coderouter-cli 2.15.0__tar.gz → 2.16.0__tar.gz

{coderouter_cli-2.15.0 → coderouter_cli-2.16.0}/CHANGELOG.md

@@ -9,6 +9,128 @@ are kept verbatim where the Japanese text itself is the subject).
 
 ---
 
+## [v2.16.0] — 2026-10-24 (Claude Code 2.1.3xx behind a gateway)
+
+Claude Code 2.1.285–2.1.309 changed several defaults for sessions behind a
+custom `ANTHROPIC_BASE_URL`: tool search is on, Opus 4.7+ / Sonnet 5+ /
+Fable ids are treated as 1M-context models, auto mode asks the gateway for a
+server-side classifier, and recovery from rejected features or an
+overflowing prompt now depends on reading the gateway's error message. This
+release makes CodeRouter answer in the shape those recoveries expect.
+
+### Added
+
+- **Anthropic-shaped errors when the whole chain fails (`/v1/messages`).**
+  A failed chain used to come back as FastAPI's `{"detail": "..."}` with
+  HTTP 502 whatever the upstreams said. It now uses Anthropic's envelope
+  (`{"type": "error", "error": {"type", "message"}}`) and a status chosen
+  from what the upstreams said (`coderouter/ingress/anthropic_errors.py`):
+  - an upstream context-overflow error (llama.cpp
+    `exceed_context_size_error`, vLLM / OpenAI "maximum context length",
+    Ollama "prompt too long", …) → **400 `invalid_request_error`, message
+    starting `prompt is too long`**, which is what triggers Claude Code's
+    reactive compaction. An overflow caused by the *completion* budget
+    ("… 32000 in the completion", `max_tokens`) is not reported this way —
+    compacting the prompt cannot fix it — and goes out as a plain 400;
+  - a non-retryable upstream 400 / 422 → 400 with the upstream's own
+    message (Claude Code drops a rejected beta header / feature only when it
+    can read why);
+  - 413 → `request_too_large`; 429 → `rate_limit_error` with the upstream's
+    `retry-after` passed through;
+  - upstream 401 / 403 and everything else → 502 `api_error` (they concern
+    CodeRouter's own upstream credentials, not the client's).
+  Upstream messages are secret-redacted and length-capped. `AdapterError`
+  now keeps the upstream body (capped at 4096 chars) and `Retry-After`.
+- **Streaming requests get a real HTTP status for pre-content failures.**
+  The route now pulls the engine's first event before committing the
+  response, so a chain that fails before producing anything is answered
+  with the status above instead of `200` plus an in-stream `error` event —
+  as api.anthropic.com does. Once the first event exists, nothing changes.
+  The wait shares the stream's overall M14 ceiling (one deadline, not two),
+  and a client that disconnects during it cancels the upstream attempt.
+- **`deferred_tools_action` (profile; `auto` | `reject` | `expand` |
+  `passthrough`, default `auto`).** Claude Code 2.1.304+ sends a server-side
+  `tool_search_tool_*` and tools marked `defer_loading: true`; only
+  api.anthropic.com runs the search. `auto` passes the request through only
+  when every provider in the chain is `kind: anthropic` on api.anthropic.com
+  and otherwise answers 400, which makes Claude Code (2.1.305+) retry
+  without tool search. `expand` drops the search tool and the flags and
+  sends every tool up front.
+- **`context_budget_action: reject`.** At `context_budget_trim_threshold`
+  the guard answers 400 `prompt is too long: ~N tokens > M maximum` instead
+  of trimming, so Claude Code summarizes the conversation itself rather
+  than losing history silently. Works for streaming and non-streaming.
+- **`coderouter vscode-init` writes Claude Code gateway settings** by
