release: v1.3.0 — fix broken auto-upgrade, streaming corruption, and Claude Code compatibility

.github/workflows/release.yml

@@ -19,8 +19,34 @@ env:
   CARGO_TERM_COLOR: always
 
 jobs:
+  verify:
+    name: Verify tag matches Cargo.toml
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v7
+        with:
+          ref: ${{ github.event.inputs.tag || github.ref }}
+
+      # A release whose binary reports a different version than its tag makes
+      # the built-in auto-upgrade re-download the same release on every start,
+      # because `current_version` never catches up with the tag. Fail loudly
+      # instead of publishing that.
+      - name: Compare tag with package version
+        shell: bash
+        run: |
+          tag="${{ github.event.inputs.tag || github.ref_name }}"
+          crate_version=$(grep -m1 '^version = ' Cargo.toml | cut -d'"' -f2)
+          echo "tag=$tag crate_version=$crate_version"
+          if [ "$tag" != "v$crate_version" ]; then
+            echo "::error::Tag $tag does not match Cargo.toml version $crate_version." \
+                 "Bump Cargo.toml (and Cargo.lock) before tagging."
+            exit 1
+          fi
+
   build:
     name: Build ${{ matrix.target }}
+    needs: verify
     runs-on: ${{ matrix.os }}
     strategy:
       fail-fast: false

CHANGELOG.md

@@ -4,6 +4,155 @@ All notable changes to this project will be documented in this file.
 
