coderouter-cli 1.8.1__py3-none-any.whl → 1.8.2__py3-none-any.whl

coderouter/data/model-capabilities.yaml

@@ -176,47 +176,59 @@ rules:
   # 「Claude Code 代替として最高」「local champ」と評価されている。
   #
   # ただし v1.8.0 までで `claude_code_suitability: ok` を declare していた
-  # のは note 記事の伝聞ベースの先回り宣言で、v1.8.1 (2026-04-26) の
-  # 実機検証 (M3 Max 32GB / Ollama 0.21.2) で次の課題が判明:
-  #   - num_ctx を declare 32768 しても Ollama 側で silent に縮められる
-  #     (canary echo-back probe 失敗)
+  # のは note 記事の伝聞ベースの先回り宣言で、v1.8.1 〜 v1.8.2
+  # (2026-04-26) の実機検証 (M3 Max 64GB / Ollama 0.21.2) で:
+  #   - num_ctx と streaming の NEEDS_TUNING は v1.8.2 で thinking モデル
+  #     用 probe バジェット拡大により偽陽性と判明 (doctor 側の課題)
   #   - tool_calls probe が native tool_calls / 修復可能 JSON のいずれも
-  #     返さず NEEDS_TUNING
-  #   - streaming probe が finish_reason='length' で 0 chars 打ち切り
-  # これらは Ollama 経由特有の問題で、HF / vLLM 直接ロードなら違う可能性。
-  # 確証ない以上、`claude_code_suitability` は撤回し `tools` 宣言だけ残す。
+  #     返さない真の課題が残る (Qwen3.6 系の Ollama 経由 tool 仕様未成熟)
+  # tool_calls 不全が解消されるまで `claude_code_suitability` は撤回。
   # 実機で動いたユーザーは `~/.coderouter/model-capabilities.yaml` で
   # `claude_code_suitability: ok` を上書きできる。
   # ------------------------------------------------------------------
 
+  # v1.8.2: thinking: true は doctor probe (num_ctx / streaming) が reasoning
+  # トークン消費分の max_tokens 余裕を確保するためのヒント。Qwen3 系は
+  # /think モードで thinking トークンを吐く設計なので true 宣言。
   - match: "qwen3.6:*"
     kind: openai_compat
     capabilities:
       tools: true
+      thinking: true
 
   - match: "qwen/qwen3.6-*"
     kind: openai_compat
     capabilities:
       tools: true
+      thinking: true
 
   # ------------------------------------------------------------------
-  # Gemma 4 family (v1.7-B 追加)
+  # Gemma 4 family (v1.7-B 追加、v1.8.2 で thinking: true 宣言)
   #
   # Google 公式 Gemma 4。Ollama 公式 tag は gemma4:e2b / e4b / 26b / 31b、
   # 全 variant が tools+vision+thinking 対応、E2B/E4B は audio もサポート。
   # MoE (26b は active 3.8B / total 25.2B)。note 記事で「日常・バランスの
   # 王者」と評価。Claude Haiku 互換性に近い簡潔な応答スタイル。
+  #
+  # v1.8.2 (2026-04-26): 実機検証 (M3 Max 64GB / Ollama 0.21.2 / gemma4:26b)
+  # で `reasoning` フィールドにかなりの量のトークンを吐く thinking モデル
+  # と確認。doctor probe の max_tokens=32 / 128 が thinking トークンに
+  # 食い切られて偽陽性 NEEDS_TUNING を出していた。registry で
+  # `thinking: true` を宣言すると doctor が probe バジェットを 1024 まで
+  # 引き上げて偽陽性を回避する。
   # ------------------------------------------------------------------
 
   - match: "gemma4:*"
     kind: openai_compat
     capabilities:
       tools: true
+      thinking: true
 
   - match: "google/gemma-4*"
     kind: openai_compat
     capabilities:
       tools: true
+      thinking: true
 