+  default: `ENABLE_TOOL_SEARCH=false`, `CLAUDE_CODE_AUTO_MODE_SERVER=0`,
+  and — with the new `--max-context-tokens N` —
+  `CLAUDE_CODE_MAX_CONTEXT_TOKENS=N`. `--no-claude-code-env` opts out. Same
+  merge / conflict rules as the existing variables; an export of one of
+  them outside the `.envrc` fence conflicts only when this run writes it.
+- **`coderouter doctor --check-claude-code`.** Local-only check of the
+  Claude Code install: warns on versions with known gateway bugs (before
+  2.1.301: compaction can repeat; 2.1.304: every request fails when the
+  gateway rejects tool search) and on missing `ENABLE_TOOL_SEARCH`,
+  `CLAUDE_CODE_MAX_CONTEXT_TOKENS` (2.1.285+) and
+  `CLAUDE_CODE_AUTO_MODE_SERVER`. Exit 0 OK / 2 needs attention.
+
+### Changed
+
+- **Breaking for clients that parse error bodies:** `/v1/messages` errors
+  for a failed chain are no longer `{"detail": "..."}`, and a streaming
+  request whose chain fails before the first event is no longer
+  `200` + `event: error`. Both now match the Anthropic API. The
+  `/v1/chat/completions` (OpenAI) route is unchanged.
+
+---
+
+## [v2.15.1] — 2026-10-24 (docs and defaults for the October 2026 ecosystem)
+
+### Added
+
+- **`doctor` names llama.cpp's default-port change.** llama.cpp b11521
+  (2026-10-09, ggml-org/llama.cpp#30159) moved `llama-server`'s default
+  listen port from 8080 to 9931. When the `auth+basic-chat` probe cannot
+  connect to a loopback provider on :8080 or :9931, `doctor` now checks the
+  other port's `/health` (1.5 s, IPv4 literal) and, if it answers, says so
+  in the transport-error detail with the two possible fixes. The config is
+  never rewritten. Other hosts and ports send no extra request.
+- **SBOM on every release.** `release.yml` builds a CycloneDX JSON SBOM
+  (`cyclonedx-bom==7.4.0`) from a clean venv holding only the wheel and its
+  runtime dependencies, and attaches `coderouter-cli-<version>.cdx.json` to
+  the GitHub Release draft. It is written outside `dist/`, so the PyPI
+  upload is unchanged.
+
+### Changed
+
+- **Codex CLI instructions corrected.** The quickstart (Pattern B), README
+  and the "do I need CodeRouter?" guide told users to set
+  `wire_api = "chat"`. Codex removed that option in 2026-02 and now fails
+  while loading such a config; it speaks only the Responses API, which
+  CodeRouter does not serve yet. The docs now say so and give the
+  direct-connect workaround (`codex --oss`, or a `wire_api = "responses"`
+  provider pointing at `llama-server`).
+- **llama.cpp port notes.** Backend guides and examples keep their explicit
+  `--port 8080` (correct on every build) and now explain the 9931 default
+  for servers started without `--port`. The Launcher is unaffected: it
+  always passes `--port`, so its 8080 default stays.
+- **README: package-name warning.** The PyPI project named `coderouter` is
+  a different product; the install section now says to install
+  `coderouter-cli`.
+- **README: repair benchmark and release protections.** The tool-call
+  repair row links the benchmark and its numbers, and the Security section
+  lists the release protections as layers rather than resting on
+  Trusted Publishing alone.
+
+### Security
+
+- **Lockfile bumps for newly published advisories.** `anyio` 4.13.0 →
+  4.14.2 (PYSEC-2026-4024, PYSEC-2026-4025), `certifi` 2026.2.25 →
+  2026.10.23 (PYSEC-2026-4320), `fsspec` 2026.4.0 → 2026.9.0
+  (PYSEC-2026-4217; reached only through the optional `[accuracy]` extra
+  via `huggingface-hub`). Found by the CI dependency audit.
+
+---
+
 ## [v2.15.0] — 2026-08-21 (fallback explainability + stream truncation detection)
 