 ## [Unreleased]
 
+## [1.3.0] - 2026-07-27
+
+### Added
+- `GET /health` liveness/readiness probe reporting version, uptime, Copilot
+  token status and remaining lifetime, loaded model count, requests served, and
+  whether API-key auth is enabled. Answers without contacting upstream and is
+  never guarded by the API key; `?strict=true` returns `503` when not ready
+- `GET /v1/models/{model}` (and `/models/{model}`) OpenAI-compatible single
+  model retrieval, including `capabilities` and `supported_endpoints`. Model
+  aliases from `model_mappings` are resolved; unknown ids return `404`
+- Graceful shutdown on Ctrl-C and `SIGTERM`, draining in-flight requests and SSE
+  streams instead of dropping them
+- `ghc_proxy_uptime_seconds` gauge on `/metrics`
+- `model=` filter on `GET /api/audit`, matching the requested or translated model
+- `/health` and `/v1/models/{model}` documented in `openapi.json`
+- **SSE keepalive.** Streams emit a `: keepalive` comment after 15s of silence
+  so an idle connection is not dropped at the ~60s timeout GitHub's upstream
+  load balancer and most intermediaries enforce. Extended thinking can easily
+  exceed that with no tokens emitted, which previously surfaced as
+  `user_request_timeout` or a stream that died mid-answer. The comment is only
+  injected at an event boundary, so the verbatim passthrough paths are never
+  spliced mid-event
+- **`anthropic-beta` request headers are forwarded.** The client's flags are
+  preserved and merged with the ones the proxy derives, instead of being
+  overwritten by the 1M-context flag
+- Missing `max_tokens` on `/v1/chat/completions` is filled from the model
+  catalog's `capabilities.limits.max_output_tokens`, which several Copilot
+  models reject the request without
+- Requesting a `/responses`-only model on `/v1/chat/completions` now returns an
+  actionable `unsupported_api_for_model` error naming the right endpoint,
+  instead of an opaque upstream 400
+
+### Fixed
+- **`context_management` was silently stripped, disabling Claude Code's context
+  editing.** The field was not in the Anthropic passthrough allowlist, and the
+  upstream rejects it with a misleading `Extra inputs are not permitted` 400
+  unless the `context-management-2025-06-27` beta is requested. It is now
+  forwarded and the matching beta flag is derived automatically
+- **`tool_result` content arrays were forwarded raw, breaking every MCP tool
+  that returns anything but a bare string.** Anthropic allows `content` to be an
+  array of blocks, which MCP servers routinely use; passing it through made the
+  upstream reject the whole request with
+  `type has to be either 'image_url' or 'text'`. Text blocks are now joined and
+  image blocks rewritten as `image_url` data URLs
+- **Images nested inside a `tool_result` did not set `Copilot-Vision-Request`.**
+  `has_image` only inspected top-level content blocks, so screenshots returned
+  by MCP tools were sent as non-vision requests
+- `image` blocks with a `url` source (rather than `base64`) are now translated
+  instead of being turned into an empty data URL
+- Models that replaced `max_tokens` with `max_completion_tokens` (`gpt-5.3-codex`
+  and friends) previously failed outright; the parameter is renamed and the
+  request retried once, on both the streaming and non-streaming paths
+- **Stopped scraping the Arch User Repository for the latest VS Code version.**
+  The AUR maintainers asked proxies of this kind to stop, having become the
+  single most-requested endpoint on their service. `dynamic_vscode_version` now
+  uses Microsoft's own `update.code.visualstudio.com` release API and ignores
+  non-`major.minor.patch` builds
+- The local `/v1/messages/count_tokens` estimate ignored per-message framing and
+  tool schemas, under-reporting badly enough that Claude Code compacted too late
+  and then hit a hard `prompt token count exceeds the limit` failure. Tool
+  definitions are now counted as serialized JSON, plus per-message and
+  per-request overhead
+- **`/v1/messages/count_tokens` always failed with HTTP 400.** The handler only
+  tried the upstream when the catalog advertised `/v1/messages/count_tokens`,
+  which Copilot never does, and then returned an error. It now forwards to the
+  upstream for any model exposing the native `/v1/messages` surface (returning
+  exact counts) and falls back to a local tiktoken estimate — marked
+  `"estimated": true` — instead of erroring. This unblocks Claude Code, which
+  calls the endpoint before every request
+- **Streaming `/v1/responses` returned upstream errors as a `200` SSE body.**
+  A 401/429/5xx from upstream was wrapped in a success stream, so clients saw a
+  broken response instead of the real failure. Non-2xx upstream responses are
+  now surfaced with their status code, matching the other streaming paths
+- **Cost estimates were wrong for mapped models and for `gpt-4o`.** Costs were
+  computed from the client-supplied model name rather than the model actually
+  served, so an alias such as `opus` priced at the fallback rate (~38x too low).
+  The `gpt-4o` rate was also unreachable because the broader `gpt-4` arm matched
+  first. Rates now key off the translated model, order specific families before
+  general ones, strip `publisher/` prefixes for GitHub Models ids, and cover the
+  gpt-4.1/gpt-5/o-series/Gemini families
+- Dashboard pagination could overflow `usize` (panicking the handler) and an
+  unbounded `per_page` forced a full clone of the request store; `page`/`per_page`
+  are now parsed with saturating arithmetic and clamped to 500
+- `error.log` grew without limit; it is now rotated to `error.log.1` past 8 MB
+- The upstream HTTP client had no connect timeout, so a dead upstream could wedge
+  a request indefinitely (30s connect timeout, 90s pool idle timeout; no overall
+  request timeout so SSE streams are unaffected)
+- The model catalog was fetched twice at startup because the periodic refresh
+  timer fired immediately on its first tick
+- `/metrics` and `/api/audit*` cloned every retained record (including captured
+  request/response bodies in debug mode) on each call; aggregation and filtering
+  now happen in place under the store lock
+- Removed a dead `is_github_models` binding in the translated Anthropic path
+
+#### Streaming integrity
+- **Streaming responses could be silently corrupted mid-text.** Every SSE chunk
+  was decoded with `String::from_utf8_lossy` as it arrived, but upstream chunks
+  do not respect UTF-8 character boundaries. Any multi-byte character (CJK,
+  emoji, `—`, box drawing) split across two chunks became `U+FFFD`, and because
+  the damaged payload then failed to parse as JSON it was **dropped entirely**,
+  deleting text from the middle of a response with no error. The corrupted or
+  missing text was then replayed into the next request as conversation history.
+  SSE parsing now buffers raw bytes and only decodes complete lines
+  (`util::SseLineBuffer`), matching the `TextDecoderStream` + `TextLineStream`
+  pipeline the TypeScript `copilot-api` proxy gets from its platform
+- **Truncated streams were delivered as if they were complete.** A transport
+  error mid-stream only did `break`, so the client received a partial answer
+  with no terminator and the request was still recorded as `200`. All five
+  streaming paths now detect the interruption, emit a protocol-appropriate
+  terminator, and record the request as `502`:
+  - OpenAI: `data: {"error": …, "code": "stream_interrupted"}` followed by `[DONE]`
+  - Anthropic: `event: error` with an `api_error` payload
+  - Responses: `event: error` with `code: "stream_interrupted"`
+  - Gemini: final chunk with `finishReason: "OTHER"`
+- **Anthropic streams that ended without a `finish_reason` never sent
+  `message_stop`.** Claude Code either blocks waiting for it or keeps the
+  partial text as a finished assistant turn. `AnthropicStreamState::finish()`
+  now closes any open content block and emits `message_delta` + `message_stop`
+- A final SSE event arriving without a trailing newline was discarded; the line
+  buffer is now flushed when the stream ends
+- `/v1/chat/completions` streaming dropped any `data:` payload that failed to
+  parse as JSON instead of forwarding it
+- `/v1/responses` streaming now also flags a stream that ends before
+  `response.completed`
+- Direct-Anthropic streaming extracted `stop_reason` and `tools_called` by
+  re-scanning the captured response body, so those audit fields were only ever
+  populated in debug mode; they are now collected while the stream is parsed
+- SSE `data:` parsing accepts `data:{…}` (no space) per the SSE spec, not only
+  `data: {…}`
+
+### Changed
+- `/api/audit/summary` now rounds cost figures to 4 decimals so small per-request
+  costs are no longer reported as `0.0`
+- Default `vscode_version` bumped to `1.130.0`
+- The release workflow now fails if the tag does not match the `Cargo.toml`
+  version. `v1.2.3` shipped with `version = "1.2.2"`, so binaries from that
+  release report `1.2.2` and `auto_upgrade` re-downloaded it on every start
+
+## [1.2.3] - 2026-07-20
+
+### Fixed
+- `--setup --claudecode` no longer overwrites unrelated Claude Code settings
+  ([#24](https://github.com/MartinForReal/ghc-proxy/pull/24))
+
+### Known issue
+- Released with `Cargo.toml` still at `1.2.2`, so binaries from this release
+  identify themselves as `1.2.2`. With `auto_upgrade` enabled they re-download
+  this release on every start. Upgrade to 1.3.0 to clear it.
+
 ## [1.2.2] - 2026-07-14
 
