fix: reject conflicting raw reasoning overrides before sending

2026-09-09 01:26:34 -04:00
parent 1ee74c35a8
commit 8e61a66342
4 changed files with 213 additions and 110 deletions
@@ -247,52 +247,35 @@ class TestReasoningEffortPriority:
        assert "reasoning_effort" not in captured[0]

    @pytest.mark.parametrize(
-        ("provider", "model", "fragment"),
+        "provider,model,tier",
        [
-            ("qwen", "qwen-max", {"enable_thinking": True}),
-            ("deepseek", "deepseek-v4-pro", {"thinking": {"type": "enabled"}}),
-            ("zhipu", "glm-5.3", {"thinking": {"type": "enabled"}}),
-            ("moonshot", "kimi-k3", {"thinking": {"type": "enabled"}}),
+            ("deepseek", "deepseek-v4-pro", "high"),
+            ("zhipu", "glm-5.3", "low"),
+            ("moonshot", "kimi-k3", "low"),
+            ("minimax", "MiniMax-M3", "medium"),
        ],
    )
-    async def test_legacy_on_tier_matches_old_fragment(self, provider, model, fragment):
-        """存量 `ENABLE_THINKING=true` 的回归门: 发出去的字节逐字不变。
+    async def test_legacy_auto_requires_explicit_migration(self, provider, model, tier):
+        """旧糖配置明确拒绝，显式选择才能恢复可执行请求。"""
+        from dataclasses import replace

-        **只覆盖 `on_base` 自己就说全了"开"的四段**。openai/anthropic/google 的开档
-        旧版硬编码 `{"reasoning_effort": "medium"}`,新版不注入任何档位——那是设计
-        §4.2 声明过的**有意变更**(medium 在 GLM/kimi/deepseek 的档位表里根本不存在,
-        是库替下游做的档位判断),不是本门要守的不变量;这三家的模型经 OpenRouter
-        登记均为默认推理,不注入也仍是"开"。minimax 不在此列: 它的模型不满足该前提,
-        已按 issue #21 改回 medium,由下一条用例单独守。
-
-        qwen/deepseek 两条字面量逐字取自升级前的 `ProviderProfile.thinking_on`;
-        zhipu/moonshot 升级前没有对应段,断言的是它们 2026-09-04 登记的形态。
-        """
        captured = []
        source = _source(provider=provider, model=model, enable_thinking=True)
-        async with self._capturing_client(captured, sources=[source]) as client:
-            await client.chat([{"role": "user", "content": "hi"}])
-        body = captured[0]
-        assert {k: body[k] for k in fragment} == fragment
-        # `auto` = 开启但不指定强度: 语法糖不得替调用方挑一个档
-        assert "reasoning_effort" not in body
-
-    async def test_legacy_minimax_on_tier_actually_turns_reasoning_on(self):
-        """回归门(issue #21): minimax 段的存量 `ENABLE_THINKING=true` 必须真开推理。
-
-        本次换代一度把这段的开启形态改成 `on_base={}`(什么参数都不注入),依据是
-        "这些模型默认就推理,不注入也仍是'开'"。T10 真实网关实测推翻了该前提:
-        MiniMax-M3 不带任何推理参数时 5/5 轮**不推理**(六个强度值则全部生效)。
-        于是存量下游从"真开推理"静默变成"不推理",而 `resolve_thinking` 的 Phase 5
-        无条件放行 `auto`、能力表也堵不住这条路。
-
-        断言落在**发出去的字节**上而非中间态: 静默不推理这件事只有在请求体里才看得见。
-        """
-        captured = []
-        source = _source(provider="minimax", model="MiniMax-M3", enable_thinking=True)
-        async with self._capturing_client(captured, sources=[source]) as client:
-            await client.chat([{"role": "user", "content": "hi"}])
-        assert captured[0]["reasoning_effort"] == "medium"
+        client = self._capturing_client(captured, sources=[source])
+        try:
+            with pytest.raises(RequestRejectedError):
+                await client.chat([])
+            assert captured == []
+        finally:
+            await client._transport.aclose()
+        client = self._capturing_client(
+            captured, sources=[replace(source, enable_thinking=None, reasoning_effort=tier)]
+        )
+        try:
+            await client.chat([])
+            assert captured[0]["reasoning_effort"] == tier
+        finally:
+            await client._transport.aclose()


 class TestEffortFallbackWiring:
@@ -1280,3 +1263,56 @@ class TestTelemetryStatusExposure:
        assert _ocr_client().telemetry_status is None
        assert _ocr_client(telemetry=_Closable()).telemetry_status is None
        self._assert_snapshot(_ocr_client(telemetry=self._recorder(tmp_path)).telemetry_status)
+
+
+class TestManagedReasoningAdmission:
+    """入口前置与实际准入边界保持明确。"""
+
+    @pytest.mark.parametrize("factory", ["env", "settings"])
+    def test_factory_rejects_conflict_before_building_backends(self, monkeypatch, factory):
+        env = {
+            **_ENV,
+            "LLM__QWEN__1__MODEL": "qwen3.7-plus",
+            "LLM__QWEN__1__ENABLE_THINKING": "true",
+            "LLM__QWEN__1__EXTRA_BODY": '{"enable_thinking":true}',
+        }
+        built = []
+
+        def forbidden(*args, **kwargs):
+            built.append(True)
+            raise AssertionError("后端不得构造")
+
+        monkeypatch.setattr("polygateway.client._build_limiter", forbidden)
+        with pytest.raises(ValueError, match="冲突"):
+            if factory == "env":
+                GatewayClient.from_env(env=env)
+            else:
+                GatewayClient.from_settings(GatewaySettings.from_env(env=env))
+        assert not built
+
+    async def test_request_conflict_does_not_enter_onion(self):
+        client = _client()
+
+        async def forbidden(request):
+            raise AssertionError("不得进入洋葱")
+
+        client._handler = forbidden
+        try:
+            with pytest.raises(ValueError, match="冲突"):
+                await client.chat([], reasoning_effort="high", overlay={"reasoning_effort": "high"})
+        finally:
+            await client._transport.aclose()
+
+    async def test_source_conflict_releases_admitted_permit(self):
+        source = _source(
+            reasoning_effort="high", provider="openai", extra_body={"reasoning_effort": "high"}
+        )
+        sent = []
+        client = _client(sources=[source], handler=lambda request: sent.append(request) or _sse())
+        try:
+            with pytest.raises(RequestRejectedError, match="冲突"):
+                await client.chat([])
+            assert sent == []
+            assert (await client._limiter_backend.source_stats(source.name)).inflight == 0
+        finally:
+            await client._transport.aclose()