://console.groq.com/keys"
+    litellm_provider_prefix = "groq/"
 
     def __init__(self, api_key: str, provider_fee: float = 1.01):
         super().__init__(

routstr/upstream/helpers.py

@@ -3,7 +3,7 @@ from __future__ import annotations
 import asyncio
 import os
 import re
-from typing import TYPE_CHECKING
+from typing import TYPE_CHECKING, Callable
 
 if TYPE_CHECKING:
     from ..core.settings import Settings
@@ -122,12 +122,16 @@ async def get_all_models_with_overrides(
 
 
 async def refresh_upstreams_models_periodically(
-    upstreams: list[BaseUpstreamProvider],
+    upstreams_provider: (
+        Callable[[], list[BaseUpstreamProvider]] | list[BaseUpstreamProvider]
+    ),
 ) -> None:
     """Background task to periodically refresh models cache for all providers.
 
     Args:
-        upstreams: List of upstream provider instances
+        upstreams_provider: Either a callable returning the live upstream list
+            (preferred — picks up providers added/changed via reinitialize_upstreams),
+            or a static list (legacy, will go stale after reinitialize_upstreams).
     """
     import asyncio
     import random
@@ -139,9 +143,14 @@ async def refresh_upstreams_models_periodically(
         logger.info("Provider models refresh disabled (interval <= 0)")
         return
 
+    def _resolve_upstreams() -> list[BaseUpstreamProvider]:
+        if callable(upstreams_provider):
+            return upstreams_provider()
+        return upstreams_provider
+
     while True:
         try:
-            for upstream in upstreams:
+            for upstream in _resolve_upstreams():
                 try:
                     await upstream.refresh_models_cache()
                 except Exception as e:
@@ -188,13 +197,14 @@ async def init_upstreams() -> list[BaseUpstreamProvider]:
         existing_providers = result.all()
 
         if not existing_providers:
-            logger.info(
-                "No upstream providers found in database, seeding from settings"
-            )
             await _seed_providers_from_settings(session, settings)
             await session.commit()
             result = await session.exec(select(UpstreamProviderRow))
             existing_providers = result.all()
+            if existing_providers:
+                logger.info(
+                    f"Seeded {len(existing_providers)} upstream providers from settings"
+                )
 
         async def _init_single_provider(
             provider_row: UpstreamProviderRow,
@@ -254,6 +264,7 @@ async def _seed_providers_from_settings(
         ("PERPLEXITY_API_KEY", "perplexity", None, None),
         ("FIREWORKS_API_KEY", "fireworks", None, None),
         ("XAI_API_KEY", "xai", None, None),
+        ("TINFOIL_API_KEY", "tinfoil", None, None),
     ]
 
     for env_key, provider_type, _, _ in env_mappings:

routstr/upstream/litellm_routing.py

@@ -0,0 +1,162 @@
+"""Map an upstream `base_url` to the correct litellm provider prefix.
+
+Used by `BaseUpstreamProvider.get_litellm_provider_prefix` so that custom /
+generic provider rows (which inherit the base class default) get routed to
+the right litellm backend instead of falling back to `openai/`.
+
+The table is compiled from litellm 1.74's
+`litellm/litellm_core_utils/get_llm_provider_logic.py` (the
+`openai_compatible_endpoints` table) plus the providers documented at
+https://docs.litellm.ai/docs/providers. Substring match is used so that
+URLs with paths, ports, regional subdomains, etc. all resolve correctly.
+
+Order matters: more specific needles must appear before more generic ones
+(e.g. ``openai.azure.com`` before ``api.openai.com``).
+"""
+
+from __future__ import annotations
+
+import os
+from urllib.parse import urlsplit
+
+import litellm
+
+DEFAULT_PREFIX = "openai/"
+
+LITELLM_HOST_PREFIX_MAP: tuple[tuple[str, str], ...] = (
+    # Azure must win over api.openai.com because the host ends with
+    # `openai.azure.com` and we don't want it picked up as plain OpenAI.
+    ("openai.azure.com", "azure/"),
+    # Google
+    ("generativelanguage.googleapis.com", "gemini/"),
+    ("aiplatform.googleapis.com", "vertex_ai/"),
+    # First-class providers with native litellm prefixes
+    ("api.openai.com", "openai/"),
+    ("api.anthropic.com", "anthropic/"),
+    ("api.groq.com", "groq/"),
+    ("api.fireworks.ai", "fireworks_ai/"),
+    ("api.x.ai", "xai/"),
+    ("api.perplexity.ai", "perplexity/"),
+    ("openrouter.ai", "openrouter/"),
+    ("api.deepseek.com", "deepseek/"),
+    ("api.together.xyz", "together_ai/"),
+    ("codestral.mistral.ai", "codestral/"),
+    ("api.mistral.ai", "mistral/"),
+    ("api.cohere.com", "cohere_chat/"),
+    ("api.cohere.ai", "cohere_chat/"),
+    ("api.deepinfra.com", "deepinfra/"),
+    ("api.endpoints.anyscale.com", "anyscale/"),
+    ("api.cerebras.ai", "cerebras/"),
+    ("inference.baseten.co", "baseten/"),
+    ("api.sambanova.ai", "sambanova/"),
+    ("api.ai21.com", "ai21_chat/"),
+    ("api.friendli.ai", "friendliai/"),
+    ("api.galadriel.com", "galadriel/"),
+    ("api.llama.com", "meta_llama/"),
+    ("api.featherless.ai", "featherless_ai/"),
+    ("inference.api.nscale.com", "nscale/"),
+    ("dashscope-intl.aliyuncs.com", "dashscope/"),
+    ("api.moonshot.ai", "moonshot/"),
+    ("api.moonshot.cn", "moonshot/"),
+    ("api.minimax.io", "minimax/"),
+    ("api.minimaxi.com", "minimax/"),
+    ("platform.publicai.co", "publicai/"),
+    ("api.synthetic.new", "synthetic/"),
+    ("api.stima.tech", "apertis/"),
+    ("nano-gpt.com", "nano-gpt/"),
+    ("api.poe.com", "poe/"),
+    ("llm.chutes.ai", "chutes/"),
+    ("api.v0.dev", "v0/"),
+    ("api.lambda.ai", "lambda_ai/"),
+    ("api.hyperbolic.xyz", "hyperbolic/"),
+    ("ai-gateway.vercel.sh", "vercel_ai_gateway/"),
+    ("api.inference.wandb.ai", "wandb/"),
+    ("integrate.api.nvidia.com", "nvidia_nim/"),
+    ("api.studio.nebius.com", "nebius/"),
+    ("api.novita.ai", "novita/"),
+    ("ark.cn-beijing.volces.com", "volcengine/"),
+    ("api.voyageai.com", "voyage/"),
+    ("api.jina.ai", "jina_ai/"),
+    ("api.aimlapi.com", "aiml/"),
+    ("api.snowflakecomputing.com", "snowflake/"),
+    ("databricks.com", "databricks/"),
+    ("huggingface.co", "huggingface/"),
+)
+
+# Substrings that indicate an Ollama deployment regardless of port/scheme.
+OLLAMA_HOST_HINTS: tuple[str, ...] = (
+    "localhost:11434",
+    "127.0.0.1:11434",
+    "ollama",
+)
+
+
+def detect_litellm_prefix(
+    base_url: str | None, default: str = DEFAULT_PREFIX
+) -> str:
+    """Return the litellm provider prefix (`"<provider>/"`) for `base_url`.
+
+    Falls back to `default` when the host doesn't match any known provider.
+    The default is `openai/` because every unmatched OpenAI-compatible
+    server is, by definition, an OpenAI-compatible server.
+    """
+    if not base_url:
+        return default
+
+    parsed = urlsplit(base_url)
+    host = parsed.netloc.lower() or base_url.lower()
+
+    for needle, prefix in LITELLM_HOST_PREFIX_MAP:
+        if needle in host:
+            return prefix
+
+    if any(hint in host for hint in OLLAMA_HOST_HINTS):
+        return "ollama_chat/"
+
+    return default
+
+
+_configured = False
+
+
+def configure_litellm() -> None:
+    """Apply litellm global settings used by the messages-dispatch path.
+
+    Idempotent: safe to call from both app startup and module-level
+    initializers without side effects on the second invocation.
+
+    Settings applied:
+
+    * ``LITELLM_DEBUG=1`` enables litellm's verbose debug logger.
+    * Forces the Anthropic-messages adapter to call OpenAI Chat Completions
+      (POST ``/chat/completions``) instead of the Responses API (POST
+      ``/responses``) for ``openai/``-prefixed providers. OpenAI-compatible
+      upstreams like Google's generativelanguage compat endpoint expose
+      ``/chat/completions`` but not ``/responses``, which would 404. Set
+      ``LITELLM_USE_RESPONSES_API_FOR_ANTHROPIC_MESSAGES=1`` to opt out.
+    * Silently drops Anthropic-Messages-only parameters (``thinking``,
+      ``cache_control``, ``context_management``, ...) when translating to
+      providers that don't accept them, instead of raising
+      ``UnsupportedParamsError``. Set ``LITELLM_STRICT_PARAMS=1`` to opt
+      out.
+    """
+    global _configured
+    if _configured:
+        return
+
+    if os.getenv("LITELLM_DEBUG") == "1":
+        try:
+            litellm._turn_on_debug()  # type: ignore[no-untyped-call]
+        except Exception:
+            pass
+
+    if os.getenv("LITELLM_USE_RESPONSES_API_FOR_ANTHROPIC_MESSAGES") != "1":
+        try:
+            litellm.use_chat_completions_url_for_anthropic_messages = True
+        except Exception:
+            pass
+
+    if os.getenv("LITELLM_STRICT_PARAMS") != "1":
+        litellm.drop_params = True
+
+    _configured = True

routstr/upstream/messages_dispatch.py

@@ -0,0 +1,575 @@
+"""Pure helpers for translating ``/v1/messages`` to upstream chat completions
+via litellm.
+
+This module owns the litellm/Anthropic-Messages translation layer:
+
+* SSE parsing (``parse_sse_blocks``, ``events_from_chunk``)
+* Payload coercion (``coerce_litellm_payload``)
+* Stream aggregation (``aggregate_anthropic_events_to_message``) — drains
+  a streamed Anthropic event sequence into a single Message dict
+* Per-event annotation for streaming (``annotate_event``,
+  ``stream_annotated_events``) — handles the model-rewrite + token-tally
+  bookkeeping shared by the bearer-key and x-cashu streaming paths
+* The dispatch entry point (``dispatch_anthropic_messages``)
+* Refund math (``compute_refund``)
+
+Nothing in here touches ``BaseUpstreamProvider``; the thin instance methods
+on the provider class forward to these functions and only retain logic that
+genuinely needs ``self`` (cost adjustment, metadata injection, refund
+sending).
+"""
+
+from __future__ import annotations
+
+import json
+from collections.abc import AsyncGenerator, AsyncIterator
+from typing import Any, Callable, NamedTuple, cast
+
+import litellm
+
+from ..core import get_logger
+from ..core.exceptions import UpstreamError
+from ..core.redaction import redact_org_ids
+from ..payment.models import Model
+from .rate_limit import classify_rate_limit
+
+logger = get_logger(__name__)
+
+# Anthropic-Messages-only fields that don't translate to OpenAI
+# Chat Completions. ``litellm.drop_params`` only filters *known*
+# unsupported params; these newer/extension fields get passed through
+# verbatim and the upstream rejects them with a 400. Pop them here so the
+# request reaches the upstream cleanly.
+ANTHROPIC_ONLY_FIELDS: tuple[str, ...] = (
+    "thinking",
+    "cache_control",
+    "context_management",
+    "output_config",
+    "mcp_servers",
+    "service_tier",
+    "anthropic_version",
+    "anthropic_beta",
+)
+
+
+def coerce_litellm_payload(payload: object) -> dict:
+    """Convert a litellm event into a plain dict.
+
+    Non-streaming responses come back as Anthropic-shaped pydantic models
+    or dicts. Streaming may yield raw bytes/str (SSE-encoded); those go
+    through ``events_from_chunk`` instead, not here.
+    """
+    if isinstance(payload, dict):
+        return dict(payload)
+    if hasattr(payload, "model_dump"):
+        return cast(dict, payload.model_dump())
+    raise TypeError(f"Cannot coerce {type(payload).__name__} to dict")
+
+
+def parse_sse_blocks(buffer: bytes) -> tuple[list[dict], bytes]:
+    """Parse complete SSE event blocks out of a byte buffer.
+
+    Returns (events, remaining_buffer). Events are JSON objects parsed from
+    one or more ``data:`` lines per block. Comments, blank lines, and
+    ``[DONE]`` sentinels are ignored. A trailing partial block is preserved
+    in remaining_buffer.
+    """
+    events: list[dict] = []
+    while True:
+        sep = buffer.find(b"\n\n")
+        if sep < 0:
+            sep_rn = buffer.find(b"\r\n\r\n")
+            if sep_rn < 0:
+                break
+            block = buffer[:sep_rn]
+            buffer = buffer[sep_rn + 4 :]
+        else:
+            block = buffer[:sep]
+            buffer = buffer[sep + 2 :]
+
+        data_lines: list[str] = []
+        for raw_line in block.replace(b"\r\n", b"\n").split(b"\n"):
+            line = raw_line.decode("utf-8", errors="replace")
+            if line.startswith(":"):
+                continue
+            if line.startswith("data:"):
+                data_lines.append(line[5:].lstrip())
+        if not data_lines:
+            continue
+        payload = "\n".join(data_lines).strip()
+        if not payload or payload == "[DONE]":
+            continue
+        try:
+            obj = json.loads(payload)
+        except json.JSONDecodeError:
+            continue
+        if isinstance(obj, dict):
+            events.append(obj)
+    return events, buffer
+
+
+def events_from_chunk(
+    chunk: object, sse_buffer: bytes
+) -> tuple[list[dict], bytes]:
+    """Normalize a stream chunk into one or more event dicts.
+
+    ``litellm.anthropic.messages.acreate(stream=True)`` yields raw SSE
+    bytes in practice; some adapters yield strings or typed events. Handle
+    all three.
+    """
+    if isinstance(chunk, (bytes, bytearray)):
+        sse_buffer += bytes(chunk)
+        events, sse_buffer = parse_sse_blocks(sse_buffer)
+        return events, sse_buffer
+    if isinstance(chunk, str):
+        sse_buffer += chunk.encode("utf-8")
+        events, sse_buffer = parse_sse_blocks(sse_buffer)
+        return events, sse_buffer
+    return [coerce_litellm_payload(chunk)], sse_buffer
+
+
+async def aggregate_anthropic_events_to_message(
+    iterator: AsyncIterator[Any],
+) -> dict:
+    """Drain an Anthropic-Messages event iterator into a single Message dict.
+
+    Produces the shape ``litellm.anthropic.messages.acreate(stream=False)``
+    would have returned. Used to transparently stream from upstream while
+    still returning a non-streaming response to the client. Lets us
+    sidestep upstream quirks (e.g. Fireworks rejects ``max_tokens > 4096``
+    unless ``stream=true``) without leaking that into client-visible
+    behavior.
+    """
+    sse_buffer = b""
+    message: dict = {}
+    blocks: list[dict] = []
+    partial_json: dict[int, str] = {}
+    final_stop_reason: str | None = None
+    final_stop_sequence: str | None = None
+    final_usage: dict[str, Any] = {}
+    final_model: str | None = None
+
+    async for chunk in iterator:
+        events, sse_buffer = events_from_chunk(chunk, sse_buffer)
+        for event in events:
+            etype = event.get("type")
+            if etype == "message_start":
+                raw = event.get("message") or {}
+                if isinstance(raw, dict):
+                    message = dict(raw)
+                    existing = message.get("content")
+                    blocks = list(existing) if isinstance(existing, list) else []
+                    usage = message.get("usage")
+                    if isinstance(usage, dict):
+                        final_usage = dict(usage)
+                    if isinstance(message.get("model"), str):
+                        final_model = message["model"]
+            elif etype == "content_block_start":
+                idx = int(event.get("index") or 0)
+                cb = event.get("content_block") or {}
+                cb_dict = dict(cb) if isinstance(cb, dict) else {}
+                while len(blocks) <= idx:
+                    blocks.append({})
+                blocks[idx] = cb_dict
+            elif etype == "content_block_delta":
+                idx = int(event.get("index") or 0)
+                if idx >= len(blocks):
+                    continue
+                delta = event.get("delta") or {}
+                if not isinstance(delta, dict):
+                    continue
+                dtype = delta.get("type")
+                block = blocks[idx]
+                if dtype == "text_delta":
+                    block["text"] = (block.get("text") or "") + (
+                        delta.get("text") or ""
+                    )
+                elif dtype == "input_json_delta":
+                    partial_json[idx] = partial_json.get(idx, "") + (
+                        delta.get("partial_json") or ""
+                    )
+                elif dtype == "thinking_delta":
+                    block["thinking"] = (block.get("thinking") or "") + (
+                        delta.get("thinking") or ""
+                    )
+                elif dtype == "signature_delta":
+                    block["signature"] = (block.get("signature") or "") + (
+                        delta.get("signature") or ""
+                    )
+            elif etype == "content_block_stop":
+                idx = int(event.get("index") or 0)
+                raw_json = partial_json.pop(idx, None)
+                if raw_json is not None and idx < len(blocks):
+                    try:
+                        blocks[idx]["input"] = (
+                            json.loads(raw_json) if raw_json else {}
+                        )
+                    except json.JSONDecodeError:
+                        blocks[idx]["input"] = raw_json
+            elif etype == "message_delta":
+                delta = event.get("delta") or {}
+                if isinstance(delta, dict):
+                    if "stop_reason" in delta:
+                        final_stop_reason = delta.get("stop_reason")
+                    if "stop_sequence" in delta:
+                        final_stop_sequence = delta.get("stop_sequence")
+                usage = event.get("usage")
+                if isinstance(usage, dict):
+                    final_usage.update(usage)
+            # message_stop: nothing to merge
+
+    if not message:
+        # Upstream returned no message_start; expose what we can so the
+        # client at least sees the assembled content.
+        message = {
+            "id": "",
+            "type": "message",
+            "role": "assistant",
+            "content": [],
+        }
+
+    message["content"] = blocks
+    if final_model and not message.get("model"):
+        message["model"] = final_model
+    if final_stop_reason is not None:
+        message["stop_reason"] = final_stop_reason
+    if final_stop_sequence is not None:
+        message["stop_sequence"] = final_stop_sequence
+    if final_usage:
+        existing_usage = message.get("usage")
+        merged = dict(existing_usage) if isinstance(existing_usage, dict) else {}
+        merged.update(final_usage)
+        message["usage"] = merged
+    return message
+
+
+class AnnotatedEvent(NamedTuple):
+    """One Anthropic SSE event after model-rewrite + token-tally bookkeeping.
+
+    ``sse_bytes`` is the wire-ready ``event:`` / ``data:`` block; the two
+    streaming paths in ``BaseUpstreamProvider`` consume ``sse_bytes`` plus
+    the tallies and only differ in whether they stream live or buffer
+    first.
+
+    ``cache_read_input_tokens`` and ``cache_creation_input_tokens`` are
+    surfaced separately so the cost path can price them against the cache
+    rate rather than fold them silently into the regular input bucket.
+
+    ``total_cost`` / ``input_cost`` / ``output_cost`` carry any
+    USD cost figures the upstream attached to this event (from
+    ``usage.cost``, ``usage.total_cost``, or ``usage.cost_details``) so the
+    streaming paths can re-embed them in the rebuilt ``usage`` dict and let
+    ``calculate_cost`` convert directly USD→sats instead of falling back to
+    token-based math.
+    """
+
+    event: dict
+    sse_bytes: bytes
+    input_tokens: int
+    output_tokens: int
+    cache_read_input_tokens: int
+    cache_creation_input_tokens: int
+    total_cost: float
+    input_cost: float
+    output_cost: float
+    model: str | None
+
+
+def annotate_event(event: dict, requested_model: str | None) -> AnnotatedEvent:
+    """Rewrite ``model`` fields and extract per-event token / model info.
+
+    Mutates ``event`` in place when ``requested_model`` is set so the
+    upstream's true model name doesn't leak to the client.
+    """
+    if requested_model:
+        msg = event.get("message")
+        if isinstance(msg, dict) and "model" in msg:
+            msg["model"] = requested_model
+        if "model" in event:
+            event["model"] = requested_model
+
+    in_tokens = 0
+    out_tokens = 0
+    cache_read_tokens = 0
+    cache_create_tokens = 0
+    total_cost = 0.0
+    input_cost = 0.0
+    output_cost = 0.0
+    model: str | None = None
+
+    def _coerce_float(value: object) -> float:
+        if value is None or isinstance(value, bool):
+            return 0.0
+        if not isinstance(value, (int, float, str)):
+            return 0.0
+        try:
+            return max(0.0, float(value))
+        except (TypeError, ValueError):
+            return 0.0
+
+    def _accumulate(usage: dict) -> None:
+        nonlocal in_tokens, out_tokens, cache_read_tokens, cache_create_tokens
+        nonlocal total_cost, input_cost, output_cost
+        in_tokens += int(usage.get("input_tokens") or 0)
+        out_tokens += int(usage.get("output_tokens") or 0)
+        cache_read_tokens += int(usage.get("cache_read_input_tokens") or 0)
+        cache_create_tokens += int(usage.get("cache_creation_input_tokens") or 0)
+        total_cost += _coerce_float(usage.get("total_cost"))
+        input_cost += _coerce_float(usage.get("input_cost"))
+        output_cost += _coerce_float(usage.get("output_cost"))
+
+    msg_for_meta = event.get("message")
+    if isinstance(msg_for_meta, dict):
+        if msg_for_meta.get("model"):
+            model = str(msg_for_meta["model"])
+        usage = msg_for_meta.get("usage")
+        if isinstance(usage, dict):
+            _accumulate(usage)
+
+    if isinstance(event.get("usage"), dict):
+        _accumulate(event["usage"])
+
+    # Some upstreams (notably OpenRouter-style proxies) attach cost fields
+    # directly at the event root rather than inside ``usage``.
+    for field in ("total_cost", "cost"):
+        total_cost = max(total_cost, _coerce_float(event.get(field)))
+    input_cost = max(input_cost, _coerce_float(event.get("input_cost")))
+    output_cost = max(output_cost, _coerce_float(event.get("output_cost")))
+    root_cost_details = event.get("cost_details")
+    if isinstance(root_cost_details, dict):
+        total_cost = max(
+            total_cost,
+            _coerce_float(root_cost_details.get("total_cost")),
+        )
+        input_cost = max(
+            input_cost,
+            _coerce_float(root_cost_details.get("input_cost")),
+        )
+        output_cost = max(
+            output_cost,
+            _coerce_float(root_cost_details.get("output_cost")),
+        )
+
+    event_type = str(event.get("type") or "")
+    payload = json.dumps(event)
+    if event_type:
+        sse_bytes = f"event: {event_type}\ndata: {payload}\n\n".encode()
+    else:
+        sse_bytes = f"data: {payload}\n\n".encode()
+
+    return AnnotatedEvent(
+        event,
+        sse_bytes,
+        in_tokens,
+        out_tokens,
+        cache_read_tokens,
+        cache_create_tokens,
+        total_cost,
+        input_cost,
+        output_cost,
+        model,
+    )
+
+
+async def stream_annotated_events(
+    iterator: AsyncIterator[Any],
+    requested_model: str | None,
+) -> AsyncGenerator[AnnotatedEvent, None]:
+    """Yield annotated, SSE-serialized events from a litellm stream.
+
+    Both streaming paths in ``BaseUpstreamProvider`` consume this; the only
+    divergence between them — yield-as-you-go vs buffer-then-replay — stays
+    in the caller.
+    """
+    sse_buffer = b""
+    async for chunk in iterator:
+        events, sse_buffer = events_from_chunk(chunk, sse_buffer)
+        for event in events:
+            yield annotate_event(event, requested_model)
+
+
+def embed_usd_costs(
+    usage: dict,
+    total_cost: float,
+    input_cost: float,
+    output_cost: float,
+) -> None:
+    """Mutate ``usage`` so ``calculate_cost`` will pick up the USD totals.
+
+    Mirrors the upstream shape: when any USD figure is present, attach
+    ``cost`` (used by the simple-fallback branch in ``calculate_cost``) and
+    a ``cost_details`` block (used by the preferred branch — also gives the
+    input/output USD split when we have one).
+    """
+    if total_cost <= 0 and input_cost <= 0 and output_cost <= 0:
+        return
+
+    cost_details: dict[str, float] = {}
+    effective_total = total_cost
+    if effective_total <= 0 and (input_cost > 0 or output_cost > 0):
+        effective_total = input_cost + output_cost
+
+    if effective_total > 0:
+        cost_details["total_cost"] = effective_total
+        usage["cost"] = effective_total
+    if input_cost > 0:
+        cost_details["input_cost"] = input_cost
+    if output_cost > 0:
+        cost_details["output_cost"] = output_cost
+    if cost_details:
+        usage["cost_details"] = cost_details
+
+
+def compute_refund(amount: int, unit: str, cost_msats: int) -> int:
+    if unit == "msat":
+        return amount - cost_msats
+    if unit == "sat":
+        return amount - (cost_msats + 999) // 1000
+    raise ValueError(f"Invalid unit: {unit}")
+
+
+async def dispatch_anthropic_messages(
+    *,
+    request_body: bytes | None,
+    model_obj: Model,
+    base_url: str,
+    api_key: str,
+    provider_prefix: str,
+    transform_model_name: Callable[[str], str],
+    log_extra: dict[str, Any] | None = None,
+) -> tuple[bool, Any, str | None]:
+    """Call ``litellm.anthropic.messages.acreate`` and return
+    ``(client_stream, result, requested_model)``.
+
+    Shared by the bearer-key and x-cashu paths. Raises :class:`UpstreamError`
+    on bad input or upstream failure.
+    """
+    if not request_body:
+        raise UpstreamError(
+            "Missing request body for /v1/messages", status_code=400
+        )
+
+    try:
+        body: dict = json.loads(request_body)
+    except json.JSONDecodeError as exc:
+        raise UpstreamError(
+            f"Invalid JSON in /v1/messages body: {exc}", status_code=400
+        ) from exc
+
+    body.pop("model", None)
+    # `stream` here is what the **client** asked for. Upstream is always
+    # streamed (see `upstream_stream` below); when the client asked for a
+    # non-streaming response we drain and aggregate the events into a
+    # single Anthropic Message dict before returning. This sidesteps
+    # provider-specific non-streaming caps (e.g. Fireworks rejects
+    # `max_tokens > 4096` unless `stream=true`).
+    client_stream = bool(body.pop("stream", False))
+    upstream_stream = True
+
+    dropped: dict[str, Any] = {}
+    for field in ANTHROPIC_ONLY_FIELDS:
+        if field in body:
+            dropped[field] = body.pop(field)
+    if dropped:
+        logger.debug(
+            "Dropped anthropic-only fields before litellm dispatch",
+            extra={"dropped_keys": sorted(dropped.keys())},
+        )
+
+    # Convention: `model.id` is the canonical upstream model name;
+    # `forwarded_model_id` is the public alias the internal API exposes
+    # and echoes back to the client.
+    requested_model = (
+        (model_obj.forwarded_model_id or model_obj.id) if model_obj else None
+    )
+    upstream_model = transform_model_name(model_obj.id)
+    litellm_model = f"{provider_prefix}{upstream_model}"
+
+    kwargs: dict = {
+        "model": litellm_model,
+        "api_base": base_url,
+        "api_key": api_key,
+        "stream": upstream_stream,
+        **body,
+    }
+
+    logger.info(
+        "Dispatching /v1/messages via litellm",
+        extra={
+            "model": litellm_model,
+            "resolved_provider": provider_prefix.rstrip("/"),
+            "client_stream": client_stream,
+            "upstream_stream": upstream_stream,
+            **(log_extra or {}),
+        },
+    )
+
+    try:
+        result = await litellm.anthropic.messages.acreate(**kwargs)
+    except Exception as exc:
+        raw_message = getattr(exc, "message", None) or str(exc) or repr(exc)
+        # Redact provider account identifiers before the message reaches logs
+        # or the surfaced error.
+        exc_message = redact_org_ids(raw_message)
+        exc_status = getattr(exc, "status_code", None)
+        exc_response = getattr(exc, "response", None)
+        response_text = None
+        if exc_response is not None:
+            try:
+                response_text = redact_org_ids(
+                    getattr(exc_response, "text", str(exc_response))
+                )
+            except Exception:
+                response_text = "<unreadable>"
+        status_for_classify = exc_status if isinstance(exc_status, int) else 502
+        rate_limit = classify_rate_limit(
+            status_for_classify, exc_message, getattr(exc, "headers", None)
+        )
+        logger.error(
+            "litellm dispatch failed",
+            extra={
+                "error": exc_message,
+                "error_type": type(exc).__name__,
+                "status_code": exc_status,
+                "error_code": rate_limit.code if rate_limit else None,
+                "llm_provider": getattr(exc, "llm_provider", None),
+                "body": redact_org_ids(str(getattr(exc, "body", "") or "")) or None,
+                "response_text": response_text,
+                "model": litellm_model,
+                "api_base": base_url,
+            },
+        )
+        raise UpstreamError(
+            f"Upstream error via litellm: {exc_message}",
+            status_code=status_for_classify,
+            code=rate_limit.code if rate_limit else None,
+            details=rate_limit.as_details() if rate_limit else None,
+        ) from exc
+
+    if not client_stream and hasattr(result, "__aiter__"):
+        # Client asked for a non-streaming response but we always stream
+        # from upstream — drain the events into a single Anthropic Message
+        # dict so the rest of the pipeline can treat it as if upstream had
+        # returned non-streaming. Some litellm adapters return a
+        # non-streaming dict even when ``stream=True``; in that case,
+        # leave the result as-is.
+        try:
+            aggregated: Any = await aggregate_anthropic_events_to_message(
+                cast(AsyncIterator[Any], result)
+            )
+        except Exception as exc:
+            logger.error(
+                "Failed to aggregate streamed events into message",
+                extra={
+                    "error": str(exc),
+                    "error_type": type(exc).__name__,
+                    "model": litellm_model,
+                },
+            )
+            raise UpstreamError(
+                f"Failed to aggregate upstream stream: {exc}",
+                status_code=502,
+            ) from exc
+        return client_stream, aggregated, requested_model
+
+    return client_stream, result, requested_model

routstr/upstream/ollama.py

@@ -21,6 +21,7 @@ class OllamaUpstreamProvider(BaseUpstreamProvider):
     provider_type = "ollama"
     default_base_url = "http://localhost:11434"
     platform_url = None
+    litellm_provider_prefix = "ollama_chat/"
 
     def __init__(
         self,
@@ -184,7 +185,7 @@ class OllamaUpstreamProvider(BaseUpstreamProvider):
             except Exception:
                 self._models_cache = models_with_fees
 