 ### Added

{coderouter_cli-2.15.0 → coderouter_cli-2.16.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: coderouter-cli
-Version: 2.15.0
+Version: 2.16.0
 Summary: Local-first, free-first, fallback-built-in LLM router. Claude Code / OpenAI compatible.
 Project-URL: Homepage, https://github.com/zephel01/CodeRouter
 Project-URL: Repository, https://github.com/zephel01/CodeRouter
@@ -66,7 +66,7 @@ Description-Content-Type: text/markdown
 ## 何ができるか — 30 秒で
 
 ```
-あなたのエージェント (Claude Code / codex / agy)
+あなたのエージェント (Claude Code / agy ほか ※)
         │
         ▼
   ┌─ CodeRouter ─┐
@@ -76,6 +76,8 @@ Description-Content-Type: text/markdown
   └──────────────┘
 ```
 
+※ 現行の codex CLI は Responses API（`/v1/responses`）専用になったため、CodeRouter 経由ではまだ使えません（対応予定）。直結の代替手順 → [quickstart Pattern B](./docs/start/quickstart.md#pattern-b-codex-cli-で使う)
+
 **やってくれること:**
 
 - ローカルモデルが壊した tool calling を Claude Code に届く前に修復する
@@ -118,6 +120,8 @@ uvx --from coderouter-cli coderouter serve --port 8088
 
 恒久インストールしたい場合: `uv tool install coderouter-cli`
 
+> **パッケージ名に注意**: PyPI のパッケージ名は **`coderouter-cli`** です。PyPI の `coderouter`（名前だけ同じ）は別の製品なので、`pip install coderouter` と打たないでください。
+
 ---
 
 ## Claude Code で使う
@@ -143,7 +147,7 @@ ANTHROPIC_BASE_URL=http://localhost:8088 ANTHROPIC_AUTH_TOKEN=dummy claude
 | Claude Code + ローカル Ollama で tool calling が壊れる | **必須** — tool 修復 (+ 必要なら wire 変換) |
 | Claude Code + ローカルで長時間回すと止まる | **必須級** — 6 系統ガード + self-healing |
 | Ollama v0.14+ / LM Studio にネイティブ直結で動いてる | **便利** — 直結に無い fallback / ガード / 診断を追加 (passthrough で翻訳ゼロ) |
-| codex / agy + Ollama 直繋ぎで動いてる | オプション — フォールバックが欲しいなら |
+| codex / agy + Ollama 直繋ぎで動いてる | オプション — フォールバックが欲しいなら(codex は `/v1/responses` 対応後) |
 | Claude API を直接叩いてて問題ない | 不要 |
 
 詳細は → [要否判定ガイド](./docs/start/when-do-i-need-coderouter.md)
@@ -156,7 +160,7 @@ ANTHROPIC_BASE_URL=http://localhost:8088 ANTHROPIC_AUTH_TOKEN=dummy claude
 
 | 機能 | 何をしてくれるか |
 |---|---|
-| **Tool-call 修復** | ローカルモデルがテキストで吐いた JSON を正しい tool_use ブロックに復元 |
+| **Tool-call 修復** | ローカルモデルがテキストで吐いた JSON を正しい tool_use ブロックに復元。実測は [修復ベンチ](./benchmarks/tool-repair/README.md)（壊れた tool call を出す 4 モデルで成功率 0% → 80〜100%、回帰コーパスで誤修復 0 件。CI でゲート） |
 | **3 層フォールバック** | ローカル → 無料クラウド → 有料の順に自動切替 |
 | **出力フィルタ** | `<think>` タグ漏れ、stop marker 漏れ、byte-fallback (`<0xNN>`) を自動除去/修復 |
 | **Wire 翻訳** | Anthropic 形式 ↔ OpenAI 形式を自動変換 (ネイティブ `/v1/messages` 対応 backend は passthrough で翻訳ゼロ) |
@@ -333,6 +337,14 @@ CodeRouter は backend ルーター層として独立して動きます。`OPENA
 
 シークレットは環境変数に置きます。[`docs/security.md`](./docs/guides/security.md) に完全な方針と報告手順があります。
 
+リリースは多層で守っています。どれか 1 つを根拠に「安全」とは言いません(正規のパイプラインでビルドされた汚染版が出た例があります)。
+
+- PyPI へは Trusted Publishing(OIDC)で公開し、長期の API トークンを置かない
+- PyPI 側の公開許可は `release.yml` の `pypi` 環境に限定している
+- タグは事前検査スクリプトを通してから打つ(version・CHANGELOG・main の一致を機械的に確認)
+- 実行時の依存は 5 個に絞っている
+- 各リリースに SBOM(CycloneDX)を添付する
+
 ## License
 
 MIT

{coderouter_cli-2.15.0 → coderouter_cli-2.16.0}/README.en.md

@@ -21,7 +21,7 @@
 ## What it does — in 30 seconds
 
 ```
-Your agent (Claude Code / codex / agy)
+Your agent (Claude Code / agy, etc. *)
         │
         ▼
   ┌─ CodeRouter ──┐
@@ -31,6 +31,8 @@ Your agent (Claude Code / codex / agy)
   └────────────────┘
 ```
 
+\* Current codex CLI speaks only the Responses API (`/v1/responses`), so it cannot go through CodeRouter yet (planned). Direct-connect workaround → [quickstart Pattern B](./docs/start/quickstart.en.md#pattern-b-use-it-with-the-codex-cli)
+
 **What it does for you:**
 