   # ------------------------------------------------------------------
   # GLM family (Z.AI / Zhipu AI、v1.7-B 追加)

coderouter/doctor.py

@@ -433,6 +433,31 @@ _STREAMING_PROBE_USER_PROMPT = (
 # truncated". "1\n2\n...\n30" is ~80 chars; 40 chars covers the halfway
 # mark (1..20) which is already obviously-truncated territory.
 _STREAMING_PROBE_MIN_EXPECTED_CHARS = 40
+
+# v1.8.2: probe response budgets.
+#
+# Both num_ctx and streaming probes ask the model for a *short* answer
+# (the canary token / "1..30"). The original budgets (32 / 128 tokens)
+# assumed a non-thinking model that emits the answer immediately. On a
+# thinking model — Gemma 4 26B, Qwen3.6, gpt-oss, deepseek-r1 — the
+# upstream burns the entire budget on a hidden ``reasoning`` field
+# *before* emitting any visible ``content``, producing a false-positive
+# NEEDS_TUNING (canary missing / 0 chars streamed). Bumping the budget
+# is the cleanest fix: non-thinking models stop early at their natural
+# stop token (no waste), thinking models get headroom for the reasoning
+# trace plus the actual answer.
+#
+# Numbers picked from the v1.8.1 reality-check session
+# (docs/articles/note-v1-8-1-reality-check.md):
+#   * Gemma 4 26B reasoning prefix observed at ~150-300 tokens before
+#     content starts → 1024 covers reasoning + 30-line count comfortably.
+#   * Non-thinking baseline kept conservative-but-non-tight (256/512) to
+#     absorb stylistic preambles ("Sure, the answer is...") without
+#     burning extra cloud quota when the operator probes a paid endpoint.
+_NUM_CTX_PROBE_MAX_TOKENS_DEFAULT = 256
+_NUM_CTX_PROBE_MAX_TOKENS_THINKING = 1024
+_STREAMING_PROBE_MAX_TOKENS_DEFAULT = 512
+_STREAMING_PROBE_MAX_TOKENS_THINKING = 1024
 # Default ``num_predict`` suggested in the emitted patch. -1 would be
 # optimal (uncapped) but "4096" communicates intent more clearly to
 # operators unfamiliar with Ollama's sentinel value, and covers Claude
@@ -475,6 +500,42 @@ def _declared_num_ctx(provider: ProviderConfig) -> int | None:
     return val if isinstance(val, int) else None
 
 
+def _is_reasoning_model(
+    provider: ProviderConfig, resolved: ResolvedCapabilities
+) -> bool:
+    """v1.8.2: True iff the model is known to emit a hidden reasoning trace.
+
+    Thinking models (Gemma 4, Qwen3-with-/think, gpt-oss, deepseek-r1,
+    Claude Sonnet 4.5+ in extended-thinking mode) burn output tokens on a
+    ``reasoning`` field before any visible ``content`` is produced. The
+    num_ctx / streaming probes use small response budgets that get fully
+    consumed by the reasoning prefix, producing a false-positive
+    NEEDS_TUNING. Callers use this to choose a generous probe budget.
+
+    Three signals fire:
+      * provider declared ``capabilities.thinking: true`` in providers.yaml
+      * provider declared ``capabilities.reasoning_passthrough: true``
+        (the operator opted in to passing the raw reasoning to the client,
+        which is only meaningful for models that emit it)
+      * registry resolved ``thinking: true`` for this (kind, model) pair
+
+    Conservative bias — when both provider declaration and registry are
+    silent, treat as non-reasoning. The probe still completes for thinking
+    models in that case (they just hit ``finish_reason='length'`` like
+    they did pre-v1.8.2), but at least the new generous default budget
+    (256 / 512) gives more headroom than the old 32 / 128.
+    """
+    if provider.capabilities.thinking is True:
+        return True
+    if provider.capabilities.reasoning_passthrough is True:
+        return True
+    if resolved.thinking is True:
+        return True
+    if resolved.reasoning_passthrough is True:
+        return True
+    return False
+
+
 _PROBE_BASIC_USER_PROMPT = "Reply with exactly the single word: PONG"
 _PROBE_TOOLS_USER_PROMPT = (
     "You have one tool named `echo`. Call it with the argument "
@@ -617,7 +678,9 @@ def _extract_openai_assistant_choice(
     return msg if isinstance(msg, dict) else None
 
 
-async def _probe_num_ctx(provider: ProviderConfig) -> ProbeResult:
+async def _probe_num_ctx(
+    provider: ProviderConfig, resolved: ResolvedCapabilities
+) -> ProbeResult:
     """v1.0-B Probe — direct detection of Ollama ``num_ctx`` truncation.
 