-            self._models_by_id = {m.id: m for m in self._models_cache}
+            self._models_by_id = {m.forwarded_model_id or m.id: m for m in self._models_cache}
             logger.info(
                 f"Refreshed models cache for {self.base_url}",
                 extra={"model_count": len(models)},

routstr/upstream/openrouter.py

@@ -1,3 +1,4 @@
+import json
 from typing import TYPE_CHECKING
 
 import httpx
@@ -15,6 +16,67 @@ class OpenRouterUpstreamProvider(BaseUpstreamProvider):
     provider_type = "openrouter"
     default_base_url = "https://openrouter.ai/api/v1"
     platform_url = "https://openrouter.ai/settings/keys"
+    supports_anthropic_messages = True
+    litellm_provider_prefix = "openrouter/"
+
+    def prepare_request_body(
+        self, body: bytes | None, model_obj: Model
+    ) -> bytes | None:
+        """Set provider.require_parameters on tool-use requests.
+
+        Without it OpenRouter can route a tool call to an endpoint that doesn't
+        support function calling and 404 with "No endpoints found that support
+        tool use". We leave a client-supplied value untouched.
+        """
+        body = super().prepare_request_body(body, model_obj)
+        if not body:
+            return body
+
+        try:
+            data = json.loads(body)
+        except json.JSONDecodeError:
+            return body
+
+        if not isinstance(data, dict) or not data.get("tools"):
+            return body
+
+        provider = data.get("provider")
+        if not isinstance(provider, dict):
+            provider = {}
+
+        if "require_parameters" in provider:
+            return body
+
+        provider["require_parameters"] = True
+        data["provider"] = provider
+        return json.dumps(data).encode()
+
+    def _apply_provider_field(self, response_json: object) -> None:
+        """Stamp the ``provider`` field for OpenRouter responses.
+
+        OpenRouter is a router, not the real serving provider, so a bare
+        ``"openrouter"`` value carries no useful information. Rules:
+
+        - Real upstream sub-provider (e.g. ``"GMICloud"``) -> ``"openrouter:GMICloud"``.
+        - Missing sub-provider, or one that merely echoes ``"openrouter"`` ->
+          ``"unknown"``.
+        - Idempotent: re-stamping never produces ``"openrouter:openrouter:..."``;
+          the ``openrouter:`` prefix appears at most once.
+        """
+        if not isinstance(response_json, dict):
+            return
+        provider_type = (self.provider_type or "").strip()
+        existing = response_json.get("provider")
+        sub = existing.strip() if isinstance(existing, str) else ""
+        # Strip any already-applied "openrouter:" prefixes (idempotency).
+        prefix = f"{provider_type}:"
+        while sub.lower().startswith(prefix.lower()):
+            sub = sub[len(prefix) :].strip()
+        # No real sub-provider, or it just echoes our own router name.
+        if not sub or sub.lower() == provider_type.lower():
+            response_json["provider"] = "unknown"
+            return
+        response_json["provider"] = f"{provider_type}:{sub}"
 
     def __init__(self, api_key: str, provider_fee: float = 1.06):
         """Initialize OpenRouter provider with API key.

routstr/upstream/perplexity.py

@@ -13,6 +13,7 @@ class PerplexityUpstreamProvider(BaseUpstreamProvider):
     provider_type = "perplexity"
     default_base_url = "https://api.perplexity.ai/"
     platform_url = "https://www.perplexity.ai/account/api/keys"
+    litellm_provider_prefix = "perplexity/"
 
     def __init__(self, api_key: str, provider_fee: float = 1.01):
         super().__init__(

routstr/upstream/ppqai.py

@@ -3,11 +3,12 @@ from __future__ import annotations
 from typing import TYPE_CHECKING, Optional
 
 import httpx
-from pydantic import BaseModel, Field
+from pydantic.v1 import BaseModel, Field
 
 from ..core.logging import get_logger
 from ..payment.models import Architecture, Model, Pricing, async_fetch_openrouter_models
 from .base import BaseUpstreamProvider, TopupData
+from .ehbp import EHBPForwardingTarget
 
 if TYPE_CHECKING:
     from ..core.db import UpstreamProviderRow
@@ -39,6 +40,10 @@ class PPQAIUpstreamProvider(BaseUpstreamProvider):
     default_base_url = "https://api.ppq.ai"
     platform_url = "https://ppq.ai/api-docs"
     IGNORED_MODEL_IDS: list[str] = ["auto"]
+    # PPQ.AI has a private encrypted endpoint, but this proxy currently has no
+    # provider-attested usage extractor/model binding for it. Keep EHBP disabled
+    # until a ConfidentialInferenceProfile can bill it without max-cost fallback.
+    supports_ehbp = False
 
     def __init__(self, api_key: str, provider_fee: float = 1.0):
         super().__init__(
@@ -70,6 +75,20 @@ class PPQAIUpstreamProvider(BaseUpstreamProvider):
     def transform_model_name(self, model_id: str) -> str:
         return model_id
 
+    def get_ehbp_forwarding_target(
+        self, path: str, model_obj: Model
+    ) -> EHBPForwardingTarget:
+        """Return the PPQ.AI private enclave target for EHBP requests.
+
+        PPQ.AI exposes EHBP-aware inference under /private/v1/... separate
+        from the public /v1/... endpoint. The encrypted body remains opaque to
+        Routstr, so PPQ.AI also needs X-Private-Model for routing/billing.
+        """
+        return EHBPForwardingTarget(
+            url=f"{self.base_url.rstrip('/')}/private/{path.lstrip('/')}",
+            headers={"X-Private-Model": model_obj.forwarded_model_id or model_obj.id},
+        )
+
     @classmethod
     async def create_account_static(cls) -> dict[str, object]:
         """Create a new PPQ.AI account without requiring an instance.

routstr/upstream/rate_limit.py

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

routstr/upstream/request_correction.py

@@ -0,0 +1,142 @@
+"""Reactive request-correction layer.
+
+When an upstream rejects a request with a recoverable 4xx error, this layer
+tries to *fix* the request body and let the caller retry the same upstream
+instead of failing outright. It is provider-agnostic: correctors key off the
+upstream's own error wording, so the same recovery works across every provider.
+
+The layer is a small pipeline of :data:`Corrector` callables. Each corrector
+inspects the parsed request body and the upstream error message and either
+returns a corrected body (plus a short label identifying the fix) or declines
+by returning ``None``. Adding a new reactive fix means writing one corrector
+and adding it to :data:`DEFAULT_CORRECTORS` — no changes to the proxy loop.
+
+All corrections are immutable: a corrector never mutates the body it is given,
+it returns a new ``dict``. The proxy threads an ``applied`` set of fix labels
+through retries so each distinct fix is applied at most once, guaranteeing the
+retry loop always terminates.
+"""
+
+from __future__ import annotations
+
+import json
+import re
+from collections.abc import Callable, Sequence
+from dataclasses import dataclass
+
+from fastapi.responses import Response
+
+from ..core import get_logger
+
+logger = get_logger(__name__)
+
+
+# Matches upstream error text that names a single rejected request parameter,
+# e.g. "`temperature` is deprecated for this model." or
+# "parameter 'top_p' is not supported". Keys off the upstream's own wording so
+# a 400 about an unsupported sampling/option field can be recovered by stripping
+# that field and retrying the same upstream.
+_UNSUPPORTED_PARAM_RE = re.compile(
+    r"[`'\"]?(?P<param>[a-zA-Z_][a-zA-Z0-9_]*)[`'\"]?\s+is\s+"
+    r"(?:deprecated|not\s+supported|unsupported|no\s+longer\s+supported)",
+    re.IGNORECASE,
+)
+
+
+# A corrector inspects the parsed request body and the upstream error message
+# and returns ``(new_body_dict, label)`` for a fix it can apply, or ``None`` to
+# decline. ``label`` identifies the fix so it is applied at most once per request.
+Corrector = Callable[[dict, str], "tuple[dict, str] | None"]
+
+
+@dataclass(frozen=True)
+class Correction:
+    """A successful request correction ready to retry.
+
+    ``body`` is the corrected JSON body (encoded), ``label`` identifies the fix
+    that was applied (e.g. the stripped param name) so the caller can guard
+    against applying the same fix twice.
+    """
+
+    body: bytes
+    label: str
+
+
+def extract_error_message(response: Response) -> str:
+    """Best-effort extraction of an error message string from a proxy Response."""
+    body_bytes = getattr(response, "body", None)
+    if not body_bytes:
+        return ""
+    try:
+        data = json.loads(body_bytes)
+    except Exception:
+        return body_bytes.decode("utf-8", errors="ignore")[:500]
+    if isinstance(data, dict):
+        err = data.get("error")
+        if isinstance(err, dict):
+            msg = err.get("message") or err.get("detail")
+            if isinstance(msg, str):
+                return msg
+        elif isinstance(err, str):
+            return err
+        if isinstance(data.get("message"), str):
+            return data["message"]
+    return ""
+
+
+def strip_unsupported_param(
+    body: dict, error_message: str
+) -> tuple[dict, str] | None:
+    """Drop a top-level param the upstream named as unsupported/deprecated.
+
+    Returns ``(new_body, param)`` (a new dict, original untouched) when the
+    error names a top-level param present in the body, otherwise ``None``.
+    """
+    match = _UNSUPPORTED_PARAM_RE.search(error_message)
+    if not match:
+        return None
+    param = match.group("param")
+    if param not in body:
+        return None
+    new_body = {k: v for k, v in body.items() if k != param}
+    return new_body, param
+
+
+# Ordered pipeline of correctors tried on each recoverable rejection.
+DEFAULT_CORRECTORS: tuple[Corrector, ...] = (strip_unsupported_param,)
+
+
+def correct_request(
+    request_body: bytes,
+    error_message: str,
+    applied: set[str],
+    correctors: Sequence[Corrector] = DEFAULT_CORRECTORS,
+) -> Correction | None:
+    """Try to correct a rejected request body so it can be retried.
+
+    Runs each corrector in order against the parsed body and ``error_message``.
+    The first corrector that proposes a fix whose ``label`` is not already in
+    ``applied`` wins; its result is returned as a :class:`Correction`. Returns
+    ``None`` when nothing parses, nothing matches, or every proposed fix was
+    already applied — the caller then treats the response as a normal failure.
+
+    ``applied`` is read-only here; the caller records the returned ``label`` to
+    bound retries and guarantee forward progress.
+    """
+    if not request_body or not error_message:
+        return None
+    try:
+        data = json.loads(request_body)
+    except Exception:
+        return None
+    if not isinstance(data, dict):
+        return None
+    for corrector in correctors:
+        result = corrector(data, error_message)
+        if result is None:
+            continue
+        new_body, label = result
+        if label in applied:
+            continue
+        return Correction(body=json.dumps(new_body).encode(), label=label)
+    return None

routstr/upstream/routstr.py

@@ -18,6 +18,10 @@ class RoutstrUpstreamProvider(BaseUpstreamProvider):
     provider_type = "routstr"
     default_base_url = None
     platform_url = None
+    # Upstream Routstr nodes serve `/v1/messages` natively, so forward the
+    # request as-is instead of round-tripping through litellm's
+    # Anthropic→OpenAI translator.
+    supports_anthropic_messages = True
 
     def __init__(
         self,
@@ -43,6 +47,13 @@ class RoutstrUpstreamProvider(BaseUpstreamProvider):
         )
         self.settings = provider_settings or {}
 
+    def normalize_request_path(
+        self, path: str, model_obj: "Model | None" = None
+    ) -> str:
+        """Preserve the ``v1/`` prefix when forwarding to an upstream Routstr.
+        """
+        return path.lstrip("/")
+
     @classmethod
     def from_db_row(
         cls, provider_row: "UpstreamProviderRow"

routstr/upstream/tinfoil.py

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

routstr/upstream/tinfoil_trailer.py

@@ -0,0 +1,161 @@
+"""h11-based HTTP client for EHBP requests that captures HTTP trailers.
+
+httpx/httpcore silently discard HTTP trailers during chunked transfer
+decoding. Tinfoil returns ``X-Tinfoil-Usage-Metrics`` as a trailer on
+streaming responses, so we need a lower-level HTTP client that preserves
+trailers from the h11 ``EndOfMessage`` event.
+
+Because EHBP response bodies are opaque encrypted blobs, buffering the full
+response is acceptable — the client decrypts the complete body regardless of
+whether it arrived streamed or buffered.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import ssl
+from dataclasses import dataclass, field
+from urllib.parse import urlsplit
+
+import h11
+
+from ..core import get_logger
+
+logger = get_logger(__name__)
+
+_READ_BUFSIZE = 65536
+_DEFAULT_TIMEOUT_SECONDS = 30.0
+_DEFAULT_CLOSE_TIMEOUT_SECONDS = 1.0
+_DEFAULT_MAX_RESPONSE_BYTES = 25 * 1024 * 1024
+
+
+@dataclass
+class TrailerResponse:
+    """Buffered HTTP response with optional trailer headers."""
+
+    status_code: int
+    headers: list[tuple[str, str]]
+    body: bytes
+    trailers: list[tuple[str, str]] = field(default_factory=list)
+
+
+def _get_header(headers: list[tuple[str, str]], name: str) -> str | None:
+    name_lower = name.lower()
+    for k, v in headers:
+        if k.lower() == name_lower:
+            return v
+    return None
+
+
+async def forward_with_trailer(
+    *,
+    method: str,
+    url: str,
+    headers: dict[str, str],
+    body: bytes,
+    timeout_seconds: float = _DEFAULT_TIMEOUT_SECONDS,
+    max_response_bytes: int = _DEFAULT_MAX_RESPONSE_BYTES,
+    close_timeout_seconds: float = _DEFAULT_CLOSE_TIMEOUT_SECONDS,
+) -> TrailerResponse:
+    """Send an HTTP/1.1 request via h11 and capture HTTP trailers.
+
+    Returns a :class:`TrailerResponse` with the full buffered body and any
+    trailer headers from the ``EndOfMessage`` event.
+    """
+    parsed = urlsplit(url)
+    host = parsed.hostname
+    if not host:
+        raise ValueError(f"Invalid URL (no hostname): {url}")
+    port = parsed.port or 443
+    path = parsed.path or "/"
+    if parsed.query:
+        path = f"{path}?{parsed.query}"
+
+    ssl_ctx = ssl.create_default_context()
+    reader, writer = await asyncio.wait_for(
+        asyncio.open_connection(host, port, ssl=ssl_ctx),
+        timeout=timeout_seconds,
+    )
+
+    try:
+        # Build HTTP/1.1 request
+        header_lines = [f"{method} {path} HTTP/1.1"]
+        has_host = any(k.lower() == "host" for k in headers)
+        if not has_host:
+            header_lines.append(f"Host: {host}")
+        header_lines.append("Connection: close")
+
+        for key, value in headers.items():
+            if key.lower() in ("host", "connection"):
+                continue
+            header_lines.append(f"{key}: {value}")
+
+        if body and not any(k.lower() == "content-length" for k in headers):
+            header_lines.append(f"Content-Length: {len(body)}")
+
+        request_data = "\r\n".join(header_lines).encode() + b"\r\n\r\n"
+        if body:
+            request_data += body
+
+        writer.write(request_data)
+        await asyncio.wait_for(writer.drain(), timeout=timeout_seconds)
+
+        # Parse response with h11
+        conn = h11.Connection(h11.CLIENT)
+        status_code = 0
+        resp_headers: list[tuple[str, str]] = []
+        body_chunks: list[bytes] = []
+        body_size = 0
+        trailers: list[tuple[str, str]] = []
+
+        while True:
+            event = conn.next_event()
+
+            if event is h11.NEED_DATA:
+                data = await asyncio.wait_for(
+                    reader.read(_READ_BUFSIZE),
+                    timeout=timeout_seconds,
+                )
+                conn.receive_data(data if data else b"")
+                continue
+
+            if isinstance(event, h11.Response):
+                status_code = event.status_code
+                resp_headers = [(k.decode(), v.decode()) for k, v in event.headers]
+
+            elif isinstance(event, h11.Data):
+                body_size += len(event.data)
+                if body_size > max_response_bytes:
+                    raise ValueError(
+                        f"EHBP response exceeded {max_response_bytes} bytes"
+                    )
+                body_chunks.append(event.data)
+
+            elif isinstance(event, h11.EndOfMessage):
+                for k, v in event.headers:
+                    trailers.append((k.decode(), v.decode()))
+                break
+
+            elif isinstance(event, h11.PAUSED):
+                # Shouldn't happen for simple request/response, but break safely
+                logger.warning("h11 PAUSED event during EHBP response parsing")
+                break
+
+            elif isinstance(event, h11.ConnectionClosed):
+                break
+
+        return TrailerResponse(
+            status_code=status_code,
+            headers=resp_headers,
+            body=b"".join(body_chunks),
+            trailers=trailers,
+        )
+    finally:
+        writer.close()
+        if close_timeout_seconds > 0:
+            try:
+                await asyncio.wait_for(
+                    writer.wait_closed(), timeout=close_timeout_seconds
+                )
+            except Exception:
+                pass

routstr/upstream/xai.py

@@ -13,6 +13,7 @@ class XAIUpstreamProvider(BaseUpstreamProvider):
     provider_type = "x-ai"
     default_base_url = "https://api.x.ai/v1"
     platform_url = "https://console.x.ai/"
+    litellm_provider_prefix = "xai/"
 
     def __init__(self, api_key: str, provider_fee: float = 1.01):
         super().__init__(

routstr/wallet.py

@@ -1,10 +1,14 @@
 import asyncio
+import re
 import time
+import typing
 from typing import TypedDict
 
 from cashu.core.base import Proof, Token
+from cashu.core.mint_info import MintInfo as _CashuMintInfo
 from cashu.wallet.helpers import deserialize_token_from_string
 from cashu.wallet.wallet import Wallet
+from pydantic_core import PydanticUndefined
 from sqlmodel import col, select, update
 
 from .core import db, get_logger
@@ -12,6 +16,18 @@ from .core.db import store_cashu_transaction
 from .core.settings import settings
 from .payment.lnurl import raw_send_to_lnurl
 
+# cashu still declares Optional[X] without explicit defaults on MintInfo.
+# Under pydantic v2 those are required, but real mints omit many of them.
+# Default Optional fields to None at import time so balance fetches don't 422.
+for _name, _field in _CashuMintInfo.model_fields.items():
+    _annot = _field.annotation
+    _is_optional = typing.get_origin(_annot) is typing.Union and type(
+        None
+    ) in typing.get_args(_annot)
+    if _is_optional and _field.default is PydanticUndefined:
+        _field.default = None
+_CashuMintInfo.model_rebuild(force=True)
+
 logger = get_logger(__name__)
 
 
@@ -20,6 +36,24 @@ async def get_balance(unit: str) -> int:
     return wallet.available_balance.amount
 
 
+async def _redeem_same_mint(
+    wallet: Wallet, token_obj: Token
+) -> tuple[int, str, str]:  # amount, unit, mint_url
+    """Redeem proofs at their own issuing mint (no cross-mint swap).
+
+    split() re-mints the incoming proofs into fresh ones we own so the sender
+    can't double-spend them. With include_fees=True the mint deducts its NUT-02
+    per-proof input fee, so we end up holding only `amount - input_fees`. Credit
+    that, not the face value, or routstr over-credits the user and its wallet
+    drifts insolvent.
+    """
+    await wallet.load_mint(keyset_id=token_obj.keysets[0])
+    wallet.verify_proofs_dleq(token_obj.proofs)
+    input_fees = wallet.get_fees_for_proofs(token_obj.proofs)
+    await wallet.split(proofs=token_obj.proofs, amount=0, include_fees=True)
+    return int(token_obj.amount) - input_fees, token_obj.unit, token_obj.mint
+
+
 async def recieve_token(
     token: str,
 ) -> tuple[int, str, str]:  # amount, unit, mint_url
@@ -33,25 +67,61 @@ async def recieve_token(
     if token_obj.mint not in settings.cashu_mints:
         return await swap_to_primary_mint(token_obj, wallet)
 
-    wallet.verify_proofs_dleq(token_obj.proofs)
-    await wallet.split(proofs=token_obj.proofs, amount=0, include_fees=True)
-
-    return token_obj.amount, token_obj.unit, token_obj.mint
+    return await _redeem_same_mint(wallet, token_obj)
 
 
 async def send(amount: int, unit: str, mint_url: str | None = None) -> tuple[int, str]:
     """Internal send function - returns amount and serialized token"""
-    wallet: Wallet = await get_wallet(mint_url or settings.primary_mint, unit)
-    proofs = get_proofs_per_mint_and_unit(
-        wallet, mint_url or settings.primary_mint, unit
+    effective_mint_url = mint_url or settings.primary_mint
+    wallet: Wallet = await get_wallet(effective_mint_url, unit)
+    proofs = get_proofs_per_mint_and_unit(wallet, effective_mint_url, unit)
+    proofs_for_mint = sum(p.amount for p in proofs)
+
+    # Fallback: proofs from untrusted source mints are swapped to primary_mint
+    # during receive, so the user's preferred refund_mint_url may have no proofs
+    # even though the global wallet has the balance.
+    if proofs_for_mint < amount and effective_mint_url != settings.primary_mint:
+        logger.info(
+            f"send: insufficient proofs at {effective_mint_url} "
+            f"(have {proofs_for_mint}, need {amount}), falling back to primary_mint={settings.primary_mint}"
+        )
+        effective_mint_url = settings.primary_mint
+        wallet = await get_wallet(effective_mint_url, unit)
+        proofs = get_proofs_per_mint_and_unit(wallet, effective_mint_url, unit)
+        proofs_for_mint = sum(p.amount for p in proofs)
+
+    all_mint_urls = list({k.mint_url for k in wallet.keysets.values()})
+    proof_summary = {
+        f"{k.mint_url}/{k.unit.name}": sum(p.amount for p in wallet.proofs if p.id == k.id)
+        for k in wallet.keysets.values()
+    }
+    # Show ALL proofs in DB by keyset_id, regardless of whether the loaded wallet
+    # knows about that keyset. This reveals proofs orphaned under stale keysets.
+    raw_proofs_by_keyset: dict[str, int] = {}
+    for p in wallet.proofs:
+        raw_proofs_by_keyset[p.id] = raw_proofs_by_keyset.get(p.id, 0) + p.amount
+    logger.info(
+        f"send: proof inventory | mint={effective_mint_url} unit={unit} amount={amount} "
+        f"primary_mint={settings.primary_mint} proofs_for_mint={proofs_for_mint} "
+        f"all_mints={all_mint_urls} by_keyset={proof_summary} "
+        f"raw_proofs_by_keyset_id={raw_proofs_by_keyset} "
+        f"total_wallet_proofs={sum(p.amount for p in wallet.proofs)}"
     )
 
+    # Reserve proofs only after serialization succeeds — if serialize_proofs or
+    # swap_to_send fails mid-way, proofs stay unreserved so dashboard balance
+    # doesn't go negative.
     send_proofs, _ = await wallet.select_to_send(
-        proofs, amount, set_reserved=True, include_fees=False
-    )
-    token = await wallet.serialize_proofs(
-        send_proofs, include_dleq=False, legacy=False, memo=None
+        proofs, amount, set_reserved=False, include_fees=False
     )
+    try:
+        token = await wallet.serialize_proofs(
+            send_proofs, include_dleq=False, legacy=False, memo=None
+        )
+    except Exception:
+        await wallet.set_reserved_for_send(send_proofs, reserved=False)
+        raise
+    await wallet.set_reserved_for_send(send_proofs, reserved=True)
     return amount, token
 
 
@@ -60,16 +130,83 @@ async def send_token(amount: int, unit: str, mint_url: str | None = None) -> str
     return token
 
 
+# A foreign mint's fee_reserve is a non-binding estimate (NUT-05): the mint may
+# demand more when re-quoting or at melt execution. Instead of padding the
+# estimate with a safety buffer (which strands the margin at the foreign mint
+# on every swap), the swap retries with the amount recomputed from the fees the
+# mint actually demands, up to this many attempts.
+_MAX_SWAP_ATTEMPTS = 3
+
+_MINT_ERROR_CODE_RE = re.compile(r"\(Code: (\d+)\)")
+_MELT_SHORTFALL_RE = re.compile(r"Provided: (\d+), needed: (\d+)")
+
+# Insufficient-melt-inputs failures differ across mint implementations. 11005 is
+# the registered "Transaction is not balanced" code (cdk), specific enough to
+# trust on the code alone. 11000 is nutshell's generic, unregistered
+# TransactionError covering many unrelated failures, so it only counts as a fee
+# shortfall alongside the "not enough inputs" detail text. With no code suffix at
+# all, that same text is the only signal.
+
+
+def _net_minted_amount(amount_msat: int, token_unit: str, fees: int) -> int:
+    """
+    Convert the token value minus fees (given in the token unit) into an
+    amount in the primary mint's unit.
+    """
+    fee_msat = fees * 1000 if token_unit == "sat" else fees
+    remaining_msat = amount_msat - fee_msat
+    if settings.primary_mint_unit == "sat":
+        return int(remaining_msat // 1000)
+    return int(remaining_msat)
+
+
+def _melt_insufficient_shortfall(error: Exception) -> int | None:
+    """
+    Classify a melt failure: return the observed shortfall (in the token unit)
+    when the mint rejected the inputs as insufficient, or None when the failure
+    is unrelated to fees and must not be retried (e.g. a Lightning payment
+    failure, where a smaller invoice would not help).
+
+    Cashu errors carry no structured amounts (NUT-00 defines only detail/code,
+    flattened to "Mint Error: <detail> (Code: <code>)" by cashu-py), so the
+    classification uses the code and the shortfall must be inferred: the
+    "Provided: X, needed: Y" amounts are nutshell-specific free text and only
+    refine the shortfall when present; otherwise shrink one unit at a time.
+    """
+    message = str(error)
+    code_match = _MINT_ERROR_CODE_RE.search(message)
+    code = code_match.group(1) if code_match is not None else None
+    has_shortfall_text = "not enough inputs" in message.lower()
+
+    match code:
+        case "11005":  # registered TransactionUnbalanced: trust the code
+            pass
+        case "11000" if has_shortfall_text:  # generic nutshell error: needs the text
+            pass
+        case None if has_shortfall_text:  # no code suffix: text is the only signal
+            pass
+        case _:  # other codes, a bare 11000, or no signal: must not retry
+            return None
+
+    amounts = _MELT_SHORTFALL_RE.search(message)
+    if amounts is not None:
+        provided, needed = int(amounts.group(1)), int(amounts.group(2))
+        if needed > provided:
+            return needed - provided
+    return 1
+
+
 async def _calculate_swap_amount(
     amount_msat: int,
     token_unit: str,
     token_mint_url: str,
     token_wallet: Wallet,
     primary_wallet: Wallet,
+    proofs: list,
 ) -> int:
     """
     Calculate the amount to mint on the primary mint after accounting for
-    potential swap fees (melt fees) on the foreign mint.
+    melt fees and NUT-02 input fees on the foreign mint.
     """
     if settings.primary_mint_unit == "sat":
         receive_amount = amount_msat // 1000
@@ -96,26 +233,20 @@ async def _calculate_swap_amount(
         dummy_melt_quote = await token_wallet.melt_quote(dummy_mint_quote.request)
 
         fee_reserve = dummy_melt_quote.fee_reserve
-        if token_unit == "sat":
-            fee_msat = fee_reserve * 1000
-        else:
-            fee_msat = fee_reserve
-
-        amount_msat_after_fee = amount_msat - fee_msat
-
-        if settings.primary_mint_unit == "sat":
-            minted_amount = int(amount_msat_after_fee // 1000)
-        else:
-            minted_amount = int(amount_msat_after_fee)
+        input_fees = token_wallet.get_fees_for_proofs(proofs)
+        total_fees = fee_reserve + input_fees
+        minted_amount = _net_minted_amount(amount_msat, token_unit, total_fees)
 
         if minted_amount <= 0:
-            raise ValueError(f"Fees ({fee_reserve} {token_unit}) exceed token amount")
+            raise ValueError(f"Fees ({total_fees} {token_unit}) exceed token amount")
 
         logger.info(
             "swap_to_primary_mint: fee estimation result",
             extra={
                 "token_amount_sat": amount_msat // 1000,
-                "estimated_fee_sat": fee_msat // 1000,
+                "estimated_fee": total_fees,
+                "estimated_fee_unit": token_unit,
+                "input_fees": input_fees,
                 "minted_amount": minted_amount,
                 "minted_unit": settings.primary_mint_unit,
             },
@@ -154,10 +285,9 @@ async def swap_to_primary_mint(
         amount_msat = token_amount
     else:
         raise ValueError("Invalid unit")
-    primary_wallet = await get_wallet(settings.primary_mint, settings.primary_mint_unit)
-
-    # If the token is already from the primary mint, we don't need to swap
-    # and we definitely don't want to calculate or pay fees.
+    # If the token is already from the primary mint, we don't need a cross-mint
+    # swap — redeem it same-mint. There's no melt/Lightning fee, but the mint's
+    # NUT-02 input fee still applies; _redeem_same_mint accounts for it.
     if token_obj.mint == settings.primary_mint:
         logger.info(
             "swap_to_primary_mint: token already on primary mint, skipping swap",
@@ -167,8 +297,9 @@ async def swap_to_primary_mint(
                 "unit": token_obj.unit,
             },
         )
-        await token_wallet.split(proofs=token_obj.proofs, amount=0, include_fees=True)
-        return token_amount, token_obj.unit, token_obj.mint
+        return await _redeem_same_mint(token_wallet, token_obj)
+
+    primary_wallet = await get_wallet(settings.primary_mint, settings.primary_mint_unit)
 
     minted_amount = await _calculate_swap_amount(
         amount_msat,
@@ -176,85 +307,185 @@ async def swap_to_primary_mint(
         token_obj.mint,
         token_wallet,
         primary_wallet,
+        token_obj.proofs,
     )
 