 - Repairs broken tool calling from local models before it reaches Claude Code
@@ -55,6 +57,8 @@ uvx --from coderouter-cli coderouter serve --port 8088
 
 For a permanent install: `uv tool install coderouter-cli`
 
+> **Mind the package name**: the PyPI package is **`coderouter-cli`**. The PyPI package named `coderouter` is a different product; don't `pip install coderouter`.
+
 ---
 
 ## Use with Claude Code
@@ -79,7 +83,7 @@ That's it. Claude Code works as usual, but your local Ollama is answering behind
 |---|---|
 | Claude Code + local Ollama, tool calling breaks | **Yes** — wire translation + tool repair |
 | Claude Code + local, dies after long sessions | **Helpful** — 6 guards + self-healing |
-| codex / agy + Ollama works fine | Optional — if you want fallback |
+| codex / agy + Ollama works fine | Optional — if you want fallback (codex: once `/v1/responses` lands) |
 | Using Claude API directly, no issues | Not needed |
 
 Full decision matrix → [Do I need CodeRouter?](./docs/start/when-do-i-need-coderouter.en.md)
@@ -93,7 +97,7 @@ Full decision matrix → [Do I need CodeRouter?](./docs/start/when-do-i-need-cod
 | Feature | What it does |
 |---|---|
 | **Wire translation** | Claude Code (Anthropic format) ↔ Ollama (OpenAI format) auto-converted |
-| **Tool-call repair** | JSON that local models emit as plain text → valid tool_use blocks |
+| **Tool-call repair** | JSON that local models emit as plain text → valid tool_use blocks. Measured in the [repair benchmark](./benchmarks/tool-repair/README.md): four models that break tool calls go from 0% to 80–100% success; zero false repairs on the regression corpus, gated in CI |
 | **3-tier fallback** | Local → free cloud → paid, automatic switching |
 | **Output filters** | Strips leaked `<think>` tags, stop markers, XML tool tags |
 
@@ -268,6 +272,14 @@ CodeRouter runs as an independent backend router layer. Point any project's `OPE
 
 Secrets go in env vars, not config files. See [`docs/security.en.md`](./docs/guides/security.en.md) for the full policy and reporting instructions.
 
+Releases are protected in layers. No single one of these is offered as proof of safety (tampered builds have shipped from legitimate pipelines before).
+
+- Published to PyPI via Trusted Publishing (OIDC); no long-lived API token exists
+- PyPI accepts uploads only from `release.yml` running in the `pypi` environment
+- Tags are cut only after a pre-flight script checks version, CHANGELOG and main
+- Five runtime dependencies
+- An SBOM (CycloneDX) is attached to every release
+
 ## License
 
 MIT

{coderouter_cli-2.15.0 → coderouter_cli-2.16.0}/README.md

@@ -21,7 +21,7 @@
 ## 何ができるか — 30 秒で
 
 ```
-あなたのエージェント (Claude Code / codex / agy)
+あなたのエージェント (Claude Code / agy ほか ※)
         │
         ▼
   ┌─ CodeRouter ─┐
@@ -31,6 +31,8 @@
   └──────────────┘
 ```
 
+※ 現行の codex CLI は Responses API（`/v1/responses`）専用になったため、CodeRouter 経由ではまだ使えません（対応予定）。直結の代替手順 → [quickstart Pattern B](./docs/start/quickstart.md#pattern-b-codex-cli-で使う)
+
 **やってくれること:**
 
 - ローカルモデルが壊した tool calling を Claude Code に届く前に修復する
@@ -73,6 +75,8 @@ uvx --from coderouter-cli coderouter serve --port 8088
 
 恒久インストールしたい場合: `uv tool install coderouter-cli`
 
+> **パッケージ名に注意**: PyPI のパッケージ名は **`coderouter-cli`** です。PyPI の `coderouter`（名前だけ同じ）は別の製品なので、`pip install coderouter` と打たないでください。
+
 ---
 
