feat: carry the per-call tier down to the transport that must send it

The Transport port took the request apart into five arguments, so a tier
placed on ChatRequest could never reach _build_payload: the field was set,
read by nobody, and silently ignored - the exact shape of failure that sent
downstream to extra_body in the first place.

complete() now takes reasoning_effort with no default, matching the
TelemetryRecorder convention: a default would turn a missing hand-off into
a silent 'no opinion'. All four fakes move with it, since @runtime_checkable
checks method names and not signatures.

EmbeddingTransport and OcrTransport are deliberately left alone - they have
no reasoning semantics - and a test now holds that line.

_build_payload drops its inline sugar conversion for effective_effort(), so
the guard and the hot path share one judgement, and passes the source's
effort_fallback for the same reason.

src/polygateway/middleware/retry.py

@@ -285,6 +285,9 @@ class RetryMW:
                 stream=request.stream,
                 overlay=request.overlay,
                 call_id=call_id,
+                # 逐次尝试原样重传: 换源不改变调用方要的档位(源级默认由 transport
+                # 自己按选中的源解析,两者在 effective_effort 里汇合)
+                reasoning_effort=request.reasoning_effort,
             )
             if result.usage_source == "unavailable":
                 # 用量不可得时按入场预扣量结算(delta==0),否则押金会被整笔退回,

src/polygateway/ports.py

@@ -12,6 +12,7 @@ from typing import Any, Protocol, runtime_checkable
 
 from .types import (
     ChatRequest,
+    Effort,
     EmbeddingTransportResult,
     LLMResponse,
     OcrLayoutResult,
@@ -36,7 +37,16 @@ class Middleware(Protocol):
 
 @runtime_checkable
 class Transport(Protocol):
-    """一次原始调用的协议细节(请求组装/流式解析/错误翻译);不含任何治理。"""
+    """一次原始调用的协议细节(请求组装/流式解析/错误翻译);不含任何治理。
+
+    `reasoning_effort` 是本次调用要求的推理档位(`None` = 不表态,随源级配置)。
+    它必须走**协议参数**而不能让 transport 自己去读 `ChatRequest`: 端口只收拆开的
+    请求要素,是为了让 transport 不依赖洋葱内部的请求类型(P7 端口最内层)。
+
+    该参数**不设默认值**,与 `TelemetryRecorder.record_llm_call` 同一既有约定:
+    库外无第三方实现者,写全签名的成本为零,而默认值会把"某一层漏传"变成静默的
+    "调用方没表态"——一次本该报错的漏配就此变成一次悄悄涨价的调用。
+    """
 
     async def complete(
         self,
@@ -46,6 +56,7 @@ class Transport(Protocol):
         stream: bool,
         overlay: dict[str, Any],
         call_id: str,
+        reasoning_effort: Effort | None,
     ) -> TransportResult: ...
 
 

src/polygateway/transports/openai_compat.py

@@ -28,6 +28,7 @@ from polygateway.streaming import StreamLivenessTimeout, stream_with_liveness_ti
 from polygateway.thinking import (
     ThinkingCapability,
     ThinkingUnsupportedError,
+    effective_effort,
     get_capability,
     observe_thinking,
     reconcile_thinking,
@@ -351,6 +352,7 @@ class OpenAICompatTransport:
         profile: ProviderProfile,
         stream: bool,
         overlay: dict[str, Any],
+        reasoning_effort: Effort | None,
     ) -> dict[str, Any]:
         payload: dict[str, Any] = {"model": source.model, "messages": messages, "stream": stream}
         if stream:
@@ -360,17 +362,19 @@ class OpenAICompatTransport:
         capability = get_capability(source.model, table=self._capabilities)
         first_time = source.model not in self._warned_models
         self._warned_models.add(source.model)
-        # `enable_thinking` 的档位语法糖(True → auto,False → none,None 不表态);
+        # 三层优先级在此汇合: 请求级 > 源级 > enable_thinking 语法糖(设计 §4.2)。
-        # 就地转换是过渡形态,T5 起由 thinking.effective_effort() 统一收口并接上
+        # 判定与装配守卫共用同一个纯函数,两处分叉就会变成"装配期放行、运行期报错"
-        # 源级/请求级档位(设计 §4.2)
-        enabled = source.enable_thinking
-        effort = None if enabled is None else (Effort.AUTO if enabled else Effort.NONE)
         payload.update(
             resolve_thinking(
                 profile,
                 capability,
-                effort,
+                effective_effort(
+                    request_effort=reasoning_effort,
+                    source_effort=source.reasoning_effort,
+                    enable_thinking=source.enable_thinking,
+                ),
                 model=source.model,
+                fallback=source.effort_fallback,
                 warn_unregistered=first_time,
             ).payload
         )
@@ -388,12 +392,22 @@ class OpenAICompatTransport:
         stream: bool,
         overlay: dict[str, Any],
         call_id: str,
+        reasoning_effort: Effort | None,
     ) -> TransportResult:
-        """一次原始调用;HTTP/线路/流式异常按 ARCH §6.2 翻译为领域错误。"""
+        """一次原始调用;HTTP/线路/流式异常按 ARCH §6.2 翻译为领域错误。
+
+        `reasoning_effort` 是**请求级**档位(`None` = 不表态);它与源级配置的优先级
+        在 `_build_payload` 里由 `effective_effort` 裁定,本层只负责把它送到。
+        """
         profile = get_provider(source.provider, registry=self._registry)
         try:
             payload = self._build_payload(
-                messages=messages, source=source, profile=profile, stream=stream, overlay=overlay
+                messages=messages,
+                source=source,
+                profile=profile,
+                stream=stream,
+                overlay=overlay,
+                reasoning_effort=reasoning_effort,
             )
         except ThinkingUnsupportedError as exc:
             # 推理开关不可满足是**请求本身**的问题: 换源重试都救不了它。只捕这个