-    mint_quote = await primary_wallet.request_mint(minted_amount)
-    logger.info(
-        "swap_to_primary_mint: mint quote received",
-        extra={"mint_quote_id": mint_quote.quote},
-    )
+    # The estimate above is non-binding: the mint may demand a higher fee on the
+    # real quote or reject the melt outright. Retry the quote/melt cycle with the
+    # amount recomputed from the fees the mint actually demands.
+    observed_extra_fee = 0
+    attempt = 0
+    while True:
+        attempt += 1
+        mint_quote = await primary_wallet.request_mint(minted_amount)
+        logger.info(
+            "swap_to_primary_mint: mint quote received",
+            extra={"mint_quote_id": mint_quote.quote, "attempt": attempt},
+        )
 
-    melt_quote = await token_wallet.melt_quote(mint_quote.request)
-    total_needed = melt_quote.amount + melt_quote.fee_reserve
-    logger.info(
-        "swap_to_primary_mint: melt quote received",
-        extra={
-            "melt_quote_id": melt_quote.quote,
-            "melt_amount": melt_quote.amount,
-            "melt_fee_reserve": melt_quote.fee_reserve,
-            "total_needed": total_needed,
-            "token_amount": token_amount,
-        },
-    )
-
-    if total_needed > token_amount:
-        logger.warning(
-            "swap_to_primary_mint: insufficient token amount for melt fees",
+        melt_quote = await token_wallet.melt_quote(mint_quote.request)
+        input_fees = token_wallet.get_fees_for_proofs(token_obj.proofs)
+        total_needed = melt_quote.amount + melt_quote.fee_reserve + input_fees
+        logger.info(
+            "swap_to_primary_mint: melt quote received",
             extra={
-                "token_amount": token_amount,
+                "melt_quote_id": melt_quote.quote,
                 "melt_amount": melt_quote.amount,
                 "melt_fee_reserve": melt_quote.fee_reserve,
+                "input_fees": input_fees,
                 "total_needed": total_needed,
-                "shortfall": total_needed - token_amount,
+                "token_amount": token_amount,
+                "attempt": attempt,
             },
         )
-        raise ValueError(
-            f"Token amount ({token_amount} {token_obj.unit}) is insufficient to cover "
-            f"melt fees. Needed: {total_needed} {token_obj.unit} "
-            f"(amount: {melt_quote.amount} + fee: {melt_quote.fee_reserve})"
-        )
 
-    try:
-        _ = await token_wallet.melt(
-            proofs=token_obj.proofs,
-            invoice=mint_quote.request,
-            fee_reserve_sat=melt_quote.fee_reserve,
-            quote_id=melt_quote.quote,
-        )
-    except Exception as e:
-        logger.error(
-            "swap_to_primary_mint: melt failed",
-            extra={
-                "error": str(e),
-                "error_type": type(e).__name__,
-                "foreign_mint": token_obj.mint,
-                "token_amount": token_amount,
-                "melt_quote_id": melt_quote.quote,
-                "total_needed": total_needed,
-            },
-        )
-        raise ValueError(
-            f"Failed to melt token from foreign mint {token_obj.mint}: {e}"
-        ) from e
+        if total_needed > token_amount:
+            recomputed = _net_minted_amount(
+                amount_msat,
+                token_obj.unit,
+                melt_quote.fee_reserve + input_fees + observed_extra_fee,
+            )
+            if attempt >= _MAX_SWAP_ATTEMPTS or recomputed <= 0:
+                logger.warning(
+                    "swap_to_primary_mint: insufficient token amount for melt fees",
+                    extra={
+                        "token_amount": token_amount,
+                        "melt_amount": melt_quote.amount,
+                        "melt_fee_reserve": melt_quote.fee_reserve,
+                        "input_fees": input_fees,
+                        "total_needed": total_needed,
+                        "shortfall": total_needed - token_amount,
+                        "attempts": attempt,
+                    },
+                )
+                raise ValueError(
+                    f"Token amount ({token_amount} {token_obj.unit}) is insufficient to cover "
+                    f"melt fees. Needed: {total_needed} {token_obj.unit} "
+                    f"(amount: {melt_quote.amount} + fee: {melt_quote.fee_reserve} + input_fees: {input_fees})"
+                )
+            logger.warning(
+                "swap_to_primary_mint: melt quote exceeds token amount, retrying",
+                extra={
+                    "total_needed": total_needed,
+                    "token_amount": token_amount,
+                    "retry_minted_amount": recomputed,
+                    "attempt": attempt,
+                },
+            )
+            minted_amount = recomputed
+            continue
+
+        try:
+            _ = await token_wallet.melt(
+                proofs=token_obj.proofs,
+                invoice=mint_quote.request,
+                fee_reserve_sat=melt_quote.fee_reserve,
+                quote_id=melt_quote.quote,
+            )
+        except Exception as e:
+            shortfall = _melt_insufficient_shortfall(e)
+            recomputed = 0
+            if shortfall is not None:
+                observed_extra_fee += shortfall
+                recomputed = _net_minted_amount(
+                    amount_msat,
+                    token_obj.unit,
+                    melt_quote.fee_reserve + input_fees + observed_extra_fee,
+                )
+            if shortfall is None or attempt >= _MAX_SWAP_ATTEMPTS or recomputed <= 0:
+                logger.error(
+                    "swap_to_primary_mint: melt failed",
+                    extra={
+                        "error": str(e),
+                        "error_type": type(e).__name__,
+                        "foreign_mint": token_obj.mint,
+                        "token_amount": token_amount,
+                        "melt_quote_id": melt_quote.quote,
+                        "total_needed": total_needed,
+                        "attempts": attempt,
+                    },
+                )
+                raise ValueError(
+                    f"Failed to melt token from foreign mint {token_obj.mint}: {e}"
+                ) from e
+            logger.warning(
+                "swap_to_primary_mint: mint demanded more than quoted at melt, retrying",
+                extra={
+                    "shortfall": shortfall,
+                    "retry_minted_amount": recomputed,
+                    "attempt": attempt,
+                },
+            )
+            minted_amount = recomputed
+            continue
+
+        break
 
     logger.info(
         "swap_to_primary_mint: melt succeeded, minting on primary",
         extra={"minted_amount": minted_amount, "mint_quote_id": mint_quote.quote},
     )
 
+    await primary_wallet.load_proofs(reload=True)
+    pre_mint_balance = primary_wallet.available_balance.amount
     try:
         _ = await primary_wallet.mint(minted_amount, quote_id=mint_quote.quote)
     except Exception as e:
-        logger.error(
-            "swap_to_primary_mint: mint on primary failed after successful melt",
-            extra={
-                "error": str(e),
-                "error_type": type(e).__name__,
-                "minted_amount": minted_amount,
-                "mint_quote_id": mint_quote.quote,
-            },
-        )
-        raise
+        if "11003" in str(e) or "outputs already signed" in str(e).lower():
+            # Previous mint call signed outputs at the mint but failed before
+            # bump_secret_derivation ran locally. Recover orphaned proofs and
+            # advance the counter so the next request derives fresh secrets.
+            logger.warning(
+                "swap_to_primary_mint: outputs already signed — recovering orphaned proofs",
+                extra={"mint_quote_id": mint_quote.quote, "minted_amount": minted_amount},
+            )
+            try:
+                for keyset_id in primary_wallet.keysets:
+                    await primary_wallet.restore_tokens_for_keyset(keyset_id, to=1, batch=25)
+                await primary_wallet.load_proofs(reload=True)
+                post_recovery_balance = primary_wallet.available_balance.amount
+                balance_gained = post_recovery_balance - pre_mint_balance
+                logger.info(
+                    "swap_to_primary_mint: recovery scan completed",
+                    extra={
+                        "pre_mint_balance": pre_mint_balance,
+                        "post_recovery_balance": post_recovery_balance,
+                        "balance_gained": balance_gained,
+                        "expected": minted_amount,
+                    },
+                )
+                if balance_gained < minted_amount:
+                    # Recovery scan ran but did NOT restore the orphaned proofs
+                    # (mint reports them as spent — they're stuck). Refuse to
+                    # credit the API key balance for proofs we don't actually hold.
+                    raise ValueError(
+                        f"Swap recovery failed: mint signed outputs but proofs are "
+                        f"unrecoverable (mint reports them spent). "
+                        f"Expected {minted_amount}, recovered {balance_gained}. "
+                        f"Local wallet DB ('.wallet/') state is corrupted — "
+                        f"the counter for keyset is stuck at a bad index range."
+                    )
+            except ValueError:
+                raise
+            except Exception as recovery_err:
+                logger.error(
+                    "swap_to_primary_mint: recovery failed",
+                    extra={"error": str(recovery_err)},
+                )
+                raise ValueError(
+                    f"Mint on primary failed and recovery unsuccessful: {e}"
+                ) from e
+        else:
+            logger.error(
+                "swap_to_primary_mint: mint on primary failed after successful melt",
+                extra={
+                    "error": str(e),
+                    "error_type": type(e).__name__,
+                    "minted_amount": minted_amount,
+                    "mint_quote_id": mint_quote.quote,
+                },
+            )
+            raise
 
     logger.info(
         "swap_to_primary_mint: completed successfully",
@@ -293,6 +524,20 @@ async def credit_balance(
                 "credit_balance: Converted to msat", extra={"amount_msat": amount}
             )
 
+        # Guard against zero/negative redemptions (empty or dust tokens, or
+        # swap-to-primary-mint amounts that net to <= 0 after fees). Raising here
+        # — before the UPDATE/commit below — leaves any freshly-created, still
+        # uncommitted ApiKey row to be rolled back when the request session
+        # closes, instead of persisting an orphan key with balance 0.
+        if amount <= 0:
+            logger.error(
+                "credit_balance: Redeemed amount is zero or negative; refusing to credit",
+                extra={"amount": amount, "unit": unit, "mint_url": mint_url},
+            )
+            raise ValueError(
+                f"Redeemed token amount must be positive, got {amount} msats"
+            )
+
         logger.info(
             "credit_balance: Updating balance",
             extra={"old_balance": key.balance, "credit_amount": amount},
@@ -304,7 +549,11 @@ async def credit_balance(
             .where(col(db.ApiKey.hashed_key) == key.hashed_key)
             .values(balance=(db.ApiKey.balance) + amount)
         )
-        await session.exec(stmt)  # type: ignore[call-overload]
+        result = await session.exec(stmt)  # type: ignore[call-overload]
+        # If pruning removed this key after redemption, do not commit a no-op
+        # balance update and pretend the top-up succeeded.
+        if (getattr(result, "rowcount", 0) or 0) == 0:
+            raise ValueError("API key disappeared before credit could be recorded")
         await session.commit()
         await session.refresh(key)
 
@@ -326,7 +575,7 @@ async def credit_balance(
         except Exception:
             pass
 
-        logger.info(
+        logger.debug(
             "Cashu token successfully redeemed and stored",
             extra={"amount": amount, "unit": unit, "mint_url": mint_url},
         )
@@ -428,7 +677,7 @@ async def fetch_all_balances(
                 "unit": unit,
                 "wallet_balance": proofs_balance,
                 "user_balance": user_balance,
-                "owner_balance": proofs_balance - user_balance,
+                "owner_balance": proofs_balance - user_balance if proofs_balance != 0 else 0,
             }
             return result
         except Exception as e:
@@ -443,11 +692,19 @@ async def fetch_all_balances(
             }
             return error_result
 
+    # Build the set of mints to inspect. Received tokens are stored against
+    # ``primary_mint`` (which defaults to a real mint even when ``cashu_mints``
+    # is empty), so include it as a fallback — otherwise a node that accepts
+    # payments would still report empty balances when ``cashu_mints`` is unset.
+    mint_urls: list[str] = list(settings.cashu_mints)
+    if settings.primary_mint and settings.primary_mint not in mint_urls:
+        mint_urls.append(settings.primary_mint)
+
     # Create tasks for all mint/unit combinations
     async with db.create_session() as session:
         tasks = [
             fetch_balance(session, mint_url, unit)
-            for mint_url in settings.cashu_mints
+            for mint_url in mint_urls
             for unit in units
         ]
 
@@ -476,7 +733,9 @@ async def fetch_all_balances(
             total_wallet_balance_sats += proofs_balance_sats
             total_user_balance_sats += user_balance_sats
 
-    owner_balance = total_wallet_balance_sats - total_user_balance_sats
+    owner_balance = 0
+    if total_wallet_balance_sats != 0:
+        owner_balance = total_wallet_balance_sats - total_user_balance_sats
 
     return (
         balance_details,
@@ -487,11 +746,11 @@ async def fetch_all_balances(
 
 
 async def periodic_payout() -> None:
-    if not settings.receive_ln_address:
-        logger.error("RECEIVE_LN_ADDRESS is not set, skipping payout")
-        return
     while True:
-        await asyncio.sleep(60 * 15)
+        await asyncio.sleep(settings.payout_interval_seconds)
+        print(settings.payout_interval_seconds)
+        if not settings.receive_ln_address:
+            continue
         try:
             async with db.create_session() as session:
                 for mint_url in settings.cashu_mints:
@@ -509,7 +768,12 @@ async def periodic_payout() -> None:
                             user_balance = user_balance // 1000
                         proofs_balance = sum(proof.amount for proof in proofs)
                         available_balance = proofs_balance - user_balance
-                        min_amount = 210 if unit == "sat" else 210000
+                        # Threshold is configured in sats; convert for msat wallets.
+                        min_amount = (
+                            settings.min_payout_sat
+                            if unit == "sat"
+                            else settings.min_payout_sat * 1000
+                        )
                         if available_balance > min_amount:
                             amount_received = await raw_send_to_lnurl(
                                 wallet,

tests/integration/conftest.py

@@ -174,17 +174,20 @@ class TestmintWallet:
         """Fallback method to create a basic test token"""
         import base64
         import json
-        import random
-        import time
+        import secrets
 
-        unique_id = int(time.time() * 1000000) + random.randint(1000, 9999)
+        # Use a cryptographically random id/secret so every minted token is
+        # guaranteed unique. Time/PRNG-based ids can collide on hosts with
+        # coarse clock resolution, producing byte-identical tokens that hash to
+        # the same api_key and silently dedupe (flaky concurrency tests).
+        unique_id = secrets.token_hex(16)
         token_data = {
             "token": [
                 {
                     "mint": self.mint_url,
                     "proofs": [
                         {
-                            "id": f"009a1f293253e41e{unique_id % 100000000:08d}",
+                            "id": f"009a1f293253e41e{unique_id[:8]}",
                             "amount": amount,
                             "secret": f"test-secret-{amount}-{unique_id}",
                             "C": "02194603ffa36356f4a56b7df9371fc3192472351453ec7398b8da8117e7c3e104",

tests/integration/test_child_keys.py

@@ -1,3 +1,4 @@
+import asyncio
 import secrets
 from typing import Any
 
@@ -7,7 +8,7 @@ from sqlmodel.ext.asyncio.session import AsyncSession
 
 from routstr.auth import adjust_payment_for_tokens, pay_for_request
 from routstr.balance import ChildKeyRequest, create_child_key
-from routstr.core.db import ApiKey
+from routstr.core.db import ApiKey, create_session
 from routstr.core.settings import settings
 
 
@@ -119,6 +120,54 @@ async def test_child_key_insufficient_balance(
     assert exc.value.status_code == 402
 
 
+@pytest.mark.asyncio
+async def test_concurrent_child_key_creation_is_atomic(
+    patched_db_engine: None,
+) -> None:
+    """Two concurrent create_child_key() calls with balance for exactly one must
+    result in exactly one success and one 402, with the parent balance deducted
+    only once."""
+    child_key_cost = 1000
+    settings.child_key_cost = child_key_cost
+
+    parent_hash = f"parent_concurrent_{secrets.token_hex(8)}"
+    async with create_session() as session:
+        parent = ApiKey(hashed_key=parent_hash, balance=child_key_cost)
+        session.add(parent)
+        await session.commit()
+
+    results: list[str] = []
+
+    async def attempt() -> None:
+        async with create_session() as session:
+            fresh_parent = await session.get(ApiKey, parent_hash)
+            assert fresh_parent is not None
+            try:
+                await create_child_key(ChildKeyRequest(count=1), fresh_parent, session)
+                results.append("success")
+            except HTTPException as exc:
+                assert exc.status_code == 402
+                results.append("blocked")
+
+    await asyncio.gather(attempt(), attempt())
+
+    assert sorted(results) == ["blocked", "success"], (
+        f"Expected exactly one success and one 402, got: {results}"
+    )
+
+    async with create_session() as session:
+        final = await session.get(ApiKey, parent_hash)
+        assert final is not None
+
+    assert final.balance == 0, (
+        f"Balance should be fully deducted once: expected 0, got {final.balance}"
+    )
+    assert final.total_spent == child_key_cost, (
+        f"total_spent should equal one deduction: expected {child_key_cost}, "
+        f"got {final.total_spent}"
+    )
+
+
 @pytest.mark.asyncio
 async def test_child_key_cannot_create_child(integration_session: AsyncSession) -> None:
     parent_key = ApiKey(

tests/integration/test_child_keys_api.py

@@ -69,9 +69,14 @@ async def test_wallet_info_child_key_no_child_keys(
     info_response = await integration_client.get("/v1/wallet/info")
     assert info_response.status_code == 200
     info_data = info_response.json()
+    parent_key = authenticated_client._test_api_key  # type: ignore[attr-defined]
+    parent_key_hash = parent_key.removeprefix("sk-")
 
     assert info_data["is_child"] is True
     assert "child_keys" not in info_data
+    assert "parent_key" not in info_data
+    assert info_data["parent_key_preview"] == parent_key_hash[:8] + "..."
+    assert info_data["parent_key_preview"] not in {parent_key, parent_key_hash}
 
 
 @pytest.mark.integration

tests/integration/test_free_response_stale_reservation.py

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

tests/integration/test_insufficient_balance.py

@@ -124,6 +124,40 @@ async def test_pay_for_request_raises_402_when_all_balance_reserved(
     assert key.reserved_balance == 50_000
 
 
+@pytest.mark.asyncio
+async def test_balance_info_matches_chat_available_balance(
+    integration_client: AsyncClient,
+    integration_session: AsyncSession,
+) -> None:
+    """
+    Regression for /v1/balance/info showing gross funds while chat admission
+    rejects with a negative available balance.
+    """
+    from routstr.auth import pay_for_request
+
+    key = _key(balance=4_404_339, reserved=4_410_636)
+    integration_session.add(key)
+    await integration_session.commit()
+
+    response = await integration_client.get(
+        "/v1/balance/info",
+        headers={"Authorization": f"Bearer sk-{key.hashed_key}"},
+    )
+
+    assert response.status_code == 200
+    body = response.json()
+    assert body["balance"] == -6_297
+    assert body["reserved"] == 4_410_636
+
+    with pytest.raises(HTTPException) as exc_info:
+        await pay_for_request(key, 1, integration_session)
+
+    assert exc_info.value.status_code == 402
+    detail = exc_info.value.detail
+    assert isinstance(detail, dict)
+    assert "-6297 available" in detail["error"]["message"]
+
+
 # ---------------------------------------------------------------------------
 # Test 4 — balance just one msat below model cost
 # ---------------------------------------------------------------------------

tests/integration/test_key_logic.py

@@ -1,12 +1,14 @@
+import asyncio
 import time
 from datetime import datetime, timedelta
 
 import pytest
+from fastapi import HTTPException
 from sqlmodel import select
 from sqlmodel.ext.asyncio.session import AsyncSession
 
 from routstr.auth import pay_for_request
-from routstr.core.db import ApiKey
+from routstr.core.db import ApiKey, create_session
 
 
 @pytest.mark.asyncio
@@ -120,6 +122,270 @@ async def test_periodic_key_reset_job(integration_session: AsyncSession) -> None
     assert key2.total_spent == 0
 
 
+@pytest.mark.asyncio
+async def test_balance_limit_enforced_atomically_under_concurrency(
+    patched_db_engine: None,
+) -> None:
+    parent_hash = "parent_limit_atomic"
+    child_hash = "child_limit_atomic"
+    cost = 300
+
+    async with create_session() as session:
+        parent = ApiKey(hashed_key=parent_hash, balance=10000)
+        child = ApiKey(
+            hashed_key=child_hash,
+            balance=0,
+            parent_key_hash=parent_hash,
+            balance_limit=cost,
+            total_spent=0,
+        )
+        session.add(parent)
+        session.add(child)
+        await session.commit()
+
+    results: list[str] = []
+
+    async def attempt() -> None:
+        async with create_session() as session:
+            fresh_child = await session.get(ApiKey, child_hash)
+            assert fresh_child is not None
+            try:
+                await pay_for_request(fresh_child, cost, session)
+                results.append("success")
+            except HTTPException as exc:
+                assert exc.status_code == 402
+                results.append("blocked")
+
+    await asyncio.gather(attempt(), attempt())
+
+    assert sorted(results) == ["blocked", "success"], (
+        f"Expected exactly one success and one 402, got: {results}"
+    )
+
+    async with create_session() as session:
+        final_child = await session.get(ApiKey, child_hash)
+        assert final_child is not None
+
+    assert final_child.reserved_balance == cost, (
+        f"Child reserved_balance should equal one reservation, "
+        f"got {final_child.reserved_balance}"
+    )
+
+
+@pytest.mark.asyncio
+async def test_parallel_payments_with_parent_and_child_key(
+    patched_db_engine: None,
+) -> None:
+    parent_hash = "parent_parallel_mixed"
+    child_hash = "child_parallel_mixed"
+    cost = 300
+
+    async with create_session() as session:
+        parent = ApiKey(hashed_key=parent_hash, balance=10000)
+        child = ApiKey(
+            hashed_key=child_hash,
+            balance=0,
+            parent_key_hash=parent_hash,
+            balance_limit=2 * cost,
+        )
+        session.add(parent)
+        session.add(child)
+        await session.commit()
+
+    async def attempt(key_hash: str) -> str:
+        async with create_session() as session:
+            fresh_key = await session.get(ApiKey, key_hash)
+            assert fresh_key is not None
+            try:
+                await pay_for_request(fresh_key, cost, session)
+                return "success"
+            except HTTPException as exc:
+                assert exc.status_code == 402
+                return "blocked"
+
+    results = await asyncio.gather(attempt(parent_hash), attempt(child_hash))
+
+    assert results == ["success", "success"], (
+        f"Both parent and child payments should succeed, got: {results}"
+    )
+
+    async with create_session() as session:
+        final_parent = await session.get(ApiKey, parent_hash)
+        final_child = await session.get(ApiKey, child_hash)
+        assert final_parent is not None
+        assert final_child is not None
+
+    # Both requests bill the parent; only the child request reserves on the child.
+    assert final_parent.reserved_balance == 2 * cost
+    assert final_parent.total_requests == 2
+    assert final_child.reserved_balance == cost
+    assert final_child.total_requests == 1
+
+
+@pytest.mark.asyncio
+async def test_balance_limit_with_existing_total_spent_under_concurrency(
+    patched_db_engine: None,
+) -> None:
+    parent_hash = "parent_total_spent"
+    child_hash = "child_total_spent"
+    cost = 300
+
+    async with create_session() as session:
+        parent = ApiKey(hashed_key=parent_hash, balance=10000)
+        # 700 already spent against a 1000 limit: only one more 300 request fits.
+        child = ApiKey(
+            hashed_key=child_hash,
+            balance=0,
+            parent_key_hash=parent_hash,
+            balance_limit=1000,
+            total_spent=700,
+        )
+        session.add(parent)
+        session.add(child)
+        await session.commit()
+
+    results: list[str] = []
+
+    async def attempt() -> None:
+        async with create_session() as session:
+            fresh_child = await session.get(ApiKey, child_hash)
+            assert fresh_child is not None
+            try:
+                await pay_for_request(fresh_child, cost, session)
+                results.append("success")
+            except HTTPException as exc:
+                assert exc.status_code == 402
+                results.append("blocked")
+
+    await asyncio.gather(attempt(), attempt())
+
+    assert sorted(results) == ["blocked", "success"], (
+        f"Expected exactly one success and one 402, got: {results}"
+    )
+
+    async with create_session() as session:
+        final_child = await session.get(ApiKey, child_hash)
+        assert final_child is not None
+
+    assert final_child.reserved_balance == cost
+    assert final_child.total_spent == 700
+
+
+@pytest.mark.asyncio
+async def test_balance_limit_with_existing_reserved_balance(
+    patched_db_engine: None,
+) -> None:
+    parent_hash = "parent_reserved_set"
+    blocked_hash = "child_reserved_blocked"
+    allowed_hash = "child_reserved_allowed"
+    cost = 300
+
+    async with create_session() as session:
+        parent = ApiKey(hashed_key=parent_hash, balance=10000)
+        # 800 already reserved against a 1000 limit: another 300 must be rejected.
+        blocked_child = ApiKey(
+            hashed_key=blocked_hash,
+            balance=0,
+            parent_key_hash=parent_hash,
+            balance_limit=1000,
+            reserved_balance=800,
+        )
+        # 500 reserved against a 1000 limit: another 300 still fits.
+        allowed_child = ApiKey(
+            hashed_key=allowed_hash,
+            balance=0,
+            parent_key_hash=parent_hash,
+            balance_limit=1000,
+            reserved_balance=500,
+        )
+        session.add(parent)
+        session.add(blocked_child)
+        session.add(allowed_child)
+        await session.commit()
+
+    async with create_session() as session:
+        fresh_blocked = await session.get(ApiKey, blocked_hash)
+        assert fresh_blocked is not None
+        with pytest.raises(HTTPException) as exc_info:
+            await pay_for_request(fresh_blocked, cost, session)
+        assert exc_info.value.status_code == 402
+
+    async with create_session() as session:
+        fresh_allowed = await session.get(ApiKey, allowed_hash)
+        assert fresh_allowed is not None
+        await pay_for_request(fresh_allowed, cost, session)
+
+    async with create_session() as session:
+        final_blocked = await session.get(ApiKey, blocked_hash)
+        final_allowed = await session.get(ApiKey, allowed_hash)
+        final_parent = await session.get(ApiKey, parent_hash)
+        assert final_blocked is not None
+        assert final_allowed is not None
+        assert final_parent is not None
+
+    assert final_blocked.reserved_balance == 800, "Rejected request must not reserve"
+    assert final_blocked.total_requests == 0
+    assert final_allowed.reserved_balance == 500 + cost
+    assert final_allowed.total_requests == 1
+    # Only the allowed request should have billed the parent.
+    assert final_parent.reserved_balance == cost
+    assert final_parent.total_requests == 1
+
+
+@pytest.mark.asyncio
+async def test_child_reservation_discarded_when_parent_balance_depleted(
+    patched_db_engine: None,
+) -> None:
+    parent_hash = "parent_depleted"
+    child_hash = "child_depleted"
+    cost = 300
+
+    async with create_session() as session:
+        # Parent can only afford one request; child has no balance_limit.
+        parent = ApiKey(hashed_key=parent_hash, balance=cost)
+        child = ApiKey(
+            hashed_key=child_hash,
+            balance=0,
+            parent_key_hash=parent_hash,
+        )
+        session.add(parent)
+        session.add(child)
+        await session.commit()
+
+    results: list[str] = []
+
+    async def attempt() -> None:
+        async with create_session() as session:
+            fresh_child = await session.get(ApiKey, child_hash)
+            assert fresh_child is not None
+            try:
+                await pay_for_request(fresh_child, cost, session)
+                results.append("success")
+            except HTTPException as exc:
+                assert exc.status_code == 402
+                results.append("blocked")
+
+    await asyncio.gather(attempt(), attempt())
+
+    assert sorted(results) == ["blocked", "success"], (
+        f"Expected exactly one success and one 402, got: {results}"
+    )
+
+    async with create_session() as session:
+        final_parent = await session.get(ApiKey, parent_hash)
+        final_child = await session.get(ApiKey, child_hash)
+        assert final_parent is not None
+        assert final_child is not None
+
+    assert final_parent.reserved_balance == cost
+    # The failed request must not leave a committed reservation on the child.
+    assert final_child.reserved_balance == cost, (
+        f"Child reserved_balance should reflect only the successful request, "
+        f"got {final_child.reserved_balance}"
+    )
+    assert final_child.total_requests == 1
+
+
 @pytest.mark.asyncio
 async def test_refund_does_not_delete_key(integration_session: AsyncSession) -> None:
     # This requires mocking the router call or testing the logic in balance.py