 ## Claude Code で使う
@@ -98,7 +102,7 @@ ANTHROPIC_BASE_URL=http://localhost:8088 ANTHROPIC_AUTH_TOKEN=dummy claude
 | Claude Code + ローカル Ollama で tool calling が壊れる | **必須** — tool 修復 (+ 必要なら wire 変換) |
 | Claude Code + ローカルで長時間回すと止まる | **必須級** — 6 系統ガード + self-healing |
 | Ollama v0.14+ / LM Studio にネイティブ直結で動いてる | **便利** — 直結に無い fallback / ガード / 診断を追加 (passthrough で翻訳ゼロ) |
-| codex / agy + Ollama 直繋ぎで動いてる | オプション — フォールバックが欲しいなら |
+| codex / agy + Ollama 直繋ぎで動いてる | オプション — フォールバックが欲しいなら(codex は `/v1/responses` 対応後) |
 | Claude API を直接叩いてて問題ない | 不要 |
 
 詳細は → [要否判定ガイド](./docs/start/when-do-i-need-coderouter.md)
@@ -111,7 +115,7 @@ ANTHROPIC_BASE_URL=http://localhost:8088 ANTHROPIC_AUTH_TOKEN=dummy claude
 
 | 機能 | 何をしてくれるか |
 |---|---|
-| **Tool-call 修復** | ローカルモデルがテキストで吐いた JSON を正しい tool_use ブロックに復元 |
+| **Tool-call 修復** | ローカルモデルがテキストで吐いた JSON を正しい tool_use ブロックに復元。実測は [修復ベンチ](./benchmarks/tool-repair/README.md)（壊れた tool call を出す 4 モデルで成功率 0% → 80〜100%、回帰コーパスで誤修復 0 件。CI でゲート） |
 | **3 層フォールバック** | ローカル → 無料クラウド → 有料の順に自動切替 |
 | **出力フィルタ** | `<think>` タグ漏れ、stop marker 漏れ、byte-fallback (`<0xNN>`) を自動除去/修復 |
 | **Wire 翻訳** | Anthropic 形式 ↔ OpenAI 形式を自動変換 (ネイティブ `/v1/messages` 対応 backend は passthrough で翻訳ゼロ) |
@@ -288,6 +292,14 @@ CodeRouter は backend ルーター層として独立して動きます。`OPENA
 
 シークレットは環境変数に置きます。[`docs/security.md`](./docs/guides/security.md) に完全な方針と報告手順があります。
 
+リリースは多層で守っています。どれか 1 つを根拠に「安全」とは言いません(正規のパイプラインでビルドされた汚染版が出た例があります)。
+
+- PyPI へは Trusted Publishing(OIDC)で公開し、長期の API トークンを置かない
+- PyPI 側の公開許可は `release.yml` の `pypi` 環境に限定している
+- タグは事前検査スクリプトを通してから打つ(version・CHANGELOG・main の一致を機械的に確認)
+- 実行時の依存は 5 個に絞っている
+- 各リリースに SBOM(CycloneDX)を添付する
+
 ## License
 
 MIT

{coderouter_cli-2.15.0 → coderouter_cli-2.16.0}/coderouter/adapters/anthropic_native.py

@@ -458,6 +458,8 @@ class AnthropicAdapter(BaseAdapter):
                 provider=self.name,
                 status_code=resp.status_code,
                 retryable=resp.status_code in _RETRYABLE_STATUSES,
+                upstream_body=resp.text,
+                retry_after=resp.headers.get("retry-after"),
             )
 
         try:
@@ -570,6 +572,8 @@ class AnthropicAdapter(BaseAdapter):
                         provider=self.name,
                         status_code=resp.status_code,
                         retryable=resp.status_code in _RETRYABLE_STATUSES,
+                        upstream_body=body.decode("utf-8", errors="replace"),
+                        retry_after=resp.headers.get("retry-after"),
                     )
 
                 # Anthropic SSE block shape:

{coderouter_cli-2.15.0 → coderouter_cli-2.16.0}/coderouter/adapters/base.py

@@ -101,6 +101,12 @@ class StreamChunk(BaseModel):
     usage: dict[str, Any] | None = None
 
 
+#: v2.16.0: max chars of an upstream error body an :class:`AdapterError`
+#: keeps. Error bodies are small; the cap only bounds a misbehaving upstream
+#: that answers an error status with a large HTML page.
+UPSTREAM_BODY_CAP = 4096
+
+
 class AdapterError(CodeRouterError):
     """Raised when a provider call fails in a way the fallback engine should retry on."""
 
@@ -111,6 +117,8 @@ class AdapterError(CodeRouterError):
         provider: str,
         status_code: int | None = None,
         retryable: bool = True,
+        upstream_body: str | None = None,
+        retry_after: str | None = None,
     ) -> None:
         """Construct an AdapterError.
 