     Addresses plan.md §9.4 symptom #1 (空応答 / 意味不明応答). Prior to
@@ -683,11 +746,21 @@ async def _probe_num_ctx(provider: ProviderConfig) -> ProbeResult:
     # whatever ``options.num_ctx`` the operator has declared. Request
     # fields win over extra_body, matching the adapter's merge order.
     body: dict[str, Any] = dict(provider.extra_body)
+    # v1.8.2: thinking models burn output tokens on a hidden ``reasoning``
+    # trace before emitting any ``content``. The pre-v1.8.2 default of 32
+    # was tight for any preamble at all; on Gemma 4 26B it caused
+    # ``finish_reason='length'`` with content="" before the canary could
+    # surface, producing a false-positive NEEDS_TUNING.
+    max_tokens = (
+        _NUM_CTX_PROBE_MAX_TOKENS_THINKING
+        if _is_reasoning_model(provider, resolved)
+        else _NUM_CTX_PROBE_MAX_TOKENS_DEFAULT
+    )
     body.update(
         {
             "model": provider.model,
             "messages": [{"role": "user", "content": user_prompt}],
-            "max_tokens": 32,
+            "max_tokens": max_tokens,
             "temperature": 0,
         }
     )
@@ -799,7 +872,9 @@ async def _probe_num_ctx(provider: ProviderConfig) -> ProbeResult:
     )
 
 
-async def _probe_streaming(provider: ProviderConfig) -> ProbeResult:
+async def _probe_streaming(
+    provider: ProviderConfig, resolved: ResolvedCapabilities
+) -> ProbeResult:
     """v1.0-C Probe — streaming completion path integrity.
 
     Addresses plan.md §9.4 symptom #1 from the **output** side. The v1.0-B
@@ -868,11 +943,18 @@ async def _probe_streaming(provider: ProviderConfig) -> ProbeResult:
     # probing. Top-level probe fields win on collision, matching adapter
     # merge order.
     body: dict[str, Any] = dict(provider.extra_body)
+    # v1.8.2: same thinking-model rationale as num_ctx probe — give
+    # reasoning a budget so the visible content has a chance to surface.
+    max_tokens = (
+        _STREAMING_PROBE_MAX_TOKENS_THINKING
+        if _is_reasoning_model(provider, resolved)
+        else _STREAMING_PROBE_MAX_TOKENS_DEFAULT
+    )
     body.update(
         {
             "model": provider.model,
             "messages": [{"role": "user", "content": _STREAMING_PROBE_USER_PROMPT}],
-            "max_tokens": 128,
+            "max_tokens": max_tokens,
             "temperature": 0,
             "stream": True,
         }
@@ -1506,11 +1588,11 @@ async def check_model(
     # declaration probes (tool_calls / thinking / reasoning-leak) should
     # dominate the report — streaming is the output-side sibling of
     # num_ctx and its NEEDS_TUNING verdict is orthogonal to the others.
-    report.results.append(await _probe_num_ctx(provider))
+    report.results.append(await _probe_num_ctx(provider, resolved))
     report.results.append(await _probe_tool_calls(provider, resolved))
     report.results.append(await _probe_thinking(provider, resolved))
     report.results.append(await _probe_reasoning_leak(provider, resolved))
-    report.results.append(await _probe_streaming(provider))
+    report.results.append(await _probe_streaming(provider, resolved))
     return report
 
 

{coderouter_cli-1.8.1.dist-info → coderouter_cli-1.8.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: coderouter-cli
-Version: 1.8.1
+Version: 1.8.2
 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