tests/integration/test_lightning_invoice_constraints.py

@@ -0,0 +1,159 @@
+"""Integration tests for Lightning invoice key constraint fields.
+
+Covers two things:
+- The three constraint fields (balance_limit, balance_limit_reset, validity_date)
+  are persisted on LightningInvoice and survive a DB round-trip.
+- create_api_key_from_invoice propagates those fields to the created ApiKey,
+  so the constraints are actually enforced when the key is used.
+"""
+
+from __future__ import annotations
+
+import time
+from unittest.mock import AsyncMock, patch
+
+import pytest
+from sqlmodel.ext.asyncio.session import AsyncSession
+
+from routstr.core.db import ApiKey, LightningInvoice
+from routstr.lightning import create_api_key_from_invoice
+
+
+def _make_invoice(**kwargs: object) -> LightningInvoice:
+    base = dict(
+        id="inv_test_001",
+        bolt11="lnbc1000n1test",
+        amount_sats=1000,
+        description="test invoice",
+        payment_hash="deadbeef" * 8,
+        status="paid",
+        purpose="create",
+        expires_at=int(time.time()) + 3600,
+        paid_at=int(time.time()),
+    )
+    base.update(kwargs)
+    return LightningInvoice(**base)  # type: ignore[arg-type]
+
+
+@pytest.fixture(autouse=True)
+def mock_wallet_mint() -> object:
+    with patch("routstr.lightning.get_wallet") as mock_get_wallet:
+        wallet = AsyncMock()
+        wallet.mint = AsyncMock(return_value=[])
+        mock_get_wallet.return_value = wallet
+        yield mock_get_wallet
+
+
+# ---------------------------------------------------------------------------
+# Persistence
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_invoice_persists_balance_limit(
+    integration_session: AsyncSession,
+) -> None:
+    invoice = _make_invoice(balance_limit=5000)
+    integration_session.add(invoice)
+    await integration_session.commit()
+
+    stored = await integration_session.get(LightningInvoice, invoice.id)
+    assert stored is not None
+    assert stored.balance_limit == 5000
+
+
+@pytest.mark.asyncio
+async def test_invoice_persists_balance_limit_reset(
+    integration_session: AsyncSession,
+) -> None:
+    invoice = _make_invoice(balance_limit=5000, balance_limit_reset="daily")
+    integration_session.add(invoice)
+    await integration_session.commit()
+
+    stored = await integration_session.get(LightningInvoice, invoice.id)
+    assert stored is not None
+    assert stored.balance_limit_reset == "daily"
+
+
+@pytest.mark.asyncio
+async def test_invoice_persists_validity_date(
+    integration_session: AsyncSession,
+) -> None:
+    expiry = int(time.time()) + 86400
+    invoice = _make_invoice(validity_date=expiry)
+    integration_session.add(invoice)
+    await integration_session.commit()
+
+    stored = await integration_session.get(LightningInvoice, invoice.id)
+    assert stored is not None
+    assert stored.validity_date == expiry
+
+
+# ---------------------------------------------------------------------------
+# Propagation to ApiKey
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_created_key_receives_balance_limit(
+    integration_session: AsyncSession,
+) -> None:
+    invoice = _make_invoice(balance_limit=8000)
+    integration_session.add(invoice)
+    await integration_session.flush()
+
+    api_key = await create_api_key_from_invoice(invoice, integration_session)
+    await integration_session.commit()
+
+    stored_key = await integration_session.get(ApiKey, api_key.hashed_key)
+    assert stored_key is not None
+    assert stored_key.balance_limit == 8000
+
+
+@pytest.mark.asyncio
+async def test_created_key_receives_balance_limit_reset(
+    integration_session: AsyncSession,
+) -> None:
+    invoice = _make_invoice(balance_limit=8000, balance_limit_reset="monthly")
+    integration_session.add(invoice)
+    await integration_session.flush()
+
+    api_key = await create_api_key_from_invoice(invoice, integration_session)
+    await integration_session.commit()
+
+    stored_key = await integration_session.get(ApiKey, api_key.hashed_key)
+    assert stored_key is not None
+    assert stored_key.balance_limit_reset == "monthly"
+
+
+@pytest.mark.asyncio
+async def test_created_key_receives_validity_date(
+    integration_session: AsyncSession,
+) -> None:
+    expiry = int(time.time()) + 86400
+    invoice = _make_invoice(validity_date=expiry)
+    integration_session.add(invoice)
+    await integration_session.flush()
+
+    api_key = await create_api_key_from_invoice(invoice, integration_session)
+    await integration_session.commit()
+
+    stored_key = await integration_session.get(ApiKey, api_key.hashed_key)
+    assert stored_key is not None
+    assert stored_key.validity_date == expiry
+
+
+@pytest.mark.asyncio
+async def test_created_key_without_constraints_has_none_fields(
+    integration_session: AsyncSession,
+) -> None:
+    invoice = _make_invoice()
+    integration_session.add(invoice)
+    await integration_session.flush()
+
+    api_key = await create_api_key_from_invoice(invoice, integration_session)
+    await integration_session.commit()
+
+    stored_key = await integration_session.get(ApiKey, api_key.hashed_key)
+    assert stored_key is not None
+    assert stored_key.balance_limit is None
+    assert stored_key.balance_limit_reset is None
+    assert stored_key.validity_date is None

tests/integration/test_lightning_invoice_rip08.py

@@ -0,0 +1,198 @@
+"""RIP-08 lightning invoice endpoint tests.
+
+Verifies both the spec-compliant path (`POST /lightning/invoice` with
+`Authorization: Bearer sk-...`) and the legacy path
+(`POST /v1/balance/lightning/invoice` with `api_key` in body).
+"""
+
+from __future__ import annotations
+
+from typing import Any
+from unittest.mock import patch
+
+import pytest
+import pytest_asyncio
+from httpx import AsyncClient
+from sqlmodel.ext.asyncio.session import AsyncSession
+
+from routstr.core.db import ApiKey
+
+RIP08_PATH = "/lightning/invoice"
+LEGACY_PATH = "/v1/balance/lightning/invoice"
+
+
+@pytest_asyncio.fixture
+async def patch_invoice_generation() -> Any:
+    """Stub out `generate_lightning_invoice` so no mint round-trip is needed."""
+    counter = {"n": 0}
+
+    async def fake_generate(amount_sats: int, description: str) -> tuple[str, str]:
+        counter["n"] += 1
+        return (
+            f"lnbc{amount_sats}n1pfakeinvoice{counter['n']}",
+            f"payment_hash_{counter['n']}",
+        )
+
+    with patch(
+        "routstr.lightning.generate_lightning_invoice",
+        side_effect=fake_generate,
+    ) as m:
+        yield m
+
+
+@pytest_asyncio.fixture
+async def seeded_topup_key(integration_session: AsyncSession) -> str:
+    """Insert an ApiKey row and return the public `sk-...` form."""
+    hashed = "0" * 64
+    key = ApiKey(
+        hashed_key=hashed,
+        balance=0,
+        refund_currency="sat",
+        refund_mint_url="http://localhost:3338",
+    )
+    integration_session.add(key)
+    await integration_session.commit()
+    return f"sk-{hashed}"
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+@pytest.mark.parametrize("path", [RIP08_PATH, LEGACY_PATH])
+async def test_create_invoice_purpose_create(
+    integration_client: AsyncClient,
+    patch_invoice_generation: Any,
+    path: str,
+) -> None:
+    """`purpose=create` works on both paths and requires no auth."""
+    resp = await integration_client.post(
+        path,
+        json={"amount_sats": 1000, "purpose": "create"},
+    )
+    assert resp.status_code == 200, resp.text
+    body = resp.json()
+    assert body["amount_sats"] == 1000
+    assert body["bolt11"].startswith("lnbc")
+    assert body["invoice_id"]
+    assert body["payment_hash"]
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+@pytest.mark.parametrize("path", [RIP08_PATH, LEGACY_PATH])
+async def test_topup_with_authorization_header(
+    integration_client: AsyncClient,
+    patch_invoice_generation: Any,
+    seeded_topup_key: str,
+    path: str,
+) -> None:
+    """RIP-08: topup using `Authorization: Bearer sk-...` header (no api_key in body)."""
+    resp = await integration_client.post(
+        path,
+        json={"amount_sats": 500, "purpose": "topup"},
+        headers={"Authorization": f"Bearer {seeded_topup_key}"},
+    )
+    assert resp.status_code == 200, resp.text
+    body = resp.json()
+    assert body["amount_sats"] == 500
+    assert body["bolt11"].startswith("lnbc")
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+@pytest.mark.parametrize("path", [RIP08_PATH, LEGACY_PATH])
+async def test_topup_with_legacy_api_key_in_body(
+    integration_client: AsyncClient,
+    patch_invoice_generation: Any,
+    seeded_topup_key: str,
+    path: str,
+) -> None:
+    """Legacy: topup with `api_key` in body still accepted on both paths."""
+    resp = await integration_client.post(
+        path,
+        json={
+            "amount_sats": 250,
+            "purpose": "topup",
+            "api_key": seeded_topup_key,
+        },
+    )
+    assert resp.status_code == 200, resp.text
+    assert resp.json()["amount_sats"] == 250
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+@pytest.mark.parametrize("path", [RIP08_PATH, LEGACY_PATH])
+async def test_topup_missing_auth_returns_401(
+    integration_client: AsyncClient,
+    patch_invoice_generation: Any,
+    path: str,
+) -> None:
+    """Topup without any credential is rejected on both paths."""
+    resp = await integration_client.post(
+        path,
+        json={"amount_sats": 100, "purpose": "topup"},
+    )
+    assert resp.status_code == 401
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+@pytest.mark.parametrize("path", [RIP08_PATH, LEGACY_PATH])
+async def test_topup_unknown_api_key_returns_404(
+    integration_client: AsyncClient,
+    patch_invoice_generation: Any,
+    path: str,
+) -> None:
+    resp = await integration_client.post(
+        path,
+        json={"amount_sats": 100, "purpose": "topup"},
+        headers={"Authorization": "Bearer sk-deadbeef"},
+    )
+    assert resp.status_code == 404
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+@pytest.mark.parametrize("path", [RIP08_PATH, LEGACY_PATH])
+async def test_invoice_status_404_for_unknown_id(
+    integration_client: AsyncClient,
+    path: str,
+) -> None:
+    base = path.rsplit("/invoice", 1)[0] + "/invoice"
+    resp = await integration_client.get(f"{base}/does-not-exist/status")
+    assert resp.status_code == 404
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_purpose_defaults_to_create(
+    integration_client: AsyncClient,
+    patch_invoice_generation: Any,
+) -> None:
+    """Per RIP-08, `purpose` may be omitted and defaults to `create`."""
+    resp = await integration_client.post(
+        RIP08_PATH,
+        json={"amount_sats": 100},
+    )
+    assert resp.status_code == 200, resp.text
+    assert resp.json()["amount_sats"] == 100
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_authorization_header_overrides_body_api_key(
+    integration_client: AsyncClient,
+    patch_invoice_generation: Any,
+    seeded_topup_key: str,
+) -> None:
+    """Header api_key wins over body api_key: bogus body must not cause 404."""
+    resp = await integration_client.post(
+        RIP08_PATH,
+        json={
+            "amount_sats": 100,
+            "purpose": "topup",
+            "api_key": "sk-" + "f" * 64,  # bogus body key
+        },
+        headers={"Authorization": f"Bearer {seeded_topup_key}"},
+    )
+    assert resp.status_code == 200, resp.text

tests/integration/test_model_test_endpoint_security.py

@@ -0,0 +1,224 @@
+import time
+from types import TracebackType
+from typing import Any
+from unittest.mock import patch
+
+import pytest
+from httpx import AsyncClient
+
+from routstr.core.admin import admin_sessions
+from routstr.core.db import ModelRow, UpstreamProviderRow
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_model_test_endpoint_requires_admin_auth(
+    integration_client: AsyncClient,
+) -> None:
+    with patch("httpx.AsyncClient") as mock_async_client:
+        response = await integration_client.post(
+            "/api/models/test",
+            json={
+                "model_id": "model-a",
+                "endpoint_type": "chat-completions",
+                "request_data": {"messages": []},
+            },
+        )
+
+    assert response.status_code == 403
+    mock_async_client.assert_not_called()
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_model_test_endpoint_rejects_unsupported_endpoint_type(
+    integration_client: AsyncClient,
+    integration_session: Any,
+) -> None:
+    admin_token = "test-admin-token-model-test"
+    admin_sessions[admin_token] = int(time.time()) + 3600
+    integration_client.headers["Authorization"] = f"Bearer {admin_token}"
+
+    provider = UpstreamProviderRow(
+        provider_type="custom",
+        base_url="https://api.example.com/v1",
+        api_key="sk-upstream-test",
+        enabled=True,
+        provider_fee=1.01,
+    )
+    integration_session.add(provider)
+    await integration_session.commit()
+    await integration_session.refresh(provider)
+    assert provider.id is not None
+
+    model = ModelRow(
+        id="model-a",
+        name="Model A",
+        created=1,
+        description="desc",
+        context_length=100,
+        architecture='{"modality": "text", "input_modalities": ["text"], "output_modalities": ["text"], "tokenizer": "tiktoken", "instruct_type": "chat"}',
+        pricing='{"prompt": 1.0, "completion": 1.0}',
+        upstream_provider_id=provider.id,
+        enabled=True,
+    )
+    integration_session.add(model)
+    await integration_session.commit()
+
+    try:
+        with patch("httpx.AsyncClient") as mock_async_client:
+            response = await integration_client.post(
+                "/api/models/test",
+                json={
+                    "model_id": "model-a",
+                    "endpoint_type": "../../abuse",
+                    "request_data": {"messages": []},
+                },
+            )
+
+        assert response.status_code == 400
+        assert response.json()["detail"] == "Unsupported endpoint_type"
+        mock_async_client.assert_not_called()
+    finally:
+        admin_sessions.pop(admin_token, None)
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_model_test_endpoint_rejects_oversized_request_data(
+    integration_client: AsyncClient,
+    integration_session: Any,
+) -> None:
+    admin_token = "test-admin-token-model-test-oversized"
+    admin_sessions[admin_token] = int(time.time()) + 3600
+    integration_client.headers["Authorization"] = f"Bearer {admin_token}"
+
+    provider = UpstreamProviderRow(
+        provider_type="custom",
+        base_url="https://api.example.com/v1",
+        api_key="sk-upstream-test",
+        enabled=True,
+        provider_fee=1.01,
+    )
+    integration_session.add(provider)
+    await integration_session.commit()
+    await integration_session.refresh(provider)
+    assert provider.id is not None
+
+    model = ModelRow(
+        id="model-a",
+        name="Model A",
+        created=1,
+        description="desc",
+        context_length=100,
+        architecture='{"modality": "text", "input_modalities": ["text"], "output_modalities": ["text"], "tokenizer": "tiktoken", "instruct_type": "chat"}',
+        pricing='{"prompt": 1.0, "completion": 1.0}',
+        upstream_provider_id=provider.id,
+        enabled=True,
+    )
+    integration_session.add(model)
+    await integration_session.commit()
+
+    oversized = "x" * (64 * 1024 + 1)
+
+    try:
+        with patch("httpx.AsyncClient") as mock_async_client:
+            response = await integration_client.post(
+                "/api/models/test",
+                json={
+                    "model_id": "model-a",
+                    "endpoint_type": "chat-completions",
+                    "request_data": {"blob": oversized},
+                },
+            )
+
+        assert response.status_code == 413
+        assert response.json()["detail"] == "request_data too large"
+        mock_async_client.assert_not_called()
+    finally:
+        admin_sessions.pop(admin_token, None)
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_model_test_endpoint_admin_uses_allowed_upstream_path(
+    integration_client: AsyncClient,
+    integration_session: Any,
+) -> None:
+    admin_token = "test-admin-token-model-test-success"
+    admin_sessions[admin_token] = int(time.time()) + 3600
+    integration_client.headers["Authorization"] = f"Bearer {admin_token}"
+
+    provider = UpstreamProviderRow(
+        provider_type="custom",
+        base_url="https://api.example.com/v1",
+        api_key="sk-upstream-test",
+        enabled=True,
+        provider_fee=1.01,
+    )
+    integration_session.add(provider)
+    await integration_session.commit()
+    await integration_session.refresh(provider)
+    assert provider.id is not None
+
+    model = ModelRow(
+        id="model-a",
+        name="Model A",
+        created=1,
+        description="desc",
+        context_length=100,
+        architecture='{"modality": "text", "input_modalities": ["text"], "output_modalities": ["text"], "tokenizer": "tiktoken", "instruct_type": "chat"}',
+        pricing='{"prompt": 1.0, "completion": 1.0}',
+        upstream_provider_id=provider.id,
+        enabled=True,
+        forwarded_model_id="upstream-model-a",
+    )
+    integration_session.add(model)
+    await integration_session.commit()
+
+    class MockResponse:
+        status_code = 200
+        text = '{"ok": true}'
+
+        def json(self) -> dict[str, bool]:
+            return {"ok": True}
+
+    class MockAsyncClient:
+        async def __aenter__(self) -> "MockAsyncClient":
+            return self
+
+        async def __aexit__(
+            self,
+            exc_type: type[BaseException] | None,
+            exc: BaseException | None,
+            tb: TracebackType | None,
+        ) -> None:
+            return None
+
+        async def post(
+            self, url: str, json: dict[str, Any], headers: dict[str, str]
+        ) -> MockResponse:
+            assert url == "https://api.example.com/v1/chat/completions"
+            assert json["model"] == "upstream-model-a"
+            assert headers["Authorization"] == "Bearer sk-upstream-test"
+            return MockResponse()
+
+    try:
+        with patch("httpx.AsyncClient", return_value=MockAsyncClient()):
+            response = await integration_client.post(
+                "/api/models/test",
+                json={
+                    "model_id": "model-a",
+                    "endpoint_type": "chat-completions",
+                    "request_data": {"messages": []},
+                },
+            )
+
+        assert response.status_code == 200
+        assert response.json() == {
+            "success": True,
+            "data": {"ok": True},
+            "status_code": 200,
+        }
+    finally:
+        admin_sessions.pop(admin_token, None)

tests/integration/test_prune_dead_api_keys.py

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

tests/integration/test_swap_fee_retry.py

@@ -0,0 +1,194 @@
+"""
+Integration tests for reactive swap fee retries via the wallet topup endpoint.
+
+Foreign-mint tokens are swapped to the primary mint using the foreign mint's
+melt quote, whose fee_reserve is a non-binding estimate (NUT-05): the mint may
+demand more when re-quoting or at melt execution. These tests cover the
+endpoint behaviour in those cases:
+
+1. The mint demands one sat more at melt time than every quote reported
+   (the mint.cubabitcoin.org incident): the swap retries with a smaller
+   invoice and the topup succeeds, crediting the recomputed amount.
+2. The real melt quote reports a higher fee_reserve than the estimate: the
+   swap re-quotes from the observed fee and the topup succeeds.
+3. The mint escalates its fee demands on every attempt: the retry budget is
+   exhausted and the endpoint returns 400 with a clear error (never 500),
+   without ever executing a melt.
+"""
+
+from collections.abc import Callable
+from unittest.mock import AsyncMock, Mock, patch
+
+import pytest
+from httpx import AsyncClient, Response
+
+from routstr.core.settings import settings
+
+# Captured at collection time, before the integration_app fixture replaces it
+# with the testmint stub that bypasses swapping (see conftest.py).
+from routstr.wallet import recieve_token as _real_recieve_token
+
+PRIMARY_MINT = "http://primary:3338"
+
+
+def _make_swap_mocks(
+    token_amount: int,
+    fee_reserves: list[int],
+    input_fees: int = 0,
+    mint_url: str = "http://foreign-mint:3338",
+) -> tuple[Mock, Mock, Mock]:
+    """Return (token, token_wallet, primary_wallet) mocks that act like a mint.
+
+    Mint quotes pass the requested amount through their ``request`` field and
+    melt quotes echo that amount back, so the mocks stay consistent for
+    whatever amounts the implementation requests. ``fee_reserves`` supplies the
+    fee_reserve of each successive melt quote (the first serves the estimation
+    pass); requesting more quotes than provided fails the test.
+    """
+    mock_token = Mock()
+    mock_token.mint = mint_url
+    mock_token.unit = "sat"
+    mock_token.amount = token_amount
+    mock_token.keysets = ["keyset1"]
+    mock_token.proofs = [Mock(amount=token_amount)]
+
+    mock_token_wallet = Mock()
+    mock_token_wallet.load_mint = AsyncMock()
+    mock_token_wallet.load_proofs = AsyncMock()
+    mock_token_wallet.get_fees_for_proofs = Mock(return_value=input_fees)
+
+    mock_primary_wallet = Mock()
+    mock_primary_wallet.load_mint = AsyncMock()
+    mock_primary_wallet.load_proofs = AsyncMock()
+    mock_primary_wallet.available_balance = Mock(amount=0)
+    mock_primary_wallet.mint = AsyncMock(return_value=Mock())
+
+    fees = iter(fee_reserves)
+
+    def _next_fee() -> int:
+        try:
+            return next(fees)
+        except StopIteration:
+            raise AssertionError(
+                "more melt quotes requested than fee_reserves provided"
+            ) from None
+
+    mock_primary_wallet.request_mint = AsyncMock(
+        side_effect=lambda amount: Mock(quote=f"mint_quote_{amount}", request=amount)
+    )
+    mock_token_wallet.melt_quote = AsyncMock(
+        side_effect=lambda invoice: Mock(
+            quote=f"melt_quote_{invoice}", amount=invoice, fee_reserve=_next_fee()
+        )
+    )
+    mock_token_wallet.melt = AsyncMock(return_value=Mock())
+
+    return mock_token, mock_token_wallet, mock_primary_wallet
+
+
+def _wallet_router(primary_wallet: Mock, token_wallet: Mock) -> Callable[..., Mock]:
+    """Route get_wallet calls to the primary or foreign wallet mock by URL."""
+
+    def fake_get_wallet(mint_url: str, unit: str = "sat", load: bool = True) -> Mock:
+        return primary_wallet if mint_url == PRIMARY_MINT else token_wallet
+
+    return fake_get_wallet
+
+
+async def _post_topup(
+    client: AsyncClient,
+    mock_token: Mock,
+    token_wallet: Mock,
+    primary_wallet: Mock,
+) -> Response:
+    """POST /v1/wallet/topup with the swap layer mocked at the mint boundary.
+
+    The conftest's testmint stub for recieve_token is swapped back for the
+    real implementation so the request exercises the actual swap path.
+    """
+    with patch("routstr.wallet.recieve_token", _real_recieve_token):
+        with patch(
+            "routstr.wallet.deserialize_token_from_string", return_value=mock_token
+        ):
+            with patch(
+                "routstr.wallet.get_wallet",
+                side_effect=_wallet_router(primary_wallet, token_wallet),
+            ):
+                with patch.object(settings, "primary_mint", PRIMARY_MINT):
+                    with patch.object(settings, "primary_mint_unit", "sat"):
+                        with patch.object(settings, "cashu_mints", [PRIMARY_MINT]):
+                            return await client.post(
+                                "/v1/wallet/topup",
+                                params={"cashu_token": "cashuAtest_foreign_token"},
+                            )
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_topup_retries_when_melt_demands_more_than_quoted(
+    authenticated_client: AsyncClient,
+) -> None:
+    """A 179-sat token where every quote reports fee_reserve=1 but the mint
+    rejects the first melt demanding 180. The retry shrinks the invoice to 177
+    and the topup credits 177 sats (177_000 msats)."""
+    mock_token, token_wallet, primary_wallet = _make_swap_mocks(
+        179, fee_reserves=[1, 1, 1], mint_url="http://mint.cubabitcoin.org"
+    )
+    token_wallet.melt.side_effect = [
+        Exception(
+            "Mint Error: not enough inputs provided for melt. "
+            "Provided: 179, needed: 180 (Code: 11000)"
+        ),
+        Mock(),
+    ]
+
+    response = await _post_topup(
+        authenticated_client, mock_token, token_wallet, primary_wallet
+    )
+
+    assert response.status_code == 200
+    assert response.json()["msats"] == 177_000
+    assert token_wallet.melt.call_count == 2
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_topup_retries_when_quote_fee_exceeds_estimate(
+    authenticated_client: AsyncClient,
+) -> None:
+    """A 1000-sat token estimated at fee 20, but the real quote demands 23.
+    The retry recomputes 1000 - 23 = 977, which fits, and the topup credits
+    977 sats (977_000 msats) with a single melt."""
+    mock_token, token_wallet, primary_wallet = _make_swap_mocks(
+        1000, fee_reserves=[20, 23, 23]
+    )
+
+    response = await _post_topup(
+        authenticated_client, mock_token, token_wallet, primary_wallet
+    )
+
+    assert response.status_code == 200
+    assert response.json()["msats"] == 977_000
+    assert token_wallet.melt.call_count == 1
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_topup_returns_400_when_retries_exhausted(
+    authenticated_client: AsyncClient,
+) -> None:
+    """A mint that escalates fee_reserve on every re-quote (1 → 10 → 25 → 50)
+    exhausts the retry budget: clean 400 with an actionable message, melt never
+    executed."""
+    mock_token, token_wallet, primary_wallet = _make_swap_mocks(
+        1000, fee_reserves=[1, 10, 25, 50]
+    )
+
+    response = await _post_topup(
+        authenticated_client, mock_token, token_wallet, primary_wallet
+    )
+
+    assert response.status_code == 400
+    assert "too small to cover swap fees" in response.json()["detail"]
+    assert token_wallet.melt_quote.call_count == 4  # estimation + 3 attempts
+    token_wallet.melt.assert_not_called()