@@ -125,11 +133,23 @@ class AdapterError(CodeRouterError):
             retryable: When True, the fallback engine may try the next
                 provider in the chain. When False, the engine stops
                 and surfaces the error as a terminal failure.
+            upstream_body: v2.16.0 — the upstream's error body (capped at
+                :data:`UPSTREAM_BODY_CAP` chars), kept so the ingress can
+                hand the upstream's own wording back to the client.
+                Claude Code recovers from several 400s (rejected beta
+                headers, tool search, context overflow) only when it can
+                read the upstream message. Never logged by this class.
+            retry_after: v2.16.0 — the upstream's ``Retry-After`` header
+                value, verbatim, when it sent one.
         """
         super().__init__(message)
         self.provider = provider
         self.status_code = status_code
         self.retryable = retryable
+        self.upstream_body = (
+            upstream_body[:UPSTREAM_BODY_CAP] if upstream_body is not None else None
+        )
+        self.retry_after = retry_after
 
     def __str__(self) -> str:
         """Render as ``[provider status=NNN] message`` for log trails."""

{coderouter_cli-2.15.0 → coderouter_cli-2.16.0}/coderouter/adapters/openai_compat.py

@@ -244,6 +244,8 @@ class OpenAICompatAdapter(BaseAdapter):
                 provider=self.name,
                 status_code=resp.status_code,
                 retryable=resp.status_code in _RETRYABLE_STATUSES,
+                upstream_body=resp.text,
+                retry_after=resp.headers.get("retry-after"),
             )
 
         try:
@@ -381,6 +383,8 @@ class OpenAICompatAdapter(BaseAdapter):
                         provider=self.name,
                         status_code=resp.status_code,
                         retryable=resp.status_code in _RETRYABLE_STATUSES,
+                        upstream_body=body.decode("utf-8", errors="replace"),
+                        retry_after=resp.headers.get("retry-after"),
                     )
                 async for line in resp.aiter_lines():
                     if not line:

{coderouter_cli-2.15.0 → coderouter_cli-2.16.0}/coderouter/cli.py

@@ -165,6 +165,17 @@ def _build_parser() -> argparse.ArgumentParser:
             "for live key values. Exit 0 clean / 2 needs attention / 1 leak."
         ),
     )
+    doctor.add_argument(
+        "--check-claude-code",
+        action="store_true",
+        help=(
+            "v2.16.0: check the local Claude Code install for gateway use — "
+            "known-bad versions (<2.1.301, 2.1.304) and the env vars that "
+            "matter behind a non-Anthropic gateway (ENABLE_TOOL_SEARCH, "
+            "CLAUDE_CODE_MAX_CONTEXT_TOKENS, CLAUDE_CODE_AUTO_MODE_SERVER). "
+            "No network. Exit 0 OK / 2 needs attention."
+        ),
+    )
     doctor.add_argument(
         "--config",
         default=None,
@@ -454,6 +465,28 @@ def _build_parser() -> argparse.ArgumentParser:
             "generation. Tip: keep secrets in a separate .envrc.local."
         ),
     )
+    vscode_init.add_argument(
+        "--max-context-tokens",
+        type=int,
+        default=None,
+        metavar="N",
+        help=(
+            "v2.16.0: also write CLAUDE_CODE_MAX_CONTEXT_TOKENS=N so Claude "
+            "Code compacts before your local model's context fills up "
+            "(since 2.1.285 it assumes 1M behind a gateway). Use the "
+            "smallest context window in the profile's chain."
+        ),
+    )
+    vscode_init.add_argument(
+        "--no-claude-code-env",
+        action="store_true",
+        help=(
+            "v2.16.0: do not write the Claude Code gateway settings "
+            "(ENABLE_TOOL_SEARCH=false, CLAUDE_CODE_AUTO_MODE_SERVER=0, "
+            "and --max-context-tokens). Use when the chain is "
+            "api.anthropic.com only."
+        ),
+    )
     vscode_init.add_argument(
         "--dry-run",
         action="store_true",
@@ -595,10 +628,16 @@ def _run_doctor(args: argparse.Namespace) -> int:
     risks AND broken providers can use a single command.
     """
     check_secrets = bool(getattr(args, "check_secrets", False))
-    if args.check_model is None and args.check_env is None and not check_secrets:
+    check_claude_code = bool(getattr(args, "check_claude_code", False))
+    if (
+        args.check_model is None
+        and args.check_env is None
+        and not check_secrets
+        and not check_claude_code
+    ):
         print(
             "doctor: provide --check-model PROVIDER, --check-env [PATH], "
-            "and/or --check-secrets",
+            "--check-secrets and/or --check-claude-code",
             file=sys.stderr,
         )
         return 1
@@ -614,6 +653,14 @@ def _run_doctor(args: argparse.Namespace) -> int:
     if check_secrets:
         worst_exit = max(worst_exit, _run_check_secrets(args.config))
 