 ### Added

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "ghc-proxy"
-version = "1.2.2"
+version = "1.3.0"
 edition = "2021"
 description = "GitHub Copilot API Proxy - Provides OpenAI and Anthropic compatible endpoints via GitHub Copilot (Rust port of ghc-tunnel)"
 license = "MIT"

README.md

@@ -240,11 +240,13 @@ the dashboard and model listings.
 | `POST /v1/chat/completions` | OpenAI chat completions |
 | `POST /v1/responses` | OpenAI responses API (Codex) |
 | `GET /v1/models` | List available models |
+| `GET /v1/models/{model}` | Retrieve a single model (aliases resolved) |
 | `POST /v1/messages` | Anthropic messages API |
-| `POST /v1/messages/count_tokens` | Anthropic token counting |
+| `POST /v1/messages/count_tokens` | Anthropic token counting (local estimate fallback) |
 | `POST /v1beta/models/{model}:generateContent` | Gemini generate content |
 | `POST /v1beta/models/{model}:streamGenerateContent` | Gemini streaming (SSE) |
 | `POST /v1beta/models/{model}:countTokens` | Gemini token counting |
+| `GET /health` | Liveness/readiness probe (`?strict=true` for 503 when not ready) |
 | `GET /openapi.json` | OpenAPI v3 specification |
 | `GET /` | Web dashboard |
 | `GET /metrics/dashboard` | Metrics dashboard UI |

docs/api.md

@@ -22,14 +22,16 @@ require a key on the LLM endpoints; see [Authentication](#authentication) below.
 | `POST /v1/chat/completions` | OpenAI chat completions (also `/chat/completions`) |
 | `POST /v1/responses` | OpenAI Responses API for Codex (also `/responses`) |
 | `POST /v1/messages` | Anthropic Messages API |
-| `POST /v1/messages/count_tokens` | Anthropic token counting (real BPE) |
+| `POST /v1/messages/count_tokens` | Anthropic token counting (real BPE, local estimate fallback) |
 | `POST /v1beta/models/{model}:generateContent` | Gemini generate content |
 | `POST /v1beta/models/{model}:streamGenerateContent` | Gemini streaming (SSE) |
 | `POST /v1beta/models/{model}:countTokens` | Gemini token counting |
 | `POST /v1/embeddings` | Embeddings (also `/embeddings`) |
 | `GET /v1/models` | List available models (also `/models`, `/api/models`) |
+| `GET /v1/models/{model}` | Retrieve a single model (also `/models/{model}`) |
 | `GET /v1/models/full/` | Raw upstream model catalog with capabilities |
 | `GET /usage` | Copilot plan and quota usage |
+| `GET /health` | Liveness/readiness probe |
 | `GET /` | Web analytics dashboard |
 | `GET /metrics/dashboard` | Metrics dashboard UI |
 | `GET /metrics` | OpenMetrics exposition endpoint |
@@ -45,6 +47,58 @@ Streaming (SSE) is supported on the chat, responses, and messages endpoints by
 setting `"stream": true` in the request body. The Gemini surface streams via the
 dedicated `:streamGenerateContent` action.
 