tests/integration/test_temporary_balances_api.py

@@ -0,0 +1,210 @@
+from datetime import datetime, timedelta, timezone
+
+import httpx
+import pytest
+from sqlmodel import col, update
+from sqlmodel.ext.asyncio.session import AsyncSession
+
+from routstr.core.admin import admin_sessions
+from routstr.core.db import ApiKey
+
+
+def _admin_headers() -> dict[str, str]:
+    token = "test-admin-token"
+    admin_sessions[token] = int(
+        (datetime.now(timezone.utc) + timedelta(minutes=5)).timestamp()
+    )
+    return {"Authorization": f"Bearer {token}"}
+
+
+async def _add_key(
+    session: AsyncSession,
+    hashed_key: str,
+    *,
+    balance: int = 0,
+    total_spent: int = 0,
+    total_requests: int = 0,
+    created_at: int | None = None,
+    parent_key_hash: str | None = None,
+    refund_address: str | None = None,
+) -> ApiKey:
+    key = ApiKey(
+        hashed_key=hashed_key,
+        balance=balance,
+        total_spent=total_spent,
+        total_requests=total_requests,
+        parent_key_hash=parent_key_hash,
+        refund_address=refund_address,
+    )
+    key.created_at = created_at
+    session.add(key)
+    await session.commit()
+
+    # The model's default_factory is translated into a SQLAlchemy column
+    # default that fires on INSERT whenever the value is None, so a true NULL
+    # (a legacy row created before the column existed) can only be produced by
+    # an explicit UPDATE after insert.
+    if created_at is None:
+        await session.exec(
+            update(ApiKey)  # type: ignore[call-overload]
+            .where(col(ApiKey.hashed_key) == hashed_key)
+            .values(created_at=None)
+        )
+        await session.commit()
+    return key
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_temporary_balances_envelope_and_created_at(
+    integration_client: httpx.AsyncClient,
+    integration_session: AsyncSession,
+) -> None:
+    await _add_key(integration_session, "key_a", balance=1000, created_at=1000)
+
+    response = await integration_client.get(
+        "/admin/api/temporary-balances", headers=_admin_headers()
+    )
+
+    assert response.status_code == 200
+    body = response.json()
+    assert set(body.keys()) == {"balances", "total", "totals"}
+    assert body["total"] == 1
+    assert body["balances"][0]["hashed_key"] == "key_a"
+    assert body["balances"][0]["created_at"] == 1000
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_temporary_balances_sorted_latest_first_nulls_last(
+    integration_client: httpx.AsyncClient,
+    integration_session: AsyncSession,
+) -> None:
+    await _add_key(integration_session, "older", created_at=1000)
+    await _add_key(integration_session, "newer", created_at=2000)
+    await _add_key(integration_session, "legacy", created_at=None)
+
+    response = await integration_client.get(
+        "/admin/api/temporary-balances", headers=_admin_headers()
+    )
+
+    assert response.status_code == 200
+    order = [b["hashed_key"] for b in response.json()["balances"]]
+    # Newest created first, NULL created_at (legacy) sorts last.
+    assert order == ["newer", "older", "legacy"]
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_temporary_balances_pagination(
+    integration_client: httpx.AsyncClient,
+    integration_session: AsyncSession,
+) -> None:
+    for i in range(5):
+        await _add_key(integration_session, f"key_{i}", created_at=1000 + i)
+
+    headers = _admin_headers()
+    page1 = (
+        await integration_client.get(
+            "/admin/api/temporary-balances?limit=2&offset=0", headers=headers
+        )
+    ).json()
+    page2 = (
+        await integration_client.get(
+            "/admin/api/temporary-balances?limit=2&offset=2", headers=headers
+        )
+    ).json()
+
+    assert page1["total"] == 5
+    assert page2["total"] == 5
+    assert [b["hashed_key"] for b in page1["balances"]] == ["key_4", "key_3"]
+    assert [b["hashed_key"] for b in page2["balances"]] == ["key_2", "key_1"]
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_temporary_balances_totals_exclude_child_balance(
+    integration_client: httpx.AsyncClient,
+    integration_session: AsyncSession,
+) -> None:
+    await _add_key(
+        integration_session,
+        "parent",
+        balance=5000,
+        total_spent=100,
+        total_requests=3,
+        created_at=1000,
+    )
+    # Child draws from parent's balance, so its balance must NOT be summed,
+    # but its spent/requests still count.
+    await _add_key(
+        integration_session,
+        "child",
+        balance=0,
+        total_spent=200,
+        total_requests=7,
+        created_at=1001,
+        parent_key_hash="parent",
+    )
+
+    response = await integration_client.get(
+        "/admin/api/temporary-balances", headers=_admin_headers()
+    )
+
+    totals = response.json()["totals"]
+    assert totals["total_balance"] == 5000
+    assert totals["total_spent"] == 300
+    assert totals["total_requests"] == 10
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_temporary_balances_search_filters_total_and_totals(
+    integration_client: httpx.AsyncClient,
+    integration_session: AsyncSession,
+) -> None:
+    await _add_key(
+        integration_session,
+        "alpha",
+        balance=1000,
+        created_at=1000,
+        refund_address="alice@ln.tld",
+    )
+    await _add_key(
+        integration_session,
+        "beta",
+        balance=2000,
+        created_at=1001,
+        refund_address="bob@ln.tld",
+    )
+
+    headers = _admin_headers()
+
+    # Match by hashed_key.
+    by_key = (
+        await integration_client.get(
+            "/admin/api/temporary-balances?search=alpha", headers=headers
+        )
+    ).json()
+    assert by_key["total"] == 1
+    assert by_key["balances"][0]["hashed_key"] == "alpha"
+    # totals reflect only the filtered set.
+    assert by_key["totals"]["total_balance"] == 1000
+
+    # Match by refund_address.
+    by_addr = (
+        await integration_client.get(
+            "/admin/api/temporary-balances?search=bob@ln.tld", headers=headers
+        )
+    ).json()
+    assert by_addr["total"] == 1
+    assert by_addr["balances"][0]["hashed_key"] == "beta"
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_temporary_balances_requires_admin(
+    integration_client: httpx.AsyncClient,
+) -> None:
+    response = await integration_client.get("/admin/api/temporary-balances")
+    assert response.status_code in (401, 403)

tests/integration/test_wallet_refund.py

@@ -372,17 +372,17 @@ async def test_refund_rejects_concurrent_topup_on_same_key(
     topup_amount_sat = 500
     topup_token = await testmint_wallet.mint_tokens(topup_amount_sat)
 
-    validate_called = asyncio.Event()
+    key_looked_up = asyncio.Event()
     allow_refund_to_continue = asyncio.Event()
-    original_validate_bearer_key = balance_module.validate_bearer_key
+    original_lookup = balance_module._lookup_key_no_create
     delayed_once = False
 
-    async def delayed_validate_bearer_key(*args: Any, **kwargs: Any) -> ApiKey:
+    async def delayed_lookup_key_no_create(*args: Any, **kwargs: Any) -> ApiKey | None:
         nonlocal delayed_once
-        key = await original_validate_bearer_key(*args, **kwargs)
-        if not delayed_once:
+        key = await original_lookup(*args, **kwargs)
+        if not delayed_once and key is not None:
             delayed_once = True
-            validate_called.set()
+            key_looked_up.set()
             await allow_refund_to_continue.wait()
         return key
 
@@ -390,7 +390,7 @@ async def test_refund_rejects_concurrent_topup_on_same_key(
         return await authenticated_client.post("/v1/wallet/refund")
 
     async def issue_topup() -> Any:
-        await validate_called.wait()
+        await key_looked_up.wait()
         try:
             return await authenticated_client.post(
                 "/v1/wallet/topup", params={"cashu_token": topup_token}
@@ -399,7 +399,7 @@ async def test_refund_rejects_concurrent_topup_on_same_key(
             allow_refund_to_continue.set()
 
     with patch(
-        "routstr.balance.validate_bearer_key", new=delayed_validate_bearer_key
+        "routstr.balance._lookup_key_no_create", new=delayed_lookup_key_no_create
     ):
         refund_response, topup_response = await asyncio.gather(
             issue_refund(), issue_topup()

tests/integration/test_wallet_topup.py

@@ -436,9 +436,9 @@ async def test_topup_with_zero_amount_token(  # type: ignore[no-untyped-def]
             "/v1/wallet/topup", params={"cashu_token": token}
         )
 
-        # Should succeed but add 0 msats
-        assert response.status_code == 200
-        assert response.json()["msats"] == 0
+        # Zero/negative redemptions are refused to avoid crediting empty
+        # or dust tokens (and to prevent orphan zero-balance keys).
+        assert response.status_code == 400
 
 
 @pytest.mark.integration

tests/unit/test_auth_cashu.py

@@ -0,0 +1,59 @@
+import hashlib
+from types import SimpleNamespace
+from typing import AsyncGenerator
+from unittest.mock import AsyncMock, patch
+
+import pytest
+from fastapi import HTTPException
+from sqlalchemy.ext.asyncio import AsyncEngine, create_async_engine
+from sqlalchemy.pool import StaticPool
+from sqlmodel import SQLModel
+from sqlmodel.ext.asyncio.session import AsyncSession
+
+from routstr.auth import validate_bearer_key
+from routstr.core.db import ApiKey
+
+
+def _make_engine() -> AsyncEngine:
+    return create_async_engine(
+        "sqlite+aiosqlite://",
+        poolclass=StaticPool,
+        connect_args={"check_same_thread": False},
+    )
+
+
+@pytest.fixture
+async def session() -> AsyncGenerator[AsyncSession, None]:
+    engine = _make_engine()
+    async with engine.begin() as conn:
+        await conn.run_sync(SQLModel.metadata.create_all)
+    db_session = AsyncSession(engine, expire_on_commit=False)
+    try:
+        yield db_session
+    finally:
+        await db_session.close()
+        await engine.dispose()
+
+
+@pytest.mark.asyncio
+async def test_failed_first_cashu_redemption_rolls_back_empty_api_key(
+    session: AsyncSession,
+) -> None:
+    token = "cashuAfirst_seen_but_redemption_fails"
+    hashed_key = hashlib.sha256(token.encode()).hexdigest()
+    token_obj = SimpleNamespace(mint="http://mint:3338", unit="sat")
+
+    from routstr.core.settings import settings
+
+    with (
+        patch.object(settings, "cashu_mints", ["http://mint:3338"]),
+        patch("routstr.auth.deserialize_token_from_string", return_value=token_obj),
+        patch(
+            "routstr.auth.credit_balance",
+            new=AsyncMock(side_effect=ValueError("token already spent")),
+        ),
+    ):
+        with pytest.raises(HTTPException):
+            await validate_bearer_key(token, session)
+
+    assert await session.get(ApiKey, hashed_key) is None

tests/unit/test_balance.py

@@ -168,12 +168,12 @@ async def test_apikey_refund_stores_cashu_transaction_with_apikey_source() -> No
     refund_token = "cashuArefund_apikey_token"
 
     session = MagicMock()
+    session.get = AsyncMock(return_value=key)
     session.exec = AsyncMock(return_value=_update_result(1))
     session.add = MagicMock()
     session.commit = AsyncMock()
 
     with (
-        patch("routstr.balance.validate_bearer_key", AsyncMock(return_value=key)),
         patch("routstr.balance.get_billing_key", AsyncMock(return_value=key)),
         patch("routstr.balance.send_token", AsyncMock(return_value=refund_token)),
         patch("routstr.balance.store_cashu_transaction", AsyncMock()) as mock_store,
@@ -203,12 +203,12 @@ async def test_apikey_refund_logs_token() -> None:
     refund_token = "cashuAlogged_token"
 
     session = MagicMock()
+    session.get = AsyncMock(return_value=key)
     session.exec = AsyncMock(return_value=_update_result(1))
     session.add = MagicMock()
     session.commit = AsyncMock()
 
     with (
-        patch("routstr.balance.validate_bearer_key", AsyncMock(return_value=key)),
         patch("routstr.balance.get_billing_key", AsyncMock(return_value=key)),
         patch("routstr.balance.send_token", AsyncMock(return_value=refund_token)),
         patch("routstr.balance.store_cashu_transaction", AsyncMock()),
@@ -232,12 +232,12 @@ async def test_apikey_refund_log_includes_path() -> None:
     refund_token = "cashuApath_token"
 
     session = MagicMock()
+    session.get = AsyncMock(return_value=key)
     session.exec = AsyncMock(return_value=_update_result(1))
     session.add = MagicMock()
     session.commit = AsyncMock()
 
     with (
-        patch("routstr.balance.validate_bearer_key", AsyncMock(return_value=key)),
         patch("routstr.balance.get_billing_key", AsyncMock(return_value=key)),
         patch("routstr.balance.send_token", AsyncMock(return_value=refund_token)),
         patch("routstr.balance.store_cashu_transaction", AsyncMock()),
@@ -269,6 +269,7 @@ async def test_apikey_refund_rejects_on_concurrent_balance_change() -> None:
     key = _make_api_key(balance=5000, refund_currency="sat")
 
     session = MagicMock()
+    session.get = AsyncMock(return_value=key)
     # Debit returns rowcount=0 → balance changed concurrently
     session.exec = AsyncMock(return_value=_update_result(0))
     session.commit = AsyncMock()
@@ -276,7 +277,6 @@ async def test_apikey_refund_rejects_on_concurrent_balance_change() -> None:
     mock_send_token = AsyncMock(return_value="cashuAshould_not_be_minted")
 
     with (
-        patch("routstr.balance.validate_bearer_key", AsyncMock(return_value=key)),
         patch("routstr.balance.get_billing_key", AsyncMock(return_value=key)),
         patch("routstr.balance.send_token", mock_send_token),
         patch("routstr.balance.store_cashu_transaction", AsyncMock()),
@@ -333,11 +333,11 @@ async def test_apikey_refund_restores_balance_on_mint_failure() -> None:
 
     # First exec call = debit (succeeds), second = restore
     session = MagicMock()
+    session.get = AsyncMock(return_value=key)
     session.exec = AsyncMock(side_effect=[_update_result(1), _update_result(1)])
     session.commit = AsyncMock()
 
     with (
-        patch("routstr.balance.validate_bearer_key", AsyncMock(return_value=key)),
         patch("routstr.balance.get_billing_key", AsyncMock(return_value=key)),
         patch("routstr.balance.send_token", AsyncMock(side_effect=Exception("mint down"))),
         patch("routstr.balance.store_cashu_transaction", AsyncMock()),
@@ -355,3 +355,49 @@ async def test_apikey_refund_restores_balance_on_mint_failure() -> None:
     assert exc_info.value.status_code == 503
     # Verify two exec calls: debit + restore
     assert session.exec.await_count == 2
+
+
+# ---------------------------------------------------------------------------
+# no-create guarantee: fresh Cashu/unknown sk- tokens must not create API keys
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_refund_fresh_cashu_bearer_returns_401() -> None:
+    """Fresh Cashu token not in DB must get 401, never create a new ApiKey."""
+    from fastapi import HTTPException
+
+    session = MagicMock()
+    session.get = AsyncMock(return_value=None)
+
+    with pytest.raises(HTTPException) as exc_info:
+        await refund_wallet_endpoint(
+            authorization="Bearer cashuAfresh_never_deposited_token",
+            x_cashu=None,
+            session=session,
+        )
+
+    assert exc_info.value.status_code == 401
+    session.get.assert_awaited_once()
+    # No add/commit → no key was persisted
+    session.add.assert_not_called()
+    session.commit.assert_not_called()
+
+
+@pytest.mark.asyncio
+async def test_refund_unknown_sk_bearer_returns_401() -> None:
+    """Unknown sk- key not in DB must get 401."""
+    from fastapi import HTTPException
+
+    session = MagicMock()
+    session.get = AsyncMock(return_value=None)
+
+    with pytest.raises(HTTPException) as exc_info:
+        await refund_wallet_endpoint(
+            authorization="Bearer sk-unknownhash",
+            x_cashu=None,
+            session=session,
+        )
+
+    assert exc_info.value.status_code == 401
+    session.get.assert_awaited_once()

tests/unit/test_cache_breakpoints.py

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

tests/unit/test_cache_pricing.py

@@ -0,0 +1,245 @@
+"""Tests for cache-aware pricing of cached input tokens.
+
+Specifies two things:
+
+1. ``backfill_cache_pricing`` — when the OpenRouter model feed omits cache
+   rates (it does for most DeepSeek models and e.g. openai/gpt-4o), they are
+   filled from litellm's bundled cost map instead of silently billing cache
+   reads at the full input rate. Existing OpenRouter values are never
+   overwritten, and provider fees apply to backfilled rates like any other.
+2. ``calculate_cost`` — cached tokens are billed at the cache rates from the
+   model's sats_pricing; the full input rate remains only as the documented
+   last resort when no cache rate could be resolved anywhere.
+"""
+
+import os
+from unittest.mock import Mock, patch
+
+import litellm
+import pytest
+
+os.environ.setdefault("UPSTREAM_BASE_URL", "http://test")
+os.environ.setdefault("UPSTREAM_API_KEY", "test")
+os.environ.setdefault("LIGHTNING_ADDRESS", "test@stm.to")
+
+from routstr.core.settings import settings
+from routstr.payment.cost_calculation import CostData, calculate_cost
+from routstr.payment.models import (
+    Architecture,
+    Model,
+    Pricing,
+    backfill_cache_pricing,
+)
+from routstr.upstream import GenericUpstreamProvider
+
+
+def _make_model(model_id: str, pricing: Pricing) -> Model:
+    return Model(
+        id=model_id,
+        name=model_id,
+        created=0,
+        description="",
+        context_length=64000,
+        architecture=Architecture(
+            modality="text->text",
+            input_modalities=["text"],
+            output_modalities=["text"],
+            tokenizer="Other",
+            instruct_type=None,
+        ),
+        pricing=pricing,
+    )
+
+
+# ============================================================================
+# backfill_cache_pricing — litellm as fallback source for missing cache rates
+# ============================================================================
+
+
+def test_backfill_deepseek_cache_read_from_litellm() -> None:
+    """deepseek/deepseek-chat has no input_cache_read on OpenRouter; litellm
+    knows the real rate (10x cheaper than input)."""
+    pricing = Pricing(prompt=2.8e-07, completion=4.2e-07)
+
+    result = backfill_cache_pricing("deepseek/deepseek-chat", pricing)
+
+    expected = litellm.model_cost["deepseek/deepseek-chat"][
+        "cache_read_input_token_cost"
+    ]
+    assert result.input_cache_read == expected
+    assert result.input_cache_read < pricing.prompt  # sanity: it's a discount
+
+
+def test_backfill_strips_vendor_prefix_for_litellm_lookup() -> None:
+    """OpenRouter ids are vendor-prefixed (openai/gpt-4o); litellm keys most
+    non-DeepSeek models without the prefix (gpt-4o)."""
+    pricing = Pricing(prompt=2.5e-06, completion=1e-05)
+
+    result = backfill_cache_pricing("openai/gpt-4o", pricing)
+
+    expected = litellm.model_cost["gpt-4o"]["cache_read_input_token_cost"]
+    assert result.input_cache_read == expected
+
+
+def test_backfill_fills_cache_write_rate() -> None:
+    """Anthropic cache writes cost more than input (1.25x); billing them at
+    the input rate undercharges. litellm carries the write rate."""
+    pricing = Pricing(prompt=3e-06, completion=1.5e-05)
+
+    result = backfill_cache_pricing("anthropic/claude-sonnet-4-5", pricing)
+
+    expected = litellm.model_cost["claude-sonnet-4-5"][
+        "cache_creation_input_token_cost"
+    ]
+    assert result.input_cache_write == expected
+    assert result.input_cache_write > pricing.prompt  # sanity: write premium
+
+
+def test_backfill_never_overwrites_openrouter_rates() -> None:
+    """When OpenRouter provides a cache rate, it is authoritative."""
+    pricing = Pricing(
+        prompt=2.1e-07, completion=7.9e-07, input_cache_read=1.3e-07
+    )
+
+    result = backfill_cache_pricing("deepseek/deepseek-chat", pricing)
+
+    assert result.input_cache_read == 1.3e-07
+
+
+def test_backfill_unknown_model_unchanged() -> None:
+    """Models litellm doesn't know stay untouched (last-resort fallback to
+    the input rate happens later, at billing time)."""
+    pricing = Pricing(prompt=1e-06, completion=2e-06)
+
+    result = backfill_cache_pricing("artificial-dumbness/dumb-1", pricing)
+
+    assert result.input_cache_read == 0.0
+    assert result.input_cache_write == 0.0
+
+
+def test_provider_fee_applies_to_backfilled_cache_rates() -> None:
+    """Backfill happens before the provider fee, so cache rates carry the
+    same markup as every other price component."""
+    provider = GenericUpstreamProvider(
+        base_url="http://upstream.example", provider_fee=2.0
+    )
+    model = _make_model(
+        "deepseek/deepseek-chat", Pricing(prompt=2.8e-07, completion=4.2e-07)
+    )
+
+    adjusted = provider._apply_provider_fee_to_model(model)
+
+    litellm_read = litellm.model_cost["deepseek/deepseek-chat"][
+        "cache_read_input_token_cost"
+    ]
+    assert adjusted.pricing.input_cache_read == pytest.approx(litellm_read * 2.0)
+    assert adjusted.pricing.prompt == pytest.approx(2.8e-07 * 2.0)
+
+
+# ============================================================================
+# calculate_cost — cached tokens billed at cache rates
+# ============================================================================
+
+
+@pytest.fixture(autouse=True)
+def patch_sats_usd_price() -> None:  # type: ignore[misc]
+    with patch("routstr.payment.cost_calculation.sats_usd_price", return_value=5.0e-5):
+        yield
+
+
+@pytest.fixture
+def model_pricing(monkeypatch: pytest.MonkeyPatch) -> Mock:
+    """Model-based pricing: 1 msat per input token, 2 per output token,
+    0.1 per cache-read token, 1.25 per cache-write token."""
+    monkeypatch.setattr(settings, "fixed_pricing", False)
+    model = Mock()
+    model.sats_pricing = Pricing(
+        prompt=0.001,
+        completion=0.002,
+        input_cache_read=0.0001,
+        input_cache_write=0.00125,
+    )
+    return model
+
+
+@pytest.mark.asyncio
+async def test_deepseek_cache_hits_billed_at_cache_rate(model_pricing: Mock) -> None:
+    """The reported overcharge scenario: a 10k-token prompt with 90% cache
+    hits costs 2900 msats at honest rates, not the 11000 msats that billing
+    every prompt token at the full input rate would charge."""
+    response = {
+        "model": "deepseek-chat",
+        "usage": {
+            "prompt_tokens": 10000,
+            "completion_tokens": 500,
+            "prompt_cache_hit_tokens": 9000,
+            "prompt_cache_miss_tokens": 1000,
+        },
+    }
+
+    with patch("routstr.proxy.get_model_instance", return_value=model_pricing):
+        result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    # 1000 input @ 1 msat + 9000 cache reads @ 0.1 msat + 500 output @ 2 msat.
+    # input_msats folds the cache-read cost in (1000 + 900) so a dashboard
+    # rendering I/O/T sees input + output == total; the cache portion stays
+    # visible in cache_read_msats.
+    assert result.cache_read_msats == 900
+    assert result.output_msats == 1000
+    assert result.input_msats == 1900
+    assert result.input_msats + result.output_msats == result.total_msats
+    assert result.total_msats == 2900
+
+
+@pytest.mark.asyncio
+async def test_anthropic_cache_write_billed_at_write_rate(model_pricing: Mock) -> None:
+    """Cache writes carry their premium rate (1.25x input here), instead of
+    being silently billed at the plain input rate."""
+    response = {
+        "model": "claude-sonnet-4-5",
+        "usage": {
+            "input_tokens": 300,
+            "output_tokens": 100,
+            "cache_read_input_tokens": 500,
+            "cache_creation_input_tokens": 2000,
+        },
+    }
+
+    with patch("routstr.proxy.get_model_instance", return_value=model_pricing):
+        result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    # 300 @ 1 + 500 @ 0.1 + 2000 @ 1.25 + 100 @ 2
+    assert result.cache_read_msats == 50
+    assert result.cache_creation_msats == 2500
+    assert result.total_msats == 3050
+
+
+@pytest.mark.asyncio
+async def test_missing_cache_rate_falls_back_to_input_rate(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """Documented last resort: when no cache rate could be resolved anywhere
+    (OpenRouter and litellm both silent), cache reads bill at the input rate —
+    never cheaper, never free."""
+    monkeypatch.setattr(settings, "fixed_pricing", False)
+    model = Mock()
+    model.sats_pricing = Pricing(prompt=0.001, completion=0.002)
+
+    response = {
+        "model": "dumb-1",
+        "usage": {
+            "prompt_tokens": 10000,
+            "completion_tokens": 500,
+            "prompt_cache_hit_tokens": 9000,
+            "prompt_cache_miss_tokens": 1000,
+        },
+    }
+
+    with patch("routstr.proxy.get_model_instance", return_value=model):
+        result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    # 1000 @ 1 + 9000 @ 1 (fallback) + 500 @ 2
+    assert result.total_msats == 11000