+    if check_claude_code:
+        from coderouter.doctor_claude_code import format_report as _cc_format
+        from coderouter.doctor_claude_code import run_check as _cc_run
+
+        cc_report = _cc_run()
+        print(_cc_format(cc_report))
+        worst_exit = max(worst_exit, cc_report.exit_code)
+
     if args.check_model is not None:
         worst_exit = max(worst_exit, _run_check_model(args))
 
@@ -976,6 +1023,7 @@ def _run_vscode_init(args: argparse.Namespace) -> int:
     """
     from coderouter.vscode_init import (
         DEFAULT_PORT,
+        claude_code_env,
         exit_code_for,
         format_result,
         run_vscode_init,
@@ -991,6 +1039,22 @@ def _run_vscode_init(args: argparse.Namespace) -> int:
 
     port = args.port if args.port is not None else DEFAULT_PORT
 
+    extra_env: dict[str, str] | None = None
+    max_ctx = getattr(args, "max_context_tokens", None)
+    if getattr(args, "no_claude_code_env", False):
+        if max_ctx is not None:
+            print(
+                "vscode-init: --max-context-tokens has no effect with --no-claude-code-env",
+                file=sys.stderr,
+            )
+            return 2
+    else:
+        try:
+            extra_env = claude_code_env(max_context_tokens=max_ctx)
+        except ValueError as exc:
+            print(f"vscode-init: {exc}", file=sys.stderr)
+            return 2
+
     try:
         result = run_vscode_init(
             target,
@@ -999,6 +1063,7 @@ def _run_vscode_init(args: argparse.Namespace) -> int:
             with_envrc=args.with_envrc,
             dry_run=args.dry_run,
             force=args.force,
+            extra_env=extra_env,
         )
     except FileNotFoundError as exc:
         # Raised only when target vanishes between the pre-check above

{coderouter_cli-2.15.0 → coderouter_cli-2.16.0}/coderouter/config/schemas.py

@@ -1089,7 +1089,13 @@ class FallbackChain(BaseModel):
     #                (``context_budget_preserve_last_n``) are always
     #                kept, and tool_use / tool_result pairs are preserved
     #                atomically to avoid breaking agent loops.
-    context_budget_action: Literal["off", "warn", "trim"] = Field(
+    #   * ``reject`` — v2.16.0: ``warn`` + at ``context_budget_trim_threshold``
+    #                refuse the request with Anthropic's own overflow error
+    #                (HTTP 400, ``prompt is too long: ...``). Claude Code
+    #                answers that by compacting the conversation itself
+    #                (a summary, not a silent cut), so this is the
+    #                recommended action for Claude Code sessions.
+    context_budget_action: Literal["off", "warn", "trim", "reject"] = Field(
         default="off",
         description=(
             "v2.0-F (L1): action when estimated request tokens approach "
@@ -1097,7 +1103,10 @@ class FallbackChain(BaseModel):
             "disables the guard entirely. ``warn`` emits a log and "
             "response header at ``context_budget_warn_threshold``. "
             "``trim`` additionally removes old messages at "
-            "``context_budget_trim_threshold`` to reclaim context space."
+            "``context_budget_trim_threshold`` to reclaim context space. "
+            "``reject`` (v2.16.0) instead answers HTTP 400 "
+            "``prompt is too long`` at that threshold so the client "
+            "compacts the conversation itself (Claude Code does)."
         ),
     )
     context_budget_warn_threshold: float = Field(
@@ -1222,6 +1231,31 @@ class FallbackChain(BaseModel):
     # --- v2.0-H (L6): Mid-stream partial stitching --------------------------
     #   * ``off``      — discard partial content on mid-stream failure (legacy).
     #   * ``surface``  — return partial content as a truncated-but-valid response.
+    # v2.16.0 (tool search): Claude Code 2.1.304+ turns tool search on behind
+    # a custom ANTHROPIC_BASE_URL. The request then carries a server-side
+    # tool-search tool (``type: tool_search_tool_*``) and tools marked
+    # ``defer_loading: true``, which only api.anthropic.com executes. A
+    # local model sees the search tool as an ordinary function (and may
+    # "call" it), while every deferred MCP tool lands in its context anyway.
+    # Claude Code retries without tool search when the gateway answers 400.
+    #   * ``auto``        — ``passthrough`` when every provider in the chain
+    #                       is ``kind: anthropic`` on api.anthropic.com,
+    #                       otherwise ``reject`` (default).
+    #   * ``reject``      — answer 400 so the client turns tool search off.
+    #   * ``expand``      — drop the search tool and the ``defer_loading``
+    #                       flags; every tool is sent up front.
+    #   * ``passthrough`` — forward unchanged (pre-v2.16.0 behaviour).
+    deferred_tools_action: Literal["auto", "reject", "expand", "passthrough"] = Field(
+        default="auto",
+        description=(
+            "v2.16.0: what to do with Claude Code tool search "
+            "(``defer_loading`` tools / ``tool_search_tool_*``). ``auto`` "
+            "passes it through only for chains made entirely of "
+            "api.anthropic.com providers and rejects it (HTTP 400, which "
+            "makes Claude Code retry without tool search) otherwise."
+        ),
+    )
+
     partial_stitch_action: Literal["off", "surface"] = Field(
         default="off",
         description=(