+## Health check
+
+`GET /health` answers without contacting the upstream, so it is cheap enough for
+a service supervisor or container probe to poll frequently. It is never guarded
+by the optional API key.
+
+```bash
+curl http://127.0.0.1:8314/health
+```
+
+```json
+{
+  "status": "ok",
+  "ready": true,
+  "version": "1.3.0",
+  "uptime_seconds": 128,
+  "copilot_token": { "present": true, "expires_in_seconds": 1487 },
+  "models_loaded": 77,
+  "requests_served": 42,
+  "auth_required": false
+}
+```
+
+`ready` is true once a Copilot token has been obtained **and** the model catalog
+has loaded. A degraded proxy still answers `200` with `ready: false` so probes
+can distinguish "process alive" from "able to serve traffic". Add `?strict=true`
+to get `503 Service Unavailable` instead when the proxy is not ready.
+
+## Retrieve a model
+
+`GET /v1/models/{model}` returns a single catalog entry in the OpenAI shape,
+including the raw `capabilities` and `supported_endpoints` reported upstream.
+Model aliases from `model_mappings` are resolved, so `/v1/models/opus` returns
+the mapped Copilot model. Unknown ids return `404` with an OpenAI-style error
+body.
+
+```bash
+curl http://127.0.0.1:8314/v1/models/claude-opus-4.8
+```
+
+## Token counting
+
+`POST /v1/messages/count_tokens` forwards to the upstream Anthropic
+`count_tokens` endpoint for models that expose the native `/v1/messages`
+surface, returning exact counts. For every other model (and whenever the
+upstream call fails) the proxy falls back to a local tiktoken estimate using the
+tokenizer advertised in the model catalog. Estimated responses are marked:
+
+```json
+{ "input_tokens": 812, "estimated": true }
+```
+
 ## OpenAI SDK
 
 ```python
@@ -131,6 +185,23 @@ curl http://127.0.0.1:8314/v1/models/full/
 This is the authoritative source for which models support a 1M-token context
 window (those advertising `max_context_window_tokens` greater than 200,000).
 
+## Audit API
+
+`GET /api/audit` returns recent request records with their extracted audit
+fields. All filters are optional and combine with AND:
+
+| Parameter | Effect |
+|-----------|--------|
+| `endpoint` | Substring match on the endpoint path |
+| `status` | Exact HTTP status code |
+| `tool_name` | Keeps records whose request offered a matching tool |
+| `agent` | `true`/`false` — agent- vs user-initiated requests |
+| `model` | Substring match on the requested or translated model |
+| `page`, `per_page` | Pagination (`per_page` is clamped to 500) |
+
+`GET /api/audit/summary` aggregates the same records into top tools, stop-reason
+counts, estimated cost, and prompt-cache hit rate.
+
 ## Notable behaviors
 
 - **GitHub Models routing** — when enabled (default), requests whose translated
@@ -152,5 +223,39 @@ window (those advertising `max_context_window_tokens` greater than 200,000).
 - **Adaptive-thinking migration** — when an upstream model rejects
   `thinking.type = "enabled"`, the proxy automatically retries using the
   adaptive format.
+- **Upstream errors are never disguised as streams** — a non-2xx upstream
+  response on a `"stream": true` request is returned as a normal error response
+  with the upstream status code, not as a `200` SSE body.
+- **Interrupted streams are reported, not silently truncated** — if the upstream
+  connection drops mid-response, the proxy emits a protocol-appropriate
+  terminator (`data: {"error": …}` + `[DONE]` for OpenAI, `event: error` for
+  Anthropic and Responses, `finishReason: "OTHER"` for Gemini) and records the
+  request as `502`. Anthropic streams that end without a `finish_reason` are
+  also closed with `message_stop`, so clients never block on a half-open
+  message or keep a partial answer as a completed turn.
+- **Byte-exact streaming** — SSE parsing buffers raw bytes and decodes only
+  complete lines, so multi-byte characters split across network chunks are
+  never mangled into `U+FFFD` or dropped.
+- **SSE keepalive** — a `: keepalive` comment is emitted after 15 seconds of
+  silence so extended thinking does not trip the ~60 second idle timeout
+  enforced by the upstream load balancer. Comments are ignored by every
+  spec-compliant SSE client.
+- **`anthropic-beta` passthrough** — the client's beta flags are forwarded and
+  merged with the ones the proxy derives (`context-1m-2025-08-07` for extended
+  context models, `context-management-2025-06-27` when the request uses
+  `context_management`).
+- **MCP tool results** — a `tool_result` whose `content` is an array of blocks
+  is normalized before translation: text blocks are joined and image blocks
+  become `image_url` data URLs. Images nested there also enable the vision
+  header.
+- **Parameter migration** — a model that rejects `max_tokens` in favour of
+  `max_completion_tokens` is retried once with the renamed parameter, and a
+  missing `max_tokens` is filled from the model catalog.
+- **Cost estimates use the served model** — the `estimated_cost_usd` field and
+  the `ghc_proxy_estimated_cost_usd_total` metric price requests using the model
+  actually sent upstream (after translation), not the alias the client asked
+  for.
+- **Graceful shutdown** — on Ctrl-C (or `SIGTERM` on Unix) the proxy stops
+  accepting connections and lets in-flight requests and SSE streams finish.
 - **Content filtering** — system-prompt add/remove and tool-result suffix
   removal are applied per your configuration.