tests/unit/test_cost_calculation_caching.py

@@ -0,0 +1,433 @@
+"""Tests for cache token handling in cost calculation.
+
+Covers OpenAI, Anthropic and DeepSeek caching formats, dialect precedence,
+edge cases, and billing accuracy.
+"""
+
+import os
+from unittest.mock import patch
+
+import pytest
+
+os.environ.setdefault("UPSTREAM_BASE_URL", "http://test")
+os.environ.setdefault("UPSTREAM_API_KEY", "test")
+os.environ.setdefault("LIGHTNING_ADDRESS", "test@stm.to")
+
+from routstr.core.settings import settings
+from routstr.payment.cost_calculation import CostData, MaxCostData, calculate_cost
+
+
+@pytest.fixture(autouse=True)
+def mock_fixed_pricing(monkeypatch: pytest.MonkeyPatch) -> None:
+    """Mock settings and price to use fixed pricing."""
+    monkeypatch.setattr(settings, "fixed_pricing", True)
+    monkeypatch.setattr(settings, "fixed_per_1k_input_tokens", 0.001)
+    monkeypatch.setattr(settings, "fixed_per_1k_output_tokens", 0.001)
+
+
+@pytest.fixture(autouse=True)
+def patch_sats_usd_price() -> None:  # type: ignore[misc]
+    """Patch sats_usd_price to avoid initialization issues."""
+    with patch("routstr.payment.cost_calculation.sats_usd_price", return_value=5.0e-5):
+        yield
+
+
+# ============================================================================
+# Test 1: OpenAI Cache Format
+# ============================================================================
+@pytest.mark.asyncio
+async def test_openai_cache_subtraction() -> None:
+    """OpenAI includes cached_tokens in prompt_tokens, subtract them."""
+    response = {
+        "model": "gpt-4",
+        "usage": {
+            "prompt_tokens": 2000,  # ← Includes 1000 cached
+            "completion_tokens": 100,
+            "prompt_tokens_details": {
+                "cached_tokens": 1000  # ← Extracted separately
+            }
+        }
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.input_tokens == 1000  # 2000 - 1000
+    assert result.cache_read_input_tokens == 1000
+    assert result.output_tokens == 100
+
+
+# ============================================================================
+# Test 2: Anthropic Cache Format
+# ============================================================================
+@pytest.mark.asyncio
+async def test_anthropic_cache_additive(mock_fixed_pricing: None) -> None:
+    """Anthropic cache tokens are separate (additive) from input_tokens."""
+    response = {
+        "model": "claude-3-5-sonnet",
+        "usage": {
+            "input_tokens": 500,  # ← Regular input only
+            "output_tokens": 100,
+            "cache_creation_input_tokens": 1500,  # ← Additive, not included above
+            "cache_read_input_tokens": 0,
+        }
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.input_tokens == 500
+    assert result.cache_creation_input_tokens == 1500
+    assert result.cache_read_input_tokens == 0
+    assert result.output_tokens == 100
+
+
+# ============================================================================
+# Test 3: Invalid Cache (Edge Case)
+# ============================================================================
+@pytest.mark.asyncio
+async def test_cache_read_exceeds_prompt_tokens(mock_fixed_pricing: None) -> None:
+    """Handle buggy upstream reporting cached > prompt_tokens."""
+    response = {
+        "model": "gpt-4",
+        "usage": {
+            "prompt_tokens": 100,
+            "completion_tokens": 50,
+            "prompt_tokens_details": {
+                "cached_tokens": 150  # ← Invalid! Greater than prompt
+            }
+        }
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    # Should not go negative
+    assert isinstance(result, CostData)
+    assert result.input_tokens == 0  # max(0, 100 - 150)
+    assert result.cache_read_input_tokens == 150
+    assert result.output_tokens == 50
+
+
+# ============================================================================
+# Test 4: Malformed Token Values
+# ============================================================================
+@pytest.mark.asyncio
+async def test_malformed_cache_tokens_coerce_to_zero(mock_fixed_pricing: None) -> None:
+    """Handle non-numeric cache token values."""
+    response = {
+        "model": "gpt-4",
+        "usage": {
+            "prompt_tokens": 100,
+            "completion_tokens": 50,
+            "cache_read_input_tokens": "-50",  # ← String, negative
+            "prompt_tokens_details": {
+                "cached_tokens": "invalid"  # ← Non-numeric string
+            }
+        }
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    # Both should coerce to 0
+    assert isinstance(result, CostData)
+    assert result.cache_read_input_tokens == 0
+    assert result.input_tokens == 100  # No subtraction if cache_read = 0
+
+
+# ============================================================================
+# Test 5: Anthropic Cache Not Subtracted
+# ============================================================================
+@pytest.mark.asyncio
+async def test_anthropic_cache_not_subtracted(mock_fixed_pricing: None) -> None:
+    """Anthropic cache fields should NOT be subtracted from input_tokens."""
+    response = {
+        "model": "claude-3-5-sonnet",
+        "usage": {
+            "input_tokens": 500,
+            "completion_tokens": 100,
+            "cache_read_input_tokens": 200,  # ← Additive, don't subtract
+        }
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    # Anthropic: input_tokens stays as-is
+    assert isinstance(result, CostData)
+    assert result.input_tokens == 500  # NOT 300
+    assert result.cache_read_input_tokens == 200
+
+
+# ============================================================================
+# Test 6: Only Cache Read, No Regular Input
+# ============================================================================
+@pytest.mark.asyncio
+async def test_only_cache_read_tokens(mock_fixed_pricing: None) -> None:
+    """Handle response with only cache read tokens."""
+    response = {
+        "model": "gpt-4",
+        "usage": {
+            "prompt_tokens": 0,
+            "completion_tokens": 50,
+            "prompt_tokens_details": {
+                "cached_tokens": 1000
+            }
+        }
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.input_tokens == 0  # max(0, 0 - 1000)
+    assert result.cache_read_input_tokens == 1000
+    assert result.output_tokens == 50
+
+
+# ============================================================================
+# Test 7: Only Cache Creation
+# ============================================================================
+@pytest.mark.asyncio
+async def test_only_cache_creation_tokens(mock_fixed_pricing: None) -> None:
+    """Handle response with only cache creation tokens (Anthropic)."""
+    response = {
+        "model": "claude-3-5-sonnet",
+        "usage": {
+            "input_tokens": 500,
+            "output_tokens": 100,
+            "cache_creation_input_tokens": 2000,
+            "cache_read_input_tokens": 0,
+        }
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.input_tokens == 500
+    assert result.cache_creation_input_tokens == 2000
+    assert result.cache_read_input_tokens == 0
+    assert result.output_tokens == 100
+
+
+# ============================================================================
+# Test 8: Both Cache Read and Creation
+# ============================================================================
+@pytest.mark.asyncio
+async def test_both_cache_read_and_creation(mock_fixed_pricing: None) -> None:
+    """Handle response with both cache read and creation."""
+    response = {
+        "model": "claude-3-5-sonnet",
+        "usage": {
+            "input_tokens": 300,
+            "output_tokens": 100,
+            "cache_creation_input_tokens": 2000,
+            "cache_read_input_tokens": 500,
+        }
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.input_tokens == 300
+    assert result.cache_creation_input_tokens == 2000
+    assert result.cache_read_input_tokens == 500
+    assert result.output_tokens == 100
+
+
+# ============================================================================
+# Test 9: Token Field Fallback
+# ============================================================================
+@pytest.mark.asyncio
+async def test_token_field_fallback_order(mock_fixed_pricing: None) -> None:
+    """Verify fallback order for token extraction."""
+    # When prompt_tokens is not present, fall back to input_tokens
+    response = {
+        "model": "gpt-4",
+        "usage": {
+            "input_tokens": 250,
+            "completion_tokens": 50,
+        }
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.input_tokens == 250
+    assert result.output_tokens == 50
+
+
+# ============================================================================
+# Test 10: Float Token Values
+# ============================================================================
+@pytest.mark.asyncio
+async def test_float_token_values_coerced_to_int(mock_fixed_pricing: None) -> None:
+    """Handle float token values by converting to int."""
+    response = {
+        "model": "gpt-4",
+        "usage": {
+            "prompt_tokens": 100.7,  # Float
+            "completion_tokens": 50.3,  # Float
+            "prompt_tokens_details": {"cached_tokens": 25.9},  # Float
+        }
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    # cached_tokens are part of prompt_tokens (OpenAI dialect) → subtracted: 100 - 25
+    assert result.input_tokens == 75  # Floored
+    assert result.output_tokens == 50  # Floored
+    assert result.cache_read_input_tokens == 25  # Floored
+
+
+# ============================================================================
+# Test 11: Boolean Cache Tokens
+# ============================================================================
+@pytest.mark.asyncio
+async def test_boolean_cache_tokens_coerced_to_zero(mock_fixed_pricing: None) -> None:
+    """Handle boolean cache token values by coercing to zero."""
+    response = {
+        "model": "gpt-4",
+        "usage": {
+            "prompt_tokens": 100,
+            "completion_tokens": 50,
+            "cache_read_input_tokens": True,  # Boolean
+        }
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.cache_read_input_tokens == 0  # Boolean coerced to 0
+    assert result.input_tokens == 100  # No subtraction
+
+
+# ============================================================================
+# Test 12: Zero Cache Tokens
+# ============================================================================
+@pytest.mark.asyncio
+async def test_zero_cache_tokens(mock_fixed_pricing: None) -> None:
+    """Handle explicit zero cache tokens."""
+    response = {
+        "model": "gpt-4",
+        "usage": {
+            "prompt_tokens": 100,
+            "completion_tokens": 50,
+            "prompt_tokens_details": {
+                "cached_tokens": 0
+            }
+        }
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.cache_read_input_tokens == 0
+    assert result.input_tokens == 100
+
+
+# ============================================================================
+# DeepSeek Cache Format
+# DeepSeek emits neither OpenAI's prompt_tokens_details nor Anthropic's
+# cache_read_input_tokens — only prompt_cache_hit_tokens and
+# prompt_cache_miss_tokens, with the documented guarantee
+# prompt_tokens = hit + miss. Hits are ~10x cheaper upstream, so billing
+# them as regular input is a large overcharge.
+# ============================================================================
+@pytest.mark.asyncio
+async def test_deepseek_cache_hit_tokens_extracted() -> None:
+    """DeepSeek cache hits are extracted and removed from regular input.
+
+    Payload shape verbatim from the DeepSeek API reference (usage object).
+    """
+    response = {
+        "model": "deepseek-chat",
+        "usage": {
+            "prompt_tokens": 10000,  # = hit + miss
+            "completion_tokens": 500,
+            "total_tokens": 10500,
+            "prompt_cache_hit_tokens": 9000,
+            "prompt_cache_miss_tokens": 1000,
+        },
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.input_tokens == 1000  # only the cache misses
+    assert result.cache_read_input_tokens == 9000
+    assert result.output_tokens == 500
+
+
+@pytest.mark.asyncio
+async def test_deepseek_all_tokens_cached() -> None:
+    """A fully cached DeepSeek prompt bills zero regular input tokens."""
+    response = {
+        "model": "deepseek-chat",
+        "usage": {
+            "prompt_tokens": 5000,
+            "completion_tokens": 100,
+            "prompt_cache_hit_tokens": 5000,
+            "prompt_cache_miss_tokens": 0,
+        },
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.input_tokens == 0
+    assert result.cache_read_input_tokens == 5000
+
+
+@pytest.mark.asyncio
+async def test_dialect_precedence_never_double_subtracts() -> None:
+    """If a vendor emits both OpenAI-style and DeepSeek-style cache fields for
+    the same cached tokens, they are counted once, not subtracted twice."""
+    response = {
+        "model": "deepseek-chat",
+        "usage": {
+            "prompt_tokens": 10000,
+            "completion_tokens": 500,
+            "prompt_tokens_details": {"cached_tokens": 9000},
+            "prompt_cache_hit_tokens": 9000,
+            "prompt_cache_miss_tokens": 1000,
+        },
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.input_tokens == 1000  # 10000 - 9000, applied exactly once
+    assert result.cache_read_input_tokens == 9000
+
+
+@pytest.mark.asyncio
+async def test_deepseek_malformed_hit_tokens_coerce_to_zero() -> None:
+    """Malformed DeepSeek cache fields degrade to billing all input at full
+    rate instead of crashing or going negative."""
+    response = {
+        "model": "deepseek-chat",
+        "usage": {
+            "prompt_tokens": 1000,
+            "completion_tokens": 50,
+            "prompt_cache_hit_tokens": "garbage",
+            "prompt_cache_miss_tokens": -5,
+        },
+    }
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.input_tokens == 1000
+    assert result.cache_read_input_tokens == 0
+
+
+# ============================================================================
+# Test 13: Missing Usage Block
+# ============================================================================
+@pytest.mark.asyncio
+async def test_missing_usage_block(mock_fixed_pricing: None) -> None:
+    """When usage is missing, return MaxCostData with zero tokens."""
+    response = {"model": "gpt-4", "choices": [{"message": {"content": "test"}}]}
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, MaxCostData)
+    assert result.input_tokens == 0
+    assert result.cache_read_input_tokens == 0
+    assert result.output_tokens == 0
+
+
+# ============================================================================
+# Test 14: Null Usage Block
+# ============================================================================
+@pytest.mark.asyncio
+async def test_null_usage_block(mock_fixed_pricing: None) -> None:
+    """When usage is null, return MaxCostData with zero tokens."""
+    response = {"model": "gpt-4", "usage": None}
+    result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, MaxCostData)
+    assert result.input_tokens == 0
+    assert result.cache_read_input_tokens == 0

tests/unit/test_count_tokens_local.py

@@ -0,0 +1,177 @@
+"""Unit tests for the local count_tokens shim.
+
+The shim runs whenever an upstream that does not support Anthropic's
+``/v1/messages`` endpoint is asked for a token count. It must always
+return a 200 JSON ``{"input_tokens": N}`` response and must never raise.
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+from unittest.mock import patch
+
+from routstr.payment.models import Architecture, Model, Pricing
+from routstr.upstream import count_tokens as count_tokens_module
+from routstr.upstream.count_tokens import count_tokens_locally
+
+
+def _make_model(model_id: str = "anthropic/claude-3-5-sonnet") -> Model:
+    pricing = Pricing(prompt=0.000003, completion=0.000015)
+    architecture = Architecture(
+        modality="text",
+        input_modalities=["text"],
+        output_modalities=["text"],
+        tokenizer="cl100k_base",
+        instruct_type=None,
+    )
+    return Model(
+        id=model_id,
+        name=model_id,
+        created=0,
+        description="",
+        context_length=200_000,
+        architecture=architecture,
+        pricing=pricing,
+    )
+
+
+def _body(payload: dict[str, Any]) -> bytes:
+    return json.dumps(payload).encode()
+
+
+def _read_payload(response: Any) -> dict[str, Any]:
+    body = response.body if isinstance(response.body, bytes) else bytes(response.body)
+    return json.loads(body.decode())
+
+
+def test_returns_input_tokens_for_simple_messages() -> None:
+    model = _make_model()
+    request_body = _body(
+        {
+            "model": model.id,
+            "messages": [{"role": "user", "content": "hello world"}],
+        }
+    )
+
+    response = count_tokens_locally(request_body, model)
+
+    assert response.status_code == 200
+    assert response.media_type == "application/json"
+    payload = _read_payload(response)
+    assert "input_tokens" in payload
+    assert isinstance(payload["input_tokens"], int)
+    assert payload["input_tokens"] >= 0
+
+
+def test_falls_back_to_estimator_when_litellm_raises() -> None:
+    model = _make_model()
+    request_body = _body(
+        {
+            "model": model.id,
+            "messages": [{"role": "user", "content": "this is a longer message"}],
+        }
+    )
+
+    with patch.object(
+        count_tokens_module,
+        "_count_with_litellm",
+        side_effect=RuntimeError("boom"),
+    ):
+        response = count_tokens_locally(request_body, model)
+
+    assert response.status_code == 200
+    payload = _read_payload(response)
+    assert payload["input_tokens"] >= 1
+
+
+def test_handles_missing_model_object() -> None:
+    request_body = _body(
+        {
+            "model": "anthropic/claude-3-5-sonnet",
+            "messages": [{"role": "user", "content": "hi"}],
+        }
+    )
+
+    response = count_tokens_locally(request_body, None)
+
+    assert response.status_code == 200
+    payload = _read_payload(response)
+    assert payload["input_tokens"] >= 0
+
+
+def test_handles_empty_request_body() -> None:
+    response = count_tokens_locally(b"", _make_model())
+
+    assert response.status_code == 200
+    payload = _read_payload(response)
+    assert payload["input_tokens"] >= 0
+
+
+def test_handles_malformed_json() -> None:
+    response = count_tokens_locally(b"not-json", _make_model())
+
+    assert response.status_code == 200
+    payload = _read_payload(response)
+    assert payload["input_tokens"] >= 0
+
+
+def test_includes_system_prompt_in_count() -> None:
+    model = _make_model()
+    short = _body(
+        {
+            "model": model.id,
+            "messages": [{"role": "user", "content": "hi"}],
+        }
+    )
+    with_system = _body(
+        {
+            "model": model.id,
+            "system": "You are a helpful assistant with a long preamble " * 10,
+            "messages": [{"role": "user", "content": "hi"}],
+        }
+    )
+
+    short_count = _read_payload(count_tokens_locally(short, model))["input_tokens"]
+    long_count = _read_payload(count_tokens_locally(with_system, model))["input_tokens"]
+
+    assert long_count > short_count
+
+
+def test_supports_anthropic_system_block_list() -> None:
+    model = _make_model()
+    request_body = _body(
+        {
+            "model": model.id,
+            "system": [{"type": "text", "text": "be terse" * 50}],
+            "messages": [{"role": "user", "content": "ok"}],
+        }
+    )
+
+    response = count_tokens_locally(request_body, model)
+
+    payload = _read_payload(response)
+    assert payload["input_tokens"] > 0
+
+
+def test_uses_forwarded_model_id_when_present() -> None:
+    model = _make_model("anthropic/claude-3-5-sonnet")
+    model.forwarded_model_id = "claude-3-5-sonnet-20241022"
+    request_body = _body(
+        {
+            "model": "ignored",
+            "messages": [{"role": "user", "content": "hi"}],
+        }
+    )
+
+    captured: dict[str, Any] = {}
+
+    def _capture(model_name: str, body: dict[str, Any]) -> int:
+        captured["model"] = model_name
+        return 7
+
+    with patch.object(count_tokens_module, "_count_with_litellm", side_effect=_capture):
+        response = count_tokens_locally(request_body, model)
+
+    assert captured["model"] == "claude-3-5-sonnet-20241022"
+    assert _read_payload(response)["input_tokens"] == 7

tests/unit/test_ehbp_finalize_payment.py

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

tests/unit/test_fetch_all_balances.py

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

tests/unit/test_litellm_routing.py

@@ -0,0 +1,93 @@
+"""Tests for `routstr.upstream.litellm_routing.detect_litellm_prefix`."""
+
+from __future__ import annotations
+
+import pytest
+
+from routstr.upstream.litellm_routing import detect_litellm_prefix
+
+
+@pytest.mark.parametrize(
+    "base_url,expected",
+    [
+        # The bug case: custom row pointing at Fireworks must NOT route to openai/.
+        ("https://api.fireworks.ai/inference/v1", "fireworks_ai/"),
+        # Other OpenAI-compatible providers commonly plugged into the custom slot.
+        ("https://api.groq.com/openai/v1", "groq/"),
+        ("https://api.x.ai/v1", "xai/"),
+        ("https://api.deepseek.com/v1", "deepseek/"),
+        ("https://api.together.xyz/v1", "together_ai/"),
+        ("https://api.perplexity.ai", "perplexity/"),
+        ("https://openrouter.ai/api/v1", "openrouter/"),
+        ("https://api.mistral.ai/v1", "mistral/"),
+        ("https://codestral.mistral.ai/v1", "codestral/"),
+        ("https://api.cohere.com/v1", "cohere_chat/"),
+        ("https://api.cohere.ai/v1", "cohere_chat/"),
+        ("https://api.deepinfra.com/v1/openai", "deepinfra/"),
+        ("https://api.cerebras.ai/v1", "cerebras/"),
+        ("https://api.sambanova.ai/v1", "sambanova/"),
+        ("https://api.moonshot.cn/v1", "moonshot/"),
+        ("https://api.moonshot.ai/v1", "moonshot/"),
+        ("https://api.studio.nebius.com/v1", "nebius/"),
+        ("https://api.novita.ai/v3/openai", "novita/"),
+        ("https://api.lambda.ai/v1", "lambda_ai/"),
+        ("https://api.aimlapi.com/v1", "aiml/"),
+        ("https://api.featherless.ai/v1", "featherless_ai/"),
+        ("https://integrate.api.nvidia.com/v1", "nvidia_nim/"),
+        ("https://inference.baseten.co/v1", "baseten/"),
+        ("https://ai-gateway.vercel.sh/v1", "vercel_ai_gateway/"),
+        ("https://api.inference.wandb.ai/v1", "wandb/"),
+        ("https://api.poe.com/v1", "poe/"),
+        ("https://llm.chutes.ai/v1/", "chutes/"),
+        ("https://api.v0.dev/v1", "v0/"),
+        ("https://api.hyperbolic.xyz/v1", "hyperbolic/"),
+        ("https://api.synthetic.new/openai/v1", "synthetic/"),
+        ("https://api.stima.tech/v1", "apertis/"),
+        ("https://nano-gpt.com/api/v1", "nano-gpt/"),
+        ("https://api.friendli.ai/serverless/v1", "friendliai/"),
+        ("https://api.galadriel.com/v1", "galadriel/"),
+        ("https://api.llama.com/compat/v1", "meta_llama/"),
+        ("https://api.minimax.io/v1", "minimax/"),
+        ("https://api.minimaxi.com/v1", "minimax/"),
+        ("https://platform.publicai.co/v1", "publicai/"),
+        ("https://inference.api.nscale.com/v1", "nscale/"),
+        ("https://dashscope-intl.aliyuncs.com/compatible-mode/v1", "dashscope/"),
+        ("https://api.endpoints.anyscale.com/v1", "anyscale/"),
+        ("https://api.ai21.com/studio/v1", "ai21_chat/"),
+        ("https://ark.cn-beijing.volces.com/api/v3", "volcengine/"),
+        ("https://api.voyageai.com/v1", "voyage/"),
+        ("https://api.jina.ai/v1", "jina_ai/"),
+        ("https://api.snowflakecomputing.com", "snowflake/"),
+        ("https://my-workspace.databricks.com/serving-endpoints", "databricks/"),
+        ("https://huggingface.co/api/inference", "huggingface/"),
+        # First-class providers — URL detection still produces the right prefix
+        # so subclasses without an explicit override stay correct.
+        ("https://api.openai.com/v1", "openai/"),
+        ("https://api.anthropic.com/v1", "anthropic/"),
+        ("https://generativelanguage.googleapis.com/v1beta/openai", "gemini/"),
+        ("https://us-central1-aiplatform.googleapis.com/v1", "vertex_ai/"),
+        # Azure ordering: must beat api.openai.com.
+        ("https://my-resource.openai.azure.com/openai/deployments/foo", "azure/"),
+        # Ollama hints.
+        ("http://localhost:11434/v1", "ollama_chat/"),
+        ("http://127.0.0.1:11434/v1", "ollama_chat/"),
+        ("http://my-ollama-host:11434/v1", "ollama_chat/"),
+        # Casing and trailing slash normalisation.
+        ("HTTPS://API.FIREWORKS.AI/INFERENCE/V1/", "fireworks_ai/"),
+        # Unknown host falls back to openai/ (still OpenAI-compatible by convention).
+        ("https://example.com/v1", "openai/"),
+        ("", "openai/"),
+    ],
+)
+def test_detect_litellm_prefix(base_url: str, expected: str) -> None:
+    assert detect_litellm_prefix(base_url) == expected
+
+
+def test_detect_litellm_prefix_none_uses_default() -> None:
+    assert detect_litellm_prefix(None) == "openai/"
+
+
+def test_detect_litellm_prefix_custom_default() -> None:
+    assert detect_litellm_prefix("https://example.com", default="anthropic/") == (
+        "anthropic/"
+    )

tests/unit/test_messages_dispatch_cost_accumulation.py

@@ -0,0 +1,333 @@
+"""Tests for cost accumulation in streaming message dispatch.
+
+Verifies that costs are correctly summed across multiple streaming events,
+not taking only the maximum.
+"""
+
+import os
+
+import pytest
+
+os.environ.setdefault("UPSTREAM_BASE_URL", "http://test")
+os.environ.setdefault("UPSTREAM_API_KEY", "test")
+
+from routstr.upstream.messages_dispatch import annotate_event
+
+
+# ============================================================================
+# Test 1: Cost Accumulation via AnnotatedEvent
+# ============================================================================
+@pytest.mark.unit
+def test_annotate_event_extracts_costs() -> None:
+    """Each event should report its own costs."""
+    event = {
+        "type": "content_block_start",
+        "usage": {"total_cost": 0.010, "input_cost": 0.008, "output_cost": 0.002}
+    }
+    result = annotate_event(event, None)
+
+    assert result.total_cost == 0.010
+    assert result.input_cost == 0.008
+    assert result.output_cost == 0.002
+
+
+# ============================================================================
+# Test 2: Multiple Events with Incremental Costs
+# ============================================================================
+@pytest.mark.unit
+def test_multiple_events_sum_costs() -> None:
+    """When processing multiple events, costs should accumulate (not max)."""
+    # Event 1: initial costs
+    event1 = {
+        "type": "message_start",
+        "usage": {"input_tokens": 100, "total_cost": 0.010}
+    }
+    result1 = annotate_event(event1, None)
+
+    # Event 2: additional costs
+    event2 = {
+        "type": "content_block_start",
+        "usage": {"output_tokens": 50, "total_cost": 0.005}
+    }
+    result2 = annotate_event(event2, None)
+
+    # Event 3: more costs
+    event3 = {
+        "type": "message_delta",
+        "usage": {"output_tokens": 25, "total_cost": 0.008}
+    }
+    result3 = annotate_event(event3, None)
+
+    # Each event should report its own cost (before accumulation)
+    assert result1.total_cost == 0.010
+    assert result2.total_cost == 0.005
+    assert result3.total_cost == 0.008
+
+    # When summed for billing: 0.010 + 0.005 + 0.008 = 0.023
+    # This verifies the cost accumulation fix (using += instead of max())
+    total_from_events = result1.total_cost + result2.total_cost + result3.total_cost
+    assert total_from_events == 0.023
+
+
+# ============================================================================
+# Test 3: Token Accumulation Consistency
+# ============================================================================
+@pytest.mark.unit
+def test_tokens_and_costs_extracted_independently() -> None:
+    """Tokens and costs should be extracted independently per event."""
+    event = {
+        "type": "content_block_delta",
+        "usage": {
+            "input_tokens": 100,
+            "output_tokens": 50,
+            "cache_read_input_tokens": 30,
+            "total_cost": 0.007,
+            "input_cost": 0.005,
+            "output_cost": 0.002
+        }
+    }
+    result = annotate_event(event, None)
+
+    # All values should be extracted
+    assert result.input_tokens == 100
+    assert result.output_tokens == 50
+    assert result.cache_read_input_tokens == 30
+    assert result.total_cost == 0.007
+    assert result.input_cost == 0.005
+    assert result.output_cost == 0.002
+
+
+# ============================================================================
+# Test 4: Cost in Message vs Root
+# ============================================================================
+@pytest.mark.unit
+def test_cost_extracted_from_message_usage() -> None:
+    """Costs in message.usage should be extracted correctly."""
+    event = {
+        "type": "message_start",
+        "message": {
+            "usage": {
+                "input_tokens": 100,
+                "total_cost": 0.010,
+                "input_cost": 0.008,
+                "output_cost": 0.002
+            }
+        }
+    }
+    result = annotate_event(event, None)
+
+    assert result.input_tokens == 100
+    assert result.total_cost == 0.010
+    assert result.input_cost == 0.008
+    assert result.output_cost == 0.002
+
+
+# ============================================================================
+# Test 5: Cost at Event Root Level
+# ============================================================================
+@pytest.mark.unit
+def test_cost_extracted_from_event_root() -> None:
+    """Costs at event root should be extracted (OpenRouter style)."""
+    event = {
+        "type": "content_block_delta",
+        "usage": {"output_tokens": 25},
+        "total_cost": 0.005,  # ← At root level
+        "cost": 0.005
+    }
+    result = annotate_event(event, None)
+
+    assert result.output_tokens == 25
+    assert result.total_cost == 0.005
+
+
+# ============================================================================
+# Test 6: Cost Details in Event Root
+# ============================================================================
+@pytest.mark.unit
+def test_cost_details_extracted_from_event_root() -> None:
+    """cost_details at event root should be extracted correctly."""
+    event = {
+        "type": "message_delta",
+        "cost_details": {
+            "total_cost": 0.015,
+            "input_cost": 0.010,
+            "output_cost": 0.005
+        }
+    }
+    result = annotate_event(event, None)
+
+    assert result.total_cost == 0.015
+    assert result.input_cost == 0.010
+    assert result.output_cost == 0.005
+
+
+# ============================================================================
+# Test 7: No Duplicated Dict Lookups
+# ============================================================================
+@pytest.mark.unit
+def test_annotate_event_no_duplicate_lookups() -> None:
+    """Verify that dict lookups are not duplicated (fix for copy-paste error)."""
+    # This is tested implicitly through proper extraction
+    event = {
+        "type": "message_delta",
+        "cost_details": {
+            "total_cost": 0.020,
+            "input_cost": 0.015,
+            "output_cost": 0.005
+        }
+    }
+    result = annotate_event(event, None)
+
+    # Should extract each field exactly once, correctly
+    assert result.total_cost == 0.020
+    assert result.input_cost == 0.015
+    assert result.output_cost == 0.005
+
+
+# ============================================================================
+# Test 8: Model Name Extraction
+# ============================================================================
+@pytest.mark.unit
+def test_model_extracted_from_event() -> None:
+    """Model name should be extracted from event."""
+    event = {
+        "type": "message_start",
+        "message": {"model": "claude-3-5-sonnet"}
+    }
+    result = annotate_event(event, None)
+
+    assert result.model == "claude-3-5-sonnet"
+
+
+# ============================================================================
+# Test 9: Model Name Override
+# ============================================================================
+@pytest.mark.unit
+def test_model_name_override() -> None:
+    """Requested model should override actual model in event."""
+    event = {
+        "type": "message_start",
+        "message": {"model": "actual-model"}
+    }
+    # Override with requested_model
+    result = annotate_event(event, requested_model="alias-model")
+
+    # Event should be modified to use requested_model
+    assert event["message"]["model"] == "alias-model"  # type: ignore[index]
+    assert result.model == "alias-model"
+
+
+# ============================================================================
+# Test 10: SSE Encoding
+# ============================================================================
+@pytest.mark.unit
+def test_sse_bytes_encoded() -> None:
+    """Event should be encoded as SSE bytes."""
+    event = {
+        "type": "content_block_start",
+        "index": 0,
+        "content_block": {"type": "text", "text": ""}
+    }
+    result = annotate_event(event, None)
+
+    assert result.sse_bytes is not None
+    assert isinstance(result.sse_bytes, bytes)
+    assert b"event: content_block_start" in result.sse_bytes or b"data:" in result.sse_bytes
+
+
+# ============================================================================
+# Test 11: Missing Cost Fields Default to Zero
+# ============================================================================
+@pytest.mark.unit
+def test_missing_cost_fields_default_to_zero() -> None:
+    """When cost fields are missing, should default to 0.0."""
+    event = {
+        "type": "content_block_delta",
+        "usage": {"output_tokens": 25}
+        # No cost fields
+    }
+    result = annotate_event(event, None)
+
+    assert result.total_cost == 0.0
+    assert result.input_cost == 0.0
+    assert result.output_cost == 0.0
+
+
+# ============================================================================
+# Test 12: Cache Tokens in Events
+# ============================================================================
+@pytest.mark.unit
+def test_cache_tokens_extracted_from_event() -> None:
+    """Cache tokens should be extracted from event usage."""
+    event = {
+        "type": "message_delta",
+        "usage": {
+            "input_tokens": 100,
+            "output_tokens": 50,
+            "cache_read_input_tokens": 200,
+            "cache_creation_input_tokens": 0
+        }
+    }
+    result = annotate_event(event, None)
+
+    assert result.input_tokens == 100
+    assert result.output_tokens == 50
+    assert result.cache_read_input_tokens == 200
+    assert result.cache_creation_input_tokens == 0
+
+
+# ============================================================================
+# Test 13: Malformed Cost Values
+# ============================================================================
+@pytest.mark.unit
+def test_malformed_cost_values_coerced() -> None:
+    """Malformed cost values should be coerced to 0.0."""
+    event = {
+        "type": "message_delta",
+        "usage": {
+            "input_tokens": 100,
+            "output_tokens": 50,
+            "total_cost": "invalid",  # ← Non-numeric
+        }
+    }
+    result = annotate_event(event, None)
+
+    # Invalid cost should default to 0.0
+    assert result.total_cost == 0.0
+
+
+# ============================================================================
+# Test 14: Negative Cost Values
+# ============================================================================
+@pytest.mark.unit
+def test_negative_cost_values_clamped() -> None:
+    """Negative cost values should be clamped to 0.0."""
+    event = {
+        "type": "message_delta",
+        "usage": {
+            "input_tokens": 100,
+            "output_tokens": 50,
+            "total_cost": -0.05  # ← Negative
+        }
+    }
+    result = annotate_event(event, None)
+
+    # Negative cost should be clamped to 0.0
+    assert result.total_cost == 0.0
+
+
+# ============================================================================
+# Test 15: Streaming Event Type Preserved
+# ============================================================================
+@pytest.mark.unit
+def test_event_type_preserved_in_sse() -> None:
+    """Event type should be preserved in SSE encoding."""
+    event_types = ["message_start", "content_block_start", "content_block_delta", "message_delta"]
+    for event_type in event_types:
+        event = {"type": event_type}
+        result = annotate_event(event, None)
+
+        assert result.event["type"] == event_type  # type: ignore[index]
+        # SSE should include the event type line if present
+        if event_type:
+            assert f"event: {event_type}".encode() in result.sse_bytes