{coderouter_cli-2.15.0 → coderouter_cli-2.16.0}/coderouter/doctor.py

@@ -635,6 +635,49 @@ _PROBE_TOOL_SPEC_ANTHROPIC = {
 }
 
 
+# v2.15.1: llama.cpp changed ``llama-server``'s default listen port from
+# 8080 to 9931 (ggml-org/llama.cpp#30159, build b11521, 2026-10-09). A server
+# started without ``--port`` on a new build is invisible to a provider that
+# still points at :8080, and the reverse holds for configs written against
+# the new default. When the basic probe cannot connect to one of those two
+# ports on a loopback host, check the other one so the transport error can
+# name the likely cause instead of looking like "server not running".
+_LLAMACPP_PORT_PAIR: dict[int, int] = {8080: 9931, 9931: 8080}
+_LOOPBACK_HOSTS = frozenset({"localhost", "127.0.0.1", "::1"})
+
+
+async def _llamacpp_port_hint(base_url: object) -> str | None:
+    """Return a hint when the other llama.cpp default port answers.
+
+    Only fires for loopback hosts on port 8080 or 9931. The alternate
+    port is probed at ``GET /health`` (llama-server's readiness route)
+    with a short timeout on the IPv4 literal — ``localhost`` can resolve
+    to ``::1`` first on macOS, where llama-server is not listening. The
+    config is never rewritten; the operator decides.
+    """
+    try:
+        parsed = httpx.URL(str(base_url))
+    except (httpx.InvalidURL, TypeError):
+        return None
+    if parsed.host not in _LOOPBACK_HOSTS or parsed.port not in _LLAMACPP_PORT_PAIR:
+        return None
+    other = _LLAMACPP_PORT_PAIR[parsed.port]
+    try:
+        async with httpx.AsyncClient(timeout=1.5) as client:
+            resp = await client.get(f"http://127.0.0.1:{other}/health")
+    except httpx.HTTPError:
+        return None
+    if resp.status_code >= 500:
+        return None
+    return (
+        f"Something is answering on port {other} instead. llama.cpp "
+        f"builds b11521+ listen on 9931 when started without --port "
+        f"(older builds: 8080). Either start llama-server with "
+        f"--port {parsed.port}, or change this provider's base_url to "
+        f"port {other}"
+    )
+
+
 async def _probe_auth_and_basic_chat(
     provider: ProviderConfig,
 ) -> ProbeResult:
@@ -670,10 +713,14 @@ async def _probe_auth_and_basic_chat(
     )
 
     if status is None:
+        detail = f"could not reach {url}: {raw}"
+        hint = await _llamacpp_port_hint(provider.base_url)
+        if hint:
+            detail = f"{detail}. {hint}"
         return ProbeResult(
             name="auth+basic-chat",
             verdict=ProbeVerdict.TRANSPORT_ERROR,
-            detail=f"could not reach {url}: {raw}",
+            detail=detail,
         )
 
     if status in (401, 403):