{coderouter_cli-1.8.1.dist-info → coderouter_cli-1.8.2.dist-info}/RECORD

@@ -2,7 +2,7 @@ coderouter/__init__.py,sha256=ghdjPrLtnRzY8fyQ4CJZI1UJKADyNTLtA3G7se8H7Ns,696
 coderouter/__main__.py,sha256=-LCgxJnvgUV240HjQKv7ly-mn2NuKHpC4nCpvTHjeSU,130
 coderouter/cli.py,sha256=vI1-dv10t4-xG6Zpt7zi_3U8xGgq54Qa8XIMUYpfOV8,19859
 coderouter/cli_stats.py,sha256=ae20xUr_hjX09Ms3fBZGZsUS52o44JC57EpbWLBOCO0,27750
-coderouter/doctor.py,sha256=Uf4R1t2-50hRx6UpV9UBe_UeiJ2UYBg7e7zkPUHegWE,64217
+coderouter/doctor.py,sha256=atYOr73LLI3lKHjhFDY0lea41_0jolfiY2zb15La_O8,68116
 coderouter/doctor_apply.py,sha256=r_J6xbu5-HivofPNriw4_vjNYs_VRs7GsGTS0oMEX10,24209
 coderouter/env_security.py,sha256=FEBZnXfJ0xE39kmMMn39zk0W_DRRnmcB_REmP9f4xWo,14796
 coderouter/errors.py,sha256=Xmq67lheyw8iv3Ox39jh2c4tvNI5RcUR4QkoxVDN6l4,1130
@@ -19,7 +19,7 @@ coderouter/config/env_file.py,sha256=CoMK27fuAXm-NtoLzXb8yN2E-wDFjHQuFwiIlmgTBQw
 coderouter/config/loader.py,sha256=FUEe8m4Tnmj_aul0vSctD8vKvNW-oLRoMRbTpSKqSmc,4077
 coderouter/config/schemas.py,sha256=NMWKstAzDKc6DT7qKr-GKlUFrsfE4HtEG7zkz9Pnq78,21597
 coderouter/data/__init__.py,sha256=uNyfD9jaCvTWsBAWtaw1Fr25OSxzv3psGMfBjT1z0Cc,328
-coderouter/data/model-capabilities.yaml,sha256=8g11ysbjEwt9Rc9Q9o4VCA0de2RmnH-mJEbnsueET_A,13949
+coderouter/data/model-capabilities.yaml,sha256=9f9CDIoWBJYtppMRh1BOS8FzRKzICV4uT-uANovGmzs,14718
 coderouter/ingress/__init__.py,sha256=WQsCH2CGJCAhy0mS6GSEdeYZRkkQu2OHDsP4CJWTLug,155
 coderouter/ingress/anthropic_routes.py,sha256=vuylsn7klFN-Dz3cBS7LrhnnSRGr6agipgMrr9gxq7k,8261
 coderouter/ingress/app.py,sha256=b9s6NpBGckaTQJCaw-6MZaY7Eq3rxgtX_cCkgWq-0Ig,6620
@@ -37,8 +37,8 @@ coderouter/translation/__init__.py,sha256=PYXN7XVEwpG1uC8RLy6fvnGbzEZhhrEuUapH8I
 coderouter/translation/anthropic.py,sha256=JpvIWNXHUPVqOGvps7o_6ZADhXuJuvpU7RdMqQFtwwM,6421
 coderouter/translation/convert.py,sha256=-qyzFzmmr9hhQV6_Sg75kJnvCZvHe3n7vRdaZtk_JqQ,47269
 coderouter/translation/tool_repair.py,sha256=fyxDb4kWHytO5JWq5y0i4tinJUtWqhMCkyfoCf5BjeM,8314
-coderouter_cli-1.8.1.dist-info/METADATA,sha256=sb0tldZ1vqseJCuO-oExjVXwnmOslwQt9MN9Rh5PVS8,44136
-coderouter_cli-1.8.1.dist-info/WHEEL,sha256=QccIxa26bgl1E6uMy58deGWi-0aeIkkangHcxk2kWfw,87
-coderouter_cli-1.8.1.dist-info/entry_points.txt,sha256=-dnLfD1YZ2WjH2zSdNCvlO65wYltM9bsHt9Fhg3yGss,51
-coderouter_cli-1.8.1.dist-info/licenses/LICENSE,sha256=wkEzoR86jFw33jvfOHjULqmkGEfxTFMgMaJnpR8mPRw,1065
-coderouter_cli-1.8.1.dist-info/RECORD,,
+coderouter_cli-1.8.2.dist-info/METADATA,sha256=MBkIOwnySR2wfw-JhBtGFipk_MqbV8J_4aAqWZ65A7g,44136
+coderouter_cli-1.8.2.dist-info/WHEEL,sha256=QccIxa26bgl1E6uMy58deGWi-0aeIkkangHcxk2kWfw,87
+coderouter_cli-1.8.2.dist-info/entry_points.txt,sha256=-dnLfD1YZ2WjH2zSdNCvlO65wYltM9bsHt9Fhg3yGss,51
+coderouter_cli-1.8.2.dist-info/licenses/LICENSE,sha256=wkEzoR86jFw33jvfOHjULqmkGEfxTFMgMaJnpR8mPRw,1065
+coderouter_cli-1.8.2.dist-info/RECORD,,