tests/unit/test_models_refresh_loop.py

@@ -0,0 +1,117 @@
+"""Regression tests for the periodic upstream models refresh loop."""
+
+from __future__ import annotations
+
+import asyncio
+import os
+from typing import cast
+from unittest.mock import AsyncMock
+
+import pytest
+
+os.environ.setdefault("UPSTREAM_BASE_URL", "http://test")
+os.environ.setdefault("UPSTREAM_API_KEY", "test")
+
+from routstr.upstream.base import BaseUpstreamProvider  # noqa: E402
+
+
+class _FakeUpstream:
+    """Minimal stand-in for BaseUpstreamProvider used by the refresh loop.
+
+    Only ``base_url`` (for error logging) and ``refresh_models_cache`` (the call
+    under test) are exercised; everything else stays unused.
+    """
+
+    def __init__(self, name: str) -> None:
+        self.base_url = f"http://{name}"
+        self.refresh_models_cache = AsyncMock()
+
+
+def _make_fake_upstream(name: str) -> BaseUpstreamProvider:
+    # The loop only uses duck-typed attributes — cast keeps the test type-clean
+    # without dragging in BaseUpstreamProvider's full constructor.
+    return cast(BaseUpstreamProvider, _FakeUpstream(name))
+
+
+@pytest.mark.asyncio
+async def test_refresh_loop_picks_up_providers_added_after_startup(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """If a provider is added after the loop starts (e.g. via reinitialize_upstreams),
+    the next loop iteration must refresh it. Previously the loop captured the upstream
+    list at startup and missed any later additions."""
+    from routstr.core.settings import settings as global_settings
+    from routstr.upstream.helpers import refresh_upstreams_models_periodically
+
+    # Tight interval so the test finishes quickly.
+    monkeypatch.setattr(
+        global_settings, "models_refresh_interval_seconds", 1, raising=False
+    )
+
+    initial_upstream = _make_fake_upstream("initial")
+    live_list: list[BaseUpstreamProvider] = [initial_upstream]
+
+    # Stub out the post-iteration sats-pricing refresh so the loop body has no DB deps.
+    async def _noop_pricing_refresh() -> None:  # pragma: no cover - trivial stub
+        return None
+
+    monkeypatch.setattr(
+        "routstr.payment.models._update_sats_pricing_once",
+        _noop_pricing_refresh,
+    )
+
+    task = asyncio.create_task(
+        refresh_upstreams_models_periodically(lambda: live_list)
+    )
+
+    try:
+        # Wait for the first iteration to refresh the initial upstream.
+        for _ in range(40):
+            if initial_upstream.refresh_models_cache.await_count >= 1:  # type: ignore[attr-defined]
+                break
+            await asyncio.sleep(0.05)
+        assert initial_upstream.refresh_models_cache.await_count >= 1, (  # type: ignore[attr-defined]
+            "loop did not refresh the initial upstream within the timeout"
+        )
+
+        # Simulate reinitialize_upstreams: replace the live list contents with new
+        # provider instances. The loop must observe the swap on its next tick.
+        new_upstream = _make_fake_upstream("added-after-startup")
+        live_list[:] = [new_upstream]
+
+        for _ in range(60):
+            if new_upstream.refresh_models_cache.await_count >= 1:  # type: ignore[attr-defined]
+                break
+            await asyncio.sleep(0.05)
+
+        assert new_upstream.refresh_models_cache.await_count >= 1, (  # type: ignore[attr-defined]
+            "loop did not refresh the upstream added after startup — "
+            "regression: list snapshot captured at startup"
+        )
+    finally:
+        task.cancel()
+        try:
+            await task
+        except asyncio.CancelledError:
+            pass
+
+
+@pytest.mark.asyncio
+async def test_refresh_loop_disabled_when_interval_non_positive(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    from routstr.core.settings import settings as global_settings
+    from routstr.upstream.helpers import refresh_upstreams_models_periodically
+
+    monkeypatch.setattr(
+        global_settings, "models_refresh_interval_seconds", 0, raising=False
+    )
+
+    upstream = _make_fake_upstream("never-refreshed")
+
+    # Loop must return immediately without ever touching the upstream.
+    await asyncio.wait_for(
+        refresh_upstreams_models_periodically(lambda: [upstream]),
+        timeout=1.0,
+    )
+    upstream.refresh_models_cache.assert_not_awaited()  # type: ignore[attr-defined]

tests/unit/test_provider_field_injection.py

@@ -0,0 +1,129 @@
+from routstr.upstream.anthropic import AnthropicUpstreamProvider
+from routstr.upstream.base import BaseUpstreamProvider
+from routstr.upstream.openrouter import OpenRouterUpstreamProvider
+
+
+def _make_provider(cls: type, provider_type: str) -> BaseUpstreamProvider:
+    p = cls(api_key="test_key")
+    assert p.provider_type == provider_type
+    return p
+
+
+def test_apply_provider_field_direct_upstream() -> None:
+    """For a direct upstream (no upstream-reported provider), the field
+    is just the provider_type string."""
+    p = _make_provider(AnthropicUpstreamProvider, "anthropic")
+    data: dict = {"id": "msg_1", "model": "claude-3-5-sonnet"}
+    p._apply_provider_field(data)
+    assert data["provider"] == "anthropic"
+
+
+def test_apply_provider_field_openrouter_passthrough() -> None:
+    """OpenRouter responses include an upstream ``provider`` string —
+    routstr should prefix with its own provider_type."""
+    p = _make_provider(OpenRouterUpstreamProvider, "openrouter")
+    data: dict = {
+        "id": "gen-abc",
+        "model": "anthropic/claude-3.5-sonnet",
+        "provider": "Anthropic",
+    }
+    p._apply_provider_field(data)
+    assert data["provider"] == "openrouter:Anthropic"
+
+
+def test_apply_provider_field_openrouter_no_upstream_provider() -> None:
+    """If OpenRouter omits the provider field, the real serving provider is
+    unknown — a bare ``openrouter`` value carries no information."""
+    p = _make_provider(OpenRouterUpstreamProvider, "openrouter")
+    data: dict = {"id": "gen-abc"}
+    p._apply_provider_field(data)
+    assert data["provider"] == "unknown"
+
+
+def test_apply_provider_field_openrouter_echoes_router_name() -> None:
+    """If OpenRouter reports its own name as the provider, treat as unknown."""
+    p = _make_provider(OpenRouterUpstreamProvider, "openrouter")
+    data: dict = {"provider": "openrouter"}
+    p._apply_provider_field(data)
+    assert data["provider"] == "unknown"
+
+
+def test_apply_provider_field_openrouter_idempotent_no_double_prefix() -> None:
+    """Re-stamping must never nest the prefix: openrouter only once."""
+    p = _make_provider(OpenRouterUpstreamProvider, "openrouter")
+    data: dict = {"provider": "GMICloud"}
+    p._apply_provider_field(data)
+    assert data["provider"] == "openrouter:GMICloud"
+    # Second pass (e.g. streaming) keeps a single prefix.
+    p._apply_provider_field(data)
+    assert data["provider"] == "openrouter:GMICloud"
+
+
+def test_apply_provider_field_openrouter_collapses_existing_double_prefix() -> None:
+    """A pre-existing double prefix is collapsed to a single one."""
+    p = _make_provider(OpenRouterUpstreamProvider, "openrouter")
+    data: dict = {"provider": "openrouter:openrouter:GMICloud"}
+    p._apply_provider_field(data)
+    assert data["provider"] == "openrouter:GMICloud"
+
+
+def test_apply_provider_field_strips_whitespace() -> None:
+    p = _make_provider(OpenRouterUpstreamProvider, "openrouter")
+    data: dict = {"provider": "  Fireworks  "}
+    p._apply_provider_field(data)
+    assert data["provider"] == "openrouter:Fireworks"
+
+
+def test_apply_provider_field_blank_upstream_treated_as_missing() -> None:
+    p = _make_provider(OpenRouterUpstreamProvider, "openrouter")
+    data: dict = {"provider": "   "}
+    p._apply_provider_field(data)
+    assert data["provider"] == "unknown"
+
+
+def test_apply_provider_field_non_string_upstream_treated_as_missing() -> None:
+    p = _make_provider(OpenRouterUpstreamProvider, "openrouter")
+    data: dict = {"provider": 42}
+    p._apply_provider_field(data)
+    assert data["provider"] == "unknown"
+
+
+def test_apply_provider_field_idempotent_for_direct_upstream() -> None:
+    """Calling twice on a direct upstream payload keeps the same value and
+    never nests the prefix (no ``anthropic:anthropic``)."""
+    p = _make_provider(AnthropicUpstreamProvider, "anthropic")
+    data: dict = {}
+    p._apply_provider_field(data)
+    p._apply_provider_field(data)
+    assert data["provider"] == "anthropic"
+
+
+def test_apply_provider_field_ignores_non_dict() -> None:
+    """Lists / primitives must be skipped silently."""
+    p = _make_provider(AnthropicUpstreamProvider, "anthropic")
+    # Should not raise.
+    p._apply_provider_field([1, 2, 3])  # type: ignore[arg-type]
+    p._apply_provider_field("hello")  # type: ignore[arg-type]
+    p._apply_provider_field(None)  # type: ignore[arg-type]
+
+
+def test_inject_cost_metadata_sets_provider() -> None:
+    """``inject_cost_metadata`` is the unified injection point and must
+    also stamp the provider field."""
+    from unittest.mock import MagicMock
+
+    from routstr.core.db import ApiKey
+
+    p = _make_provider(OpenRouterUpstreamProvider, "openrouter")
+    key = MagicMock(spec=ApiKey)
+    key.balance = 1000
+
+    response_json: dict = {
+        "model": "anthropic/claude-3.5-sonnet",
+        "provider": "Anthropic",
+        "usage": {"prompt_tokens": 10, "completion_tokens": 5},
+    }
+    cost_data = {"total_msats": 2500, "total_usd": 0.0025}
+    p.inject_cost_metadata(response_json, cost_data, key)
+
+    assert response_json["provider"] == "openrouter:Anthropic"

tests/unit/test_proxy_not_found.py

@@ -0,0 +1,70 @@
+"""Tests for the app-level 404 handler in routstr.core.main."""
+
+from __future__ import annotations
+
+import pytest
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+
+from routstr.core import main as core_main
+
+
+def _make_app() -> FastAPI:
+    app = FastAPI()
+    app.add_api_route(
+        "/{path:path}",
+        core_main.not_found_catch_all,
+        methods=["GET", "POST"],
+        include_in_schema=False,
+    )
+    return app
+
+
+@pytest.mark.skipif(
+    core_main._NOT_FOUND_HTML is None,
+    reason="UI bundle (ui_out/404.html) not present in this environment",
+)
+def test_unknown_path_returns_html_404_for_browser() -> None:
+    client = TestClient(_make_app())
+    response = client.get("/some/random/page", headers={"accept": "text/html"})
+    assert response.status_code == 404
+    assert response.headers["content-type"].startswith("text/html")
+    assert "404" in response.text
+
+
+def test_unknown_path_returns_json_404_for_api_client() -> None:
+    client = TestClient(_make_app())
+    response = client.get(
+        "/some/random/page", headers={"accept": "application/json"}
+    )
+    assert response.status_code == 404
+    assert response.headers["content-type"].startswith("application/json")
+    payload = response.json()
+    assert payload["error"]["type"] == "not_found"
+    assert payload["error"]["code"] == 404
+    assert "/some/random/page" in payload["error"]["message"]
+
+
+def test_root_path_returns_404() -> None:
+    client = TestClient(_make_app())
+    response = client.get("/", headers={"accept": "application/json"})
+    assert response.status_code == 404
+
+
+def test_json_returned_when_ui_html_missing(monkeypatch: pytest.MonkeyPatch) -> None:
+    from routstr.core import not_found as nf
+
+    monkeypatch.setattr(nf, "_NOT_FOUND_HTML", None)
+    client = TestClient(_make_app())
+    response = client.get("/some/random/page", headers={"accept": "text/html"})
+    assert response.status_code == 404
+    assert response.headers["content-type"].startswith("application/json")
+
+
+def test_post_unknown_path_returns_json_even_for_browser() -> None:
+    client = TestClient(_make_app())
+    response = client.post("/some/random/page", headers={"accept": "text/html"})
+    assert response.status_code == 404
+    assert response.headers["content-type"].startswith("application/json")
+    payload = response.json()
+    assert payload["error"]["type"] == "not_found"

tests/unit/test_proxy_tinfoil_attestation_routing.py

@@ -0,0 +1,118 @@
+from __future__ import annotations
+
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+from fastapi import FastAPI
+from fastapi.responses import Response
+from httpx import ASGITransport, AsyncClient
+
+from routstr import proxy as proxy_module
+
+
+@pytest.fixture
+def proxy_app() -> FastAPI:
+    app = FastAPI()
+    app.include_router(proxy_module.proxy_router)
+    return app
+
+
+@pytest.mark.asyncio
+async def test_attestation_get_routes_directly_to_tinfoil_provider(
+    monkeypatch: pytest.MonkeyPatch, proxy_app: FastAPI
+) -> None:
+    non_tinfoil = MagicMock()
+    non_tinfoil.provider_type = "openai"
+    non_tinfoil.prepare_headers = MagicMock(return_value={})
+    non_tinfoil.forward_get_request = AsyncMock(
+        return_value=Response(status_code=404, content=b"wrong upstream")
+    )
+
+    tinfoil = MagicMock()
+    tinfoil.provider_type = "tinfoil"
+    tinfoil.prepare_headers = MagicMock(return_value={"accept": "application/json"})
+    tinfoil.forward_get_request = AsyncMock(
+        return_value=Response(status_code=200, content=b'{"attestation":true}')
+    )
+
+    monkeypatch.setattr(proxy_module, "_upstreams", [non_tinfoil, tinfoil])
+
+    async with AsyncClient(
+        transport=ASGITransport(app=proxy_app), base_url="http://test"  # type: ignore[arg-type]
+    ) as client:
+        response = await client.get("/attestation")
+
+    assert response.status_code == 200
+    assert response.content == b'{"attestation":true}'
+    non_tinfoil.forward_get_request.assert_not_called()
+    tinfoil.forward_get_request.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_tee_attestation_get_routes_directly_to_tinfoil_provider(
+    monkeypatch: pytest.MonkeyPatch, proxy_app: FastAPI
+) -> None:
+    non_tinfoil = MagicMock()
+    non_tinfoil.provider_type = "openrouter"
+    non_tinfoil.prepare_headers = MagicMock(return_value={})
+    non_tinfoil.forward_get_request = AsyncMock(
+        return_value=Response(status_code=404, content=b"wrong upstream")
+    )
+
+    tinfoil = MagicMock()
+    tinfoil.provider_type = "tinfoil"
+    tinfoil.prepare_headers = MagicMock(return_value={"accept": "application/json"})
+    tinfoil.forward_get_request = AsyncMock(
+        return_value=Response(status_code=200, content=b'{"tee":true}')
+    )
+
+    monkeypatch.setattr(proxy_module, "_upstreams", [non_tinfoil, tinfoil])
+
+    async with AsyncClient(
+        transport=ASGITransport(app=proxy_app), base_url="http://test"  # type: ignore[arg-type]
+    ) as client:
+        response = await client.get("/tee/attestation")
+
+    assert response.status_code == 200
+    assert response.content == b'{"tee":true}'
+    non_tinfoil.forward_get_request.assert_not_called()
+    tinfoil.forward_get_request.assert_awaited_once()
+
+
+def test_attestation_upstream_selection_is_tinfoil_only() -> None:
+    non_tinfoil = MagicMock(provider_type="openai")
+    tinfoil = MagicMock(provider_type="tinfoil")
+
+    assert proxy_module._select_unauthenticated_get_upstreams(
+        "attestation", [non_tinfoil, tinfoil]
+    ) == [tinfoil]
+    assert proxy_module._select_unauthenticated_get_upstreams(
+        "tee/attestation", [non_tinfoil, tinfoil]
+    ) == [tinfoil]
+    assert proxy_module._select_unauthenticated_get_upstreams(
+        "tee/other", [non_tinfoil, tinfoil]
+    ) == [non_tinfoil, tinfoil]
+
+
+@pytest.mark.asyncio
+async def test_well_known_get_bypasses_model_lookup(
+    monkeypatch: pytest.MonkeyPatch, proxy_app: FastAPI
+) -> None:
+    upstream = MagicMock()
+    upstream.provider_type = "openai"
+    upstream.prepare_headers = MagicMock(return_value={})
+    upstream.forward_get_request = AsyncMock(
+        return_value=Response(status_code=200, content=b"lnurl metadata")
+    )
+
+    monkeypatch.setattr(proxy_module, "_upstreams", [upstream])
+    monkeypatch.setattr(proxy_module, "get_model_instance", MagicMock(side_effect=AssertionError))
+
+    async with AsyncClient(
+        transport=ASGITransport(app=proxy_app), base_url="http://test"  # type: ignore[arg-type]
+    ) as client:
+        response = await client.get("/.well-known/lnurlp/alice")
+
+    assert response.status_code == 200
+    assert response.content == b"lnurl metadata"
+    upstream.forward_get_request.assert_awaited_once()

tests/unit/test_request_correction.py

@@ -0,0 +1,156 @@
+"""Unit tests for the reactive request-correction layer.
+
+Covers the recovery path that lets a request survive a 400 where the upstream
+names a single unsupported request param (e.g. newer Anthropic models
+deprecating ``temperature``): the param is stripped from the JSON body and the
+same upstream is retried, provider-agnostically, keyed off the error text.
+"""
+
+from __future__ import annotations
+
+import json
+
+from fastapi.responses import Response
+
+from routstr.upstream.request_correction import (
+    Correction,
+    correct_request,
+    extract_error_message,
+    strip_unsupported_param,
+)
+
+
+def _body(**kwargs: object) -> bytes:
+    return json.dumps(kwargs).encode()
+
+
+class TestCorrectRequest:
+    def test_strips_deprecated_temperature(self) -> None:
+        body = _body(model="claude-opus-4-8", temperature=1, messages=[])
+        result = correct_request(
+            body, "`temperature` is deprecated for this model.", set()
+        )
+        assert isinstance(result, Correction)
+        assert result.label == "temperature"
+        decoded = json.loads(result.body)
+        assert "temperature" not in decoded
+        assert decoded["model"] == "claude-opus-4-8"
+
+    def test_strips_not_supported_param(self) -> None:
+        body = _body(model="m", top_p=0.9, messages=[])
+        result = correct_request(body, "Parameter 'top_p' is not supported", set())
+        assert result is not None
+        assert result.label == "top_p"
+        assert "top_p" not in json.loads(result.body)
+
+    def test_returns_none_when_label_already_applied(self) -> None:
+        body = _body(model="m", temperature=1)
+        assert (
+            correct_request(body, "`temperature` is deprecated", {"temperature"})
+            is None
+        )
+
+    def test_returns_none_when_param_absent_from_body(self) -> None:
+        body = _body(model="m", messages=[])
+        assert correct_request(body, "`temperature` is deprecated", set()) is None
+
+    def test_returns_none_when_message_does_not_match(self) -> None:
+        body = _body(model="m", temperature=1)
+        assert correct_request(body, "Insufficient balance", set()) is None
+
+    def test_returns_none_on_empty_inputs(self) -> None:
+        assert correct_request(b"", "`temperature` is deprecated", set()) is None
+        assert correct_request(_body(temperature=1), "", set()) is None
+
+    def test_returns_none_on_non_object_body(self) -> None:
+        assert correct_request(b"[1, 2, 3]", "`temperature` is deprecated", set()) is None
+
+    def test_deprecated_model_name_is_not_stripped_as_param(self) -> None:
+        """A 'model is deprecated' error must not strip an unrelated body field.
+
+        The regex matches the ``<token> is deprecated`` wording, but the
+        ``param not in body`` guard means a deprecated *model* name (not a
+        request param) yields no correction rather than a false strip.
+        """
+        body = _body(model="gpt-3", temperature=1, messages=[])
+        assert correct_request(body, "`gpt-3` is deprecated, use gpt-4", set()) is None
+
+    def test_streaming_400_buffered_error_is_correctable(self) -> None:
+        """Streaming 400s funnel through a buffered JSON Response, so the same
+        correction path applies as for non-streaming requests."""
+        # Mirrors forward_upstream_error_response's buffered JSON envelope.
+        resp = Response(
+            content=json.dumps(
+                {"error": {"message": "`temperature` is deprecated for this model"}}
+            ).encode(),
+            status_code=400,
+        )
+        body = _body(model="claude-opus-4-8", temperature=1, messages=[])
+        result = correct_request(body, extract_error_message(resp), set())
+        assert isinstance(result, Correction)
+        assert result.label == "temperature"
+        assert "temperature" not in json.loads(result.body)
+
+
+class TestStripUnsupportedParam:
+    def test_does_not_mutate_input(self) -> None:
+        body = {"model": "m", "temperature": 1}
+        result = strip_unsupported_param(body, "`temperature` is deprecated")
+        assert result is not None
+        new_body, param = result
+        assert param == "temperature"
+        assert "temperature" not in new_body
+        # original untouched (immutability)
+        assert body == {"model": "m", "temperature": 1}
+
+    def test_declines_when_no_match(self) -> None:
+        assert strip_unsupported_param({"temperature": 1}, "nope") is None
+
+
+class TestExtractErrorMessage:
+    def test_extracts_nested_error_message(self) -> None:
+        resp = Response(
+            content=json.dumps(
+                {"error": {"message": "`temperature` is deprecated", "type": "x"}}
+            ).encode(),
+            status_code=400,
+        )
+        assert extract_error_message(resp) == "`temperature` is deprecated"
+
+    def test_extracts_string_error(self) -> None:
+        resp = Response(
+            content=json.dumps({"error": "bad request"}).encode(), status_code=400
+        )
+        assert extract_error_message(resp) == "bad request"
+
+    def test_extracts_top_level_message(self) -> None:
+        resp = Response(
+            content=json.dumps({"message": "nope"}).encode(), status_code=400
+        )
+        assert extract_error_message(resp) == "nope"
+
+    def test_empty_body_returns_empty_string(self) -> None:
+        assert extract_error_message(Response(status_code=400)) == ""
+
+    def test_non_json_body_returns_preview(self) -> None:
+        resp = Response(content=b"plain text error", status_code=400)
+        assert extract_error_message(resp) == "plain text error"
+
+
+class TestEndToEndChaining:
+    def test_two_distinct_params_corrected_sequentially(self) -> None:
+        """Simulates the proxy loop: each 400 fixes one param, set guards reuse."""
+        body = _body(model="m", temperature=1, top_p=0.5, messages=[])
+        applied: set[str] = set()
+
+        first = correct_request(body, "`temperature` is deprecated", applied)
+        assert first is not None
+        body, applied = first.body, applied | {first.label}
+
+        second = correct_request(body, "`top_p` is not supported", applied)
+        assert second is not None
+        body, applied = second.body, applied | {second.label}
+
+        decoded = json.loads(body)
+        assert "temperature" not in decoded and "top_p" not in decoded
+        assert applied == {"temperature", "top_p"}

tests/unit/test_require_parameters.py

@@ -0,0 +1,95 @@
+import json
+import os
+
+os.environ.setdefault("UPSTREAM_BASE_URL", "http://test")
+os.environ.setdefault("UPSTREAM_API_KEY", "test")
+os.environ.setdefault("LIGHTNING_ADDRESS", "test@stm.to")
+
+from routstr.upstream import GenericUpstreamProvider
+from routstr.upstream.openrouter import OpenRouterUpstreamProvider
+
+
+def _model(model_id: str = "openai/gpt-4o"):  # type: ignore[no-untyped-def]
+    from routstr.payment.models import Architecture, Model, Pricing
+
+    return Model(
+        id=model_id,
+        name=model_id,
+        created=0,
+        description="",
+        context_length=128000,
+        architecture=Architecture(
+            modality="text->text",
+            input_modalities=["text"],
+            output_modalities=["text"],
+            tokenizer="GPT",
+            instruct_type=None,
+        ),
+        pricing=Pricing(prompt=0.0, completion=0.0),
+    )
+
+
+def _tool_body() -> dict:
+    return {
+        "model": "openai/gpt-4o",
+        "messages": [{"role": "user", "content": "What's the weather?"}],
+        "tools": [
+            {
+                "type": "function",
+                "function": {"name": "get_weather", "parameters": {}},
+            }
+        ],
+    }
+
+
+def _prepare(provider, body: dict) -> dict:  # type: ignore[no-untyped-def]
+    out = provider.prepare_request_body(json.dumps(body).encode(), _model())
+    assert out is not None
+    return json.loads(out)
+
+
+def test_injects_require_parameters_for_tool_request() -> None:
+    data = _prepare(OpenRouterUpstreamProvider(api_key="test"), _tool_body())
+    assert data["provider"]["require_parameters"] is True
+
+
+def test_generic_provider_on_openrouter_url_is_left_alone() -> None:
+    # Only OpenRouterUpstreamProvider injects; a generic provider pointed at the
+    # same base URL doesn't.
+    provider = GenericUpstreamProvider(base_url="https://openrouter.ai/api/v1")
+    data = _prepare(provider, _tool_body())
+    assert "provider" not in data
+
+
+def test_no_injection_without_tools() -> None:
+    body = {"model": "openai/gpt-4o", "messages": [{"role": "user", "content": "hi"}]}
+    data = _prepare(OpenRouterUpstreamProvider(api_key="test"), body)
+    assert "provider" not in data
+
+
+def test_empty_tools_list_does_not_inject() -> None:
+    body = _tool_body()
+    body["tools"] = []
+    data = _prepare(OpenRouterUpstreamProvider(api_key="test"), body)
+    assert "provider" not in data
+
+
+def test_direct_provider_does_not_inject() -> None:
+    provider = GenericUpstreamProvider(base_url="https://api.openai.com/v1")
+    data = _prepare(provider, _tool_body())
+    assert "provider" not in data
+
+
+def test_keeps_client_set_require_parameters() -> None:
+    body = _tool_body()
+    body["provider"] = {"require_parameters": False}
+    data = _prepare(OpenRouterUpstreamProvider(api_key="test"), body)
+    assert data["provider"]["require_parameters"] is False
+
+
+def test_preserves_other_provider_fields() -> None:
+    body = _tool_body()
+    body["provider"] = {"order": ["openai", "azure"]}
+    data = _prepare(OpenRouterUpstreamProvider(api_key="test"), body)
+    assert data["provider"]["order"] == ["openai", "azure"]
+    assert data["provider"]["require_parameters"] is True

tests/unit/test_settings.py

@@ -1,11 +1,12 @@
 import os
 
 import pytest
+from pydantic.v1 import ValidationError
 from sqlalchemy.ext.asyncio import create_async_engine
 from sqlmodel import text
 from sqlmodel.ext.asyncio.session import AsyncSession
 
-from routstr.core.settings import SettingsService
+from routstr.core.settings import Settings, SettingsService
 
 
 @pytest.mark.asyncio
@@ -46,6 +47,54 @@ async def test_settings_db_precedence_over_env() -> None:
         assert again.enable_analytics_sharing is False
 
 
+def test_payout_settings_have_sensible_defaults() -> None:
+    s = Settings()
+    assert s.min_payout_sat == 210
+    assert s.payout_interval_seconds == 900
+
+
+@pytest.mark.parametrize(
+    "field,bad_value",
+    [
+        ("min_payout_sat", 0),
+        ("min_payout_sat", -1),
+        ("payout_interval_seconds", 0),
+        ("payout_interval_seconds", -10),
+    ],
+)
+def test_payout_settings_reject_invalid_values(field: str, bad_value: int) -> None:
+    kwargs: dict[str, object] = {field: bad_value}
+    with pytest.raises(ValidationError):
+        Settings(**kwargs)  # type: ignore[arg-type]
+
+
+def test_payout_settings_accept_custom_positive_values() -> None:
+    s = Settings(min_payout_sat=500, payout_interval_seconds=60)
+    assert s.min_payout_sat == 500
+    assert s.payout_interval_seconds == 60
+
+
+@pytest.mark.asyncio
+async def test_payout_settings_persist_via_settings_service() -> None:
+    engine = create_async_engine("sqlite+aiosqlite:///:memory:")
+    async with AsyncSession(engine, expire_on_commit=False) as session:
+        await SettingsService.initialize(session)
+        updated = await SettingsService.update(
+            {"min_payout_sat": 1000, "payout_interval_seconds": 300}, session
+        )
+        assert updated.min_payout_sat == 1000
+        assert updated.payout_interval_seconds == 300
+
+
+@pytest.mark.asyncio
+async def test_payout_settings_update_rejects_invalid() -> None:
+    engine = create_async_engine("sqlite+aiosqlite:///:memory:")
+    async with AsyncSession(engine, expire_on_commit=False) as session:
+        await SettingsService.initialize(session)
+        with pytest.raises(ValidationError):
+            await SettingsService.update({"min_payout_sat": 0}, session)
+
+
 @pytest.mark.asyncio
 async def test_settings_initialize_discards_unknown_keys() -> None:
     engine = create_async_engine("sqlite+aiosqlite:///:memory:")

tests/unit/test_stale_reservations.py

@@ -0,0 +1,378 @@
+"""Tests for stale reserved_balance handling (issue #551).
+
+Covers:
+- pay_for_request stamping reserved_at on billing and child keys
+- release_stale_reservations sweeper semantics
+- reset_all_reserved_balances clearing reserved_at
+- refund endpoint self-healing stale/legacy reservations
+- proxy reverting the reservation when the client disconnects (CancelledError)
+"""
+
+import asyncio
+import time
+from typing import AsyncGenerator
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+from sqlalchemy.ext.asyncio import AsyncEngine, create_async_engine
+from sqlalchemy.pool import StaticPool
+from sqlmodel import SQLModel
+from sqlmodel.ext.asyncio.session import AsyncSession
+
+from routstr.auth import pay_for_request
+from routstr.balance import refund_wallet_endpoint
+from routstr.core.db import (
+    ApiKey,
+    release_stale_reservations,
+    reset_all_reserved_balances,
+)
+
+
+def _make_engine() -> AsyncEngine:
+    return create_async_engine(
+        "sqlite+aiosqlite://",
+        poolclass=StaticPool,
+        connect_args={"check_same_thread": False},
+    )
+
+
+@pytest.fixture
+async def session() -> "AsyncGenerator[AsyncSession, None]":
+    engine = _make_engine()
+    async with engine.begin() as conn:
+        await conn.run_sync(SQLModel.metadata.create_all)
+    db_session = AsyncSession(engine, expire_on_commit=False)
+    try:
+        yield db_session
+    finally:
+        await db_session.close()
+        await engine.dispose()
+
+
+# ---------------------------------------------------------------------------
+# pay_for_request stamps reserved_at
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_pay_for_request_sets_reserved_at(session: AsyncSession) -> None:
+    key = ApiKey(hashed_key="paykey", balance=10_000)
+    session.add(key)
+    await session.commit()
+
+    before = int(time.time())
+    await pay_for_request(key, 1_000, session)
+
+    await session.refresh(key)
+    assert key.reserved_balance == 1_000
+    assert key.reserved_at is not None
+    assert key.reserved_at >= before
+
+
+@pytest.mark.asyncio
+async def test_pay_for_request_sets_reserved_at_on_child_key(session: AsyncSession) -> None:
+    parent = ApiKey(hashed_key="parentkey", balance=10_000)
+    child = ApiKey(hashed_key="childkey", balance=0, parent_key_hash="parentkey")
+    session.add(parent)
+    session.add(child)
+    await session.commit()
+
+    await pay_for_request(child, 1_000, session)
+
+    await session.refresh(parent)
+    await session.refresh(child)
+    assert parent.reserved_balance == 1_000
+    assert parent.reserved_at is not None
+    assert child.reserved_balance == 1_000
+    assert child.reserved_at is not None
+
+
+@pytest.mark.asyncio
+async def test_revert_clears_reserved_at_when_fully_released(
+    session: AsyncSession,
+) -> None:
+    from routstr.auth import revert_pay_for_request
+
+    key = ApiKey(hashed_key="revertkey", balance=10_000)
+    session.add(key)
+    await session.commit()
+
+    await pay_for_request(key, 1_000, session)
+    reverted = await revert_pay_for_request(key, session, 1_000)
+
+    assert reverted is True
+    await session.refresh(key)
+    assert key.reserved_balance == 0
+    assert key.reserved_at is None
+
+
+@pytest.mark.asyncio
+async def test_revert_keeps_reserved_at_while_other_reservations_remain(
+    session: AsyncSession,
+) -> None:
+    from routstr.auth import revert_pay_for_request
+
+    key = ApiKey(hashed_key="partialrevert", balance=10_000)
+    session.add(key)
+    await session.commit()
+
+    await pay_for_request(key, 1_000, session)
+    await pay_for_request(key, 1_000, session)
+    reverted = await revert_pay_for_request(key, session, 1_000)
+
+    assert reverted is True
+    await session.refresh(key)
+    assert key.reserved_balance == 1_000
+    assert key.reserved_at is not None
+
+
+# ---------------------------------------------------------------------------
+# release_stale_reservations sweeper
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_release_stale_reservations_releases_old(session: AsyncSession) -> None:
+    key = ApiKey(
+        hashed_key="stalekey",
+        balance=5_000,
+        reserved_balance=1_000,
+        reserved_at=int(time.time()) - 1_000,
+    )
+    session.add(key)
+    await session.commit()
+
+    released = await release_stale_reservations(session, max_age_seconds=300)
+
+    assert released == 1
+    await session.refresh(key)
+    assert key.reserved_balance == 0
+    assert key.reserved_at is None
+
+
+@pytest.mark.asyncio
+async def test_release_stale_reservations_keeps_fresh(session: AsyncSession) -> None:
+    key = ApiKey(
+        hashed_key="freshkey",
+        balance=5_000,
+        reserved_balance=1_000,
+        reserved_at=int(time.time()),
+    )
+    session.add(key)
+    await session.commit()
+
+    released = await release_stale_reservations(session, max_age_seconds=300)
+
+    assert released == 0
+    await session.refresh(key)
+    assert key.reserved_balance == 1_000
+    assert key.reserved_at is not None
+
+
+@pytest.mark.asyncio
+async def test_release_stale_reservations_skips_null_reserved_at(session: AsyncSession) -> None:
+    # Reservations without a timestamp may belong to instances running older
+    # code (rolling deploy) — the background sweeper must not touch them.
+    key = ApiKey(
+        hashed_key="legacykey",
+        balance=5_000,
+        reserved_balance=1_000,
+        reserved_at=None,
+    )
+    session.add(key)
+    await session.commit()
+
+    released = await release_stale_reservations(session, max_age_seconds=300)
+
+    assert released == 0
+    await session.refresh(key)
+    assert key.reserved_balance == 1_000
+
+
+@pytest.mark.asyncio
+async def test_reset_all_reserved_balances_clears_reserved_at(session: AsyncSession) -> None:
+    key = ApiKey(
+        hashed_key="resetkey",
+        balance=5_000,
+        reserved_balance=1_000,
+        reserved_at=int(time.time()),
+    )
+    session.add(key)
+    await session.commit()
+
+    await reset_all_reserved_balances(session)
+
+    await session.refresh(key)
+    assert key.reserved_balance == 0
+    assert key.reserved_at is None
+
+
+# ---------------------------------------------------------------------------
+# Refund endpoint self-healing
+# ---------------------------------------------------------------------------
+
+
+def _refund_patches(refund_token: str = "cashuArefund"):  # type: ignore[no-untyped-def]
+    return (
+        patch("routstr.balance.send_token", AsyncMock(return_value=refund_token)),
+        patch("routstr.balance.store_cashu_transaction", AsyncMock()),
+        patch("routstr.balance._refund_cache_get", AsyncMock(return_value=None)),
+        patch("routstr.balance._refund_cache_set", AsyncMock()),
+    )
+
+
+async def _add_key(session: AsyncSession, **kwargs) -> ApiKey:  # type: ignore[no-untyped-def]
+    key = ApiKey(refund_currency="sat", **kwargs)
+    session.add(key)
+    await session.commit()
+    return key
+
+
+@pytest.mark.asyncio
+async def test_refund_self_heals_stale_reservation(session: AsyncSession) -> None:
+    key = await _add_key(
+        session,
+        hashed_key="stalerefund",
+        balance=5_000,
+        reserved_balance=2_000,
+        reserved_at=int(time.time()) - 10_000,
+    )
+
+    p1, p2, p3, p4 = _refund_patches()
+    with p1, p2, p3, p4:
+        result = await refund_wallet_endpoint(
+            authorization="Bearer sk-stalerefund",
+            x_cashu=None,
+            session=session,
+        )
+
+    assert isinstance(result, dict)
+    assert result["token"] == "cashuArefund"
+    # Full balance refunded (5000 msats -> 5 sats), reservation healed
+    assert result["sats"] == "5"
+    await session.refresh(key)
+    assert key.balance == 0
+    assert key.reserved_balance == 0
+    assert key.reserved_at is None
+
+
+@pytest.mark.asyncio
+async def test_refund_self_heals_legacy_null_reserved_at(session: AsyncSession) -> None:
+    # Keys stuck from before reserved_at existed must be refundable.
+    key = await _add_key(
+        session,
+        hashed_key="legacyrefund",
+        balance=5_000,
+        reserved_balance=2_000,
+        reserved_at=None,
+    )
+
+    p1, p2, p3, p4 = _refund_patches()
+    with p1, p2, p3, p4:
+        result = await refund_wallet_endpoint(
+            authorization="Bearer sk-legacyrefund",
+            x_cashu=None,
+            session=session,
+        )
+
+    assert isinstance(result, dict)
+    assert result["token"] == "cashuArefund"
+    await session.refresh(key)
+    assert key.balance == 0
+    assert key.reserved_balance == 0
+
+
+@pytest.mark.asyncio
+async def test_refund_rejects_recent_reservation(session: AsyncSession) -> None:
+    from fastapi import HTTPException
+
+    await _add_key(
+        session,
+        hashed_key="activerefund",
+        balance=5_000,
+        reserved_balance=2_000,
+        reserved_at=int(time.time()),
+    )
+
+    p1, p2, p3, p4 = _refund_patches()
+    with p1, p2, p3, p4:
+        with pytest.raises(HTTPException) as exc_info:
+            await refund_wallet_endpoint(
+                authorization="Bearer sk-activerefund",
+                x_cashu=None,
+                session=session,
+            )
+
+    assert exc_info.value.status_code == 400
+    assert "ongoing requests" in exc_info.value.detail
+
+
+@pytest.mark.asyncio
+async def test_refund_without_reservation_still_works(session: AsyncSession) -> None:
+    key = await _add_key(
+        session,
+        hashed_key="plainrefund",
+        balance=5_000,
+        reserved_balance=0,
+    )
+
+    p1, p2, p3, p4 = _refund_patches()
+    with p1, p2, p3, p4:
+        result = await refund_wallet_endpoint(
+            authorization="Bearer sk-plainrefund",
+            x_cashu=None,
+            session=session,
+        )
+
+    assert isinstance(result, dict)
+    assert result["token"] == "cashuArefund"
+    await session.refresh(key)
+    assert key.balance == 0
+
+
+# ---------------------------------------------------------------------------
+# Proxy reverts reservation on client disconnect
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_proxy_reverts_reservation_on_client_disconnect() -> None:
+    from routstr import proxy as proxy_module
+
+    key = ApiKey(hashed_key="cancelkey", balance=10_000)
+
+    request = MagicMock()
+    request.method = "POST"
+    request.headers = {"authorization": "Bearer sk-cancelkey"}
+    request.body = AsyncMock(return_value=b'{"model": "test-model"}')
+
+    upstream = MagicMock()
+    upstream.provider_type = "test"
+    upstream.prepare_headers = MagicMock(side_effect=lambda h: h)
+    upstream.forward_request = AsyncMock(side_effect=asyncio.CancelledError())
+
+    session = MagicMock()
+    revert_mock = AsyncMock(return_value=True)
+
+    with (
+        patch.object(proxy_module, "get_model_instance", return_value=MagicMock()),
+        patch.object(proxy_module, "get_provider_for_model", return_value=[upstream]),
+        patch.object(
+            proxy_module, "get_max_cost_for_model", AsyncMock(return_value=1_000)
+        ),
+        patch.object(
+            proxy_module,
+            "calculate_discounted_max_cost",
+            AsyncMock(return_value=1_000),
+        ),
+        patch.object(proxy_module, "check_token_balance", MagicMock()),
+        patch.object(
+            proxy_module, "get_bearer_token_key", AsyncMock(return_value=key)
+        ),
+        patch.object(proxy_module, "pay_for_request", AsyncMock(return_value=1_000)),
+        patch.object(proxy_module, "revert_pay_for_request", revert_mock),
+    ):
+        with pytest.raises(asyncio.CancelledError):
+            await proxy_module.proxy(request, "v1/chat/completions", session=session)
+
+    revert_mock.assert_awaited_once_with(key, session, 1_000)

tests/unit/test_stream_id_injection.py

@@ -51,7 +51,15 @@ async def test_stream_with_id_injection() -> None:
         base.adjust_payment_for_tokens = AsyncMock(
             return_value={"total_usd": 0.1, "total_msats": 100}
         )
-        base.create_session = MagicMock()
+        # create_session() is used as an async context manager whose entered
+        # value exposes an awaitable .get(). Build a mock that behaves that
+        # way so the post-stream cost-chunk emission can run.
+        mock_session = MagicMock()
+        mock_session.get = AsyncMock(return_value=key)
+        mock_ctx = MagicMock()
+        mock_ctx.__aenter__ = AsyncMock(return_value=mock_session)
+        mock_ctx.__aexit__ = AsyncMock(return_value=None)
+        base.create_session = MagicMock(return_value=mock_ctx)
 
         streaming_response = await provider.handle_streaming_chat_completion(
             response=mock_response,

tests/unit/test_streaming_sse_providers.py

@@ -0,0 +1,411 @@
+"""Battle-test the streaming SSE parser against real per-provider framing.
+
+Each test drives the *actual* ``handle_streaming_chat_completion`` generator
+with a mock upstream response whose ``aiter_bytes`` emits byte sequences that
+mirror what each supported provider sends on the wire (captured from the
+providers' own streaming docs):
+
+* OpenAI / Groq / Fireworks / xAI / Perplexity / Azure - plain
+  ``data: {json}\\n\\n`` + ``data: [DONE]``.
+* OpenRouter - same, but with ``: OPENROUTER PROCESSING`` keepalive comments
+  interleaved (the framing that produced the original
+  ``Unexpected token ':'`` client crash).
+* Gemini (native ``alt=sse``) - ``data:`` payloads framed with CRLF.
+
+The invariant every provider must satisfy: every line the proxy emits that
+starts with ``data: `` either equals ``[DONE]`` or is valid JSON, and no SSE
+comment ever reaches the client. That invariant is exactly what the buggy
+``re.split(b"data: ")`` parser violated for OpenRouter.
+"""
+
+import json
+from collections.abc import AsyncGenerator
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+
+from routstr.core.db import ApiKey
+from routstr.upstream import base
+from routstr.upstream.base import BaseUpstreamProvider
+
+
+def _make_response(chunks: list[bytes]) -> MagicMock:
+    async def aiter_bytes() -> AsyncGenerator[bytes, None]:
+        for chunk in chunks:
+            yield chunk
+
+    mock_response = MagicMock()
+    mock_response.status_code = 200
+    mock_response.headers = {"content-type": "text/event-stream"}
+    mock_response.aiter_bytes = aiter_bytes
+    return mock_response
+
+
+async def _drive(chunks: list[bytes], requested_model: str | None = None) -> list[bytes]:
+    """Run the real streaming generator over ``chunks`` and collect output bytes."""
+    provider = BaseUpstreamProvider(
+        base_url="https://api.example.com", api_key="test_key"
+    )
+
+    key = MagicMock(spec=ApiKey)
+    key.hashed_key = "test_hash"
+    key.balance = 1000
+
+    base.adjust_payment_for_tokens = AsyncMock(
+        return_value={"total_usd": 0.1, "total_msats": 100}
+    )
+    mock_session = MagicMock()
+    mock_session.get = AsyncMock(return_value=key)
+    mock_ctx = MagicMock()
+    mock_ctx.__aenter__ = AsyncMock(return_value=mock_session)
+    mock_ctx.__aexit__ = AsyncMock(return_value=None)
+    base.create_session = MagicMock(return_value=mock_ctx)
+
+    streaming_response = await provider.handle_streaming_chat_completion(
+        response=_make_response(chunks),
+        key=key,
+        max_cost_for_model=100,
+        background_tasks=MagicMock(),
+        requested_model=requested_model,
+    )
+
+    out: list[bytes] = []
+    async for chunk in streaming_response.body_iterator:
+        if isinstance(chunk, str):
+            out.append(chunk.encode())
+        else:
+            out.append(bytes(chunk))
+    return out
+
+
+def _data_payloads(out: list[bytes]) -> list[bytes]:
+    """Return the raw payload of every ``data: `` line across all emitted bytes."""
+    payloads: list[bytes] = []
+    for chunk in out:
+        for line in chunk.split(b"\n"):
+            if line.startswith(b"data: "):
+                payloads.append(line[len(b"data: ") :])
+    return payloads
+
+
+def _assert_clean(out: list[bytes]) -> list[dict]:
+    """Core invariant: every data line is [DONE] or valid JSON; no comments leak."""
+    blob = b"".join(out)
+    # No SSE comment line must ever reach the client.
+    for line in blob.split(b"\n"):
+        assert not line.startswith(b":"), f"comment leaked to client: {line!r}"
+        # The original bug signature: a data line whose value is itself a comment.
+        assert not line.startswith(b"data: :"), f"mangled comment frame: {line!r}"
+
+    objs: list[dict] = []
+    for payload in _data_payloads(out):
+        stripped = payload.strip()
+        if stripped == b"[DONE]":
+            continue
+        obj = json.loads(stripped)  # raises if the proxy emitted non-JSON data
+        objs.append(obj)
+    return objs
+
+
+@pytest.mark.asyncio
+async def test_openai_style_plain_stream() -> None:
+    """OpenAI / Groq / Fireworks / xAI / Perplexity: plain data + [DONE]."""
+    chunks = [
+        b'data: {"id":"x","choices":[{"delta":{"content":"Hello"}}]}\n\n',
+        b'data: {"id":"x","choices":[{"delta":{"content":" world"}}]}\n\n',
+        b"data: [DONE]\n\n",
+    ]
+    out = await _drive(chunks)
+    objs = _assert_clean(out)
+    assert any(o.get("choices") for o in objs)
+    assert b"data: [DONE]\n\n" in b"".join(out)
+
+
+@pytest.mark.asyncio
+async def test_openrouter_keepalive_comments() -> None:
+    """OpenRouter ``: OPENROUTER PROCESSING`` keepalives must never crash clients.
+
+    This is the exact regression: the old parser emitted
+    ``data: : OPENROUTER PROCESSING`` which made downstream
+    ``JSON.parse`` throw ``Unexpected token ':'``.
+    """
+    chunks = [
+        b": OPENROUTER PROCESSING\n\n",
+        b": OPENROUTER PROCESSING\n\n",
+        b'data: {"id":"x","choices":[{"delta":{"content":"Hi"}}]}\n\n',
+        b": OPENROUTER PROCESSING\n\n",
+        b'data: {"id":"x","choices":[{"delta":{"content":"!"}}]}\n\n',
+        b"data: [DONE]\n\n",
+    ]
+    out = await _drive(chunks)
+    objs = _assert_clean(out)
+    # The keepalive must be gone entirely.
+    assert b"OPENROUTER PROCESSING" not in b"".join(out)
+    # Real content survived.
+    contents = [
+        c["delta"]["content"]
+        for o in objs
+        for c in o.get("choices", [])
+        if "delta" in c
+    ]
+    assert "Hi" in contents and "!" in contents
+
+
+@pytest.mark.asyncio
+async def test_openrouter_comment_glued_to_data_chunk() -> None:
+    """Keepalive packed into the same TCP chunk as data (the harder case)."""
+    chunks = [
+        b'data: {"id":"x","choices":[{"delta":{"content":"a"}}]}\n\n'
+        b": OPENROUTER PROCESSING\n\n"
+        b'data: {"id":"x","choices":[{"delta":{"content":"b"}}]}\n\n',
+        b"data: [DONE]\n\n",
+    ]
+    out = await _drive(chunks)
+    objs = _assert_clean(out)
+    contents = [
+        c["delta"]["content"]
+        for o in objs
+        for c in o.get("choices", [])
+        if "delta" in c
+    ]
+    assert contents == ["a", "b"]
+
+
+@pytest.mark.asyncio
+async def test_json_split_across_chunk_boundary() -> None:
+    """A single event's JSON arriving in two TCP reads must reassemble."""
+    chunks = [
+        b'data: {"id":"x","choices":[{"delta":{"con',
+        b'tent":"split"}}]}\n\n',
+        b"data: [DONE]\n\n",
+    ]
+    out = await _drive(chunks)
+    objs = _assert_clean(out)
+    contents = [
+        c["delta"]["content"]
+        for o in objs
+        for c in o.get("choices", [])
+        if "delta" in c
+    ]
+    assert contents == ["split"]
+
+
+@pytest.mark.asyncio
+async def test_byte_by_byte_fragmentation() -> None:
+    """Pathological framing: one byte per chunk. Must still parse cleanly."""
+    raw = (
+        b'data: {"id":"x","choices":[{"delta":{"content":"drip"}}]}\n\n'
+        b": OPENROUTER PROCESSING\n\n"
+        b"data: [DONE]\n\n"
+    )
+    chunks = [raw[i : i + 1] for i in range(len(raw))]
+    out = await _drive(chunks)
+    objs = _assert_clean(out)
+    assert objs and objs[0]["choices"][0]["delta"]["content"] == "drip"
+
+
+@pytest.mark.asyncio
+async def test_gemini_crlf_framing() -> None:
+    """Gemini native (alt=sse) frames events with CRLF."""
+    chunks = [
+        b'data: {"id":"g","choices":[{"delta":{"content":"hej"}}]}\r\n\r\n',
+        b'data: {"id":"g","choices":[{"delta":{"content":"!"}}]}\r\n\r\n',
+    ]
+    out = await _drive(chunks)
+    objs = _assert_clean(out)
+    contents = [
+        c["delta"]["content"]
+        for o in objs
+        for c in o.get("choices", [])
+        if "delta" in c
+    ]
+    assert contents == ["hej", "!"]
+
+
+@pytest.mark.asyncio
+async def test_azure_leading_role_chunk() -> None:
+    """Azure OpenAI opens with a content-filter / role-only chunk."""
+    chunks = [
+        b'data: {"id":"az","choices":[],"prompt_filter_results":[]}\n\n',
+        b'data: {"id":"az","choices":[{"delta":{"role":"assistant"}}]}\n\n',
+        b'data: {"id":"az","choices":[{"delta":{"content":"ok"}}]}\n\n',
+        b"data: [DONE]\n\n",
+    ]
+    out = await _drive(chunks)
+    _assert_clean(out)
+
+
+@pytest.mark.asyncio
+async def test_openrouter_mid_stream_error_event() -> None:
+    """OpenRouter mid-stream errors arrive as a normal data JSON event."""
+    err = {
+        "id": "x",
+        "object": "chat.completion.chunk",
+        "model": "openai/gpt-4o",
+        "error": {"code": "server_error", "message": "Provider disconnected"},
+        "choices": [{"index": 0, "delta": {"content": ""}, "finish_reason": "error"}],
+    }
+    chunks = [
+        b'data: {"id":"x","choices":[{"delta":{"content":"partial"}}]}\n\n',
+        b"data: " + json.dumps(err).encode() + b"\n\n",
+    ]
+    out = await _drive(chunks)
+    objs = _assert_clean(out)
+    assert any("error" in o for o in objs), "error event must be forwarded intact"
+
+
+@pytest.mark.asyncio
+async def test_gemini_combined_content_and_usage_chunk() -> None:
+    """Gemini thinking models pack usage into the final *content* chunk.
+
+    Regression: the parser swallowed any chunk carrying a ``usage`` dict, so
+    when content + usage arrived together the assistant text was dropped and
+    the client saw "no assistant messages" despite a 200 + token accounting.
+    """
+    chunks = [
+        b'data: {"id":"g","choices":[{"delta":{"content":"the answer"},'
+        b'"finish_reason":"stop"}],"usage":{"prompt_tokens":3,'
+        b'"completion_tokens":2,"total_tokens":5}}\n\n',
+        b"data: [DONE]\n\n",
+    ]
+    out = await _drive(chunks)
+    objs = _assert_clean(out)
+    contents = [
+        c["delta"]["content"]
+        for o in objs
+        for c in o.get("choices", [])
+        if "delta" in c
+    ]
+    # Content delivered exactly once (not dropped, not duplicated by the trailer).
+    assert contents == ["the answer"]
+
+
+@pytest.mark.asyncio
+async def test_separate_usage_chunk_not_forwarded_as_content() -> None:
+    """A pure usage chunk (choices: []) is still swallowed, content intact."""
+    chunks = [
+        b'data: {"id":"x","choices":[{"delta":{"content":"hello"}}]}\n\n',
+        b'data: {"id":"x","choices":[],"usage":{"total_tokens":4}}\n\n',
+        b"data: [DONE]\n\n",
+    ]
+    out = await _drive(chunks)
+    objs = _assert_clean(out)
+    contents = [
+        c["delta"]["content"]
+        for o in objs
+        for c in o.get("choices", [])
+        if "delta" in c
+    ]
+    assert contents == ["hello"]
+
+
+@pytest.mark.asyncio
+async def test_requested_model_override_applied() -> None:
+    """Model rewriting still works through the buffered parser."""
+    chunks = [
+        b'data: {"id":"x","model":"upstream-model","choices":[{"delta":{"content":"hi"}}]}\n\n',
+        b"data: [DONE]\n\n",
+    ]
+    out = await _drive(chunks, requested_model="routstr-model")
+    objs = _assert_clean(out)
+    # The upstream content chunk carried model "upstream-model"; the parser must
+    # rewrite it to the requested model. (The trailing routstr-generated usage
+    # chunk is excluded - it is not an upstream-forwarded chunk.)
+    content_chunks = [o for o in objs if o.get("choices")]
+    assert content_chunks, "expected at least one forwarded content chunk"
+    assert all(o.get("model") == "routstr-model" for o in content_chunks)
+
+
+@pytest.mark.asyncio
+async def test_multiline_non_json_data_each_line_prefixed() -> None:
+    """A multi-line non-JSON ``data`` block must keep a ``data:`` prefix per line.
+
+    Two ``data:`` lines in one event reassemble to ``line one\\nline two``, which
+    is not JSON, so it takes the raw-forward path. The parser must re-prefix each
+    line; a bare second line would reach the client without its ``data:`` field
+    and break naive SSE parsers.
+    """
+    chunks = [
+        b"data: line one\ndata: line two\n\n",
+        b"data: [DONE]\n\n",
+    ]
+    out = await _drive(chunks)
+    blob = b"".join(out)
+    for line in blob.split(b"\n"):
+        stripped = line.strip()
+        if not stripped or stripped == b"[DONE]":
+            continue
+        assert line.startswith(b"data: "), f"bare line leaked to client: {line!r}"
+    assert b"data: line one" in blob and b"data: line two" in blob
+
+
+@pytest.mark.asyncio
+async def test_crlf_delimiter_split_across_chunk_boundary() -> None:
+    """CRLF event delimiter straddling two TCP reads must not merge events.
+
+    Regression: a per-chunk ``replace(b"\\r\\n", b"\\n")`` left a stray ``\\r``
+    when a ``\\r\\n`` of the ``\\r\\n\\r\\n`` delimiter landed at the very end of
+    one ``aiter_bytes`` chunk and the matching ``\\n`` opened the next. The
+    ``\\n\\n`` split then missed the boundary, glued two events into one frame
+    with two ``data:`` lines, and the client's ``JSON.parse`` threw on the
+    concatenated payload (the "unexpected token"/"Extra data" crash).
+    """
+    e1 = b'data: {"id":"x","choices":[{"delta":{"content":"a"}}]}'
+    e2 = b'data: {"id":"x","choices":[{"delta":{"content":"b"}}]}'
+    chunks = [
+        e1 + b"\r\n\r",  # delimiter cut mid-CRLF
+        b"\n" + e2 + b"\r\n\r\n",
+        b"data: [DONE]\r\n\r\n",
+    ]
+    out = await _drive(chunks)
+
+    # Client-accurate check: a real SSE client concatenates all ``data:`` lines
+    # *within one event* (events are ``\n\n``-delimited) before parsing. A
+    # merged frame would surface here as two objects glued into one payload,
+    # which ``_assert_clean`` (per-line) would miss.
+    blob = b"".join(out)
+    contents: list[str] = []
+    for event in blob.split(b"\n\n"):
+        datas = [
+            ln[len(b"data: ") :]
+            for ln in event.split(b"\n")
+            if ln.startswith(b"data: ")
+        ]
+        if not datas:
+            continue
+        payload = b"".join(datas)
+        if payload.strip() == b"[DONE]":
+            continue
+        obj = json.loads(payload)  # raises if two events were merged into one
+        for c in obj.get("choices", []):
+            if "delta" in c:
+                contents.append(c["delta"]["content"])
+    assert contents == ["a", "b"]
+
+
+@pytest.mark.asyncio
+async def test_truncated_json_tail_on_connection_close() -> None:
+    """A stream that drops mid-event must not emit the partial JSON downstream.
+
+    Regression: the end-of-stream flush ran ``_process_event`` on the leftover
+    buffer unconditionally. When the upstream connection closed mid-event the
+    leftover was incomplete JSON, which fell through to the raw-forward path and
+    handed the client a ``data: {partial`` frame -> ``Unterminated string`` parse
+    error. The truncated tail must be dropped instead.
+    """
+    chunks = [
+        b'data: {"id":"x","choices":[{"delta":{"content":"ok"}}]}\n\n',
+        b'data: {"id":"x","choices":[{"delta":{"con',  # connection dies here
+    ]
+    out = await _drive(chunks)
+    objs = _assert_clean(out)  # raises if the partial tail leaked as a data frame
+    contents = [
+        c["delta"]["content"]
+        for o in objs
+        for c in o.get("choices", [])
+        if "delta" in c
+    ]
+    # The one complete chunk is delivered; the truncated fragment is dropped
+    # entirely (no second delta), and _assert_clean above guarantees nothing
+    # non-JSON ever reached the client.
+    assert contents == ["ok"]