tests/e2e/test_thinking_live.py

@@ -514,6 +514,7 @@ class TestAssemblyGuardAgainstRealConfig:
                     stream=True,
                     overlay={},
                     call_id="e2e-guard",
+                    reasoning_effort=None,
                 )
         finally:
             await client.aclose()

tests/integration/test_redis_cross_connection.py

@@ -73,7 +73,7 @@ class ScriptedTransport:
         self.hang = hang
         self.calls: list[str] = []
 
-    async def complete(self, *, messages, source, stream, overlay, call_id):
+    async def complete(self, *, messages, source, stream, overlay, call_id, reasoning_effort):
         self.calls.append(source.name)
         if self.hang:
             await asyncio.Event().wait()

tests/unit/test_backpressure.py

@@ -210,7 +210,7 @@ class ClockAdvancingTransport:
         self.clock = clock
         self.calls = []
 
-    async def complete(self, *, messages, source, stream, overlay, call_id):
+    async def complete(self, *, messages, source, stream, overlay, call_id, reasoning_effort):
         self.calls.append((source.name, call_id))
         advance, action = self.script.pop(0)
         self.clock.advance(advance)

tests/unit/test_client.py

@@ -24,6 +24,7 @@ from polygateway.transports.openai_compat import OpenAICompatTransport
 from polygateway.types import (
     BackpressurePolicy,
     BreakerConfig,
+    Effort,
     GlobalLimits,
     RetryPolicy,
     SourceConfig,
@@ -219,6 +220,31 @@ class TestReasoningEffortPriority:
         overrides.setdefault("model", "glm-5.3")
         return _source(provider="zhipu", **overrides)
 
+    async def test_request_effort_wins_over_source(self):
+        captured = []
+        source = self._zhipu(reasoning_effort=Effort.LOW)
+        async with self._capturing_client(captured, sources=[source]) as client:
+            await client.chat([{"role": "user", "content": "hi"}], reasoning_effort=Effort.MAX)
+        assert captured[0]["reasoning_effort"] == "max"
+        assert captured[0]["thinking"] == {"type": "enabled"}
+
+    async def test_none_request_does_not_clear_source(self):
+        """请求级不表态 ≠ 请求级要求"不推理": 前者必须让源级默认继续生效。"""
+        captured = []
+        source = self._zhipu(reasoning_effort=Effort.LOW)
+        async with self._capturing_client(captured, sources=[source]) as client:
+            await client.chat([{"role": "user", "content": "hi"}])
+        assert captured[0]["reasoning_effort"] == "low"
+
+    async def test_request_none_tier_is_an_opinion_not_an_absence(self):
+        """请求级 `none` 是"要求不推理",不得被当成"没表态"而回落到源级档位。"""
+        captured = []
+        source = self._zhipu(model="glm-5.2", reasoning_effort=Effort.MAX)
+        async with self._capturing_client(captured, sources=[source]) as client:
+            await client.chat([{"role": "user", "content": "hi"}], reasoning_effort=Effort.NONE)
+        assert captured[0]["thinking"] == {"type": "disabled"}
+        assert "reasoning_effort" not in captured[0]
+
     @pytest.mark.parametrize(
         ("provider", "model", "fragment"),
         [

tests/unit/test_openai_compat.py

@@ -69,13 +69,14 @@ def _transport_for(handler, *, registry=None):
     )
 
 
-async def _complete(transport, source, *, stream=True, overlay=None):
+async def _complete(transport, source, *, stream=True, overlay=None, reasoning_effort=None):
     return await transport.complete(
         messages=[{"role": "user", "content": "hi"}],
         source=source,
         stream=stream,
         overlay=overlay or {},
         call_id="cid-1",
+        reasoning_effort=reasoning_effort,
     )
 
 

tests/unit/test_ports.py

@@ -1,15 +1,18 @@
 """ports.py 端口冻结测试(M1 设计 §4): Protocol 结构性检查 + Gate 快照校验。"""
 
+import inspect
 from typing import Any
 
 import pytest
 
 from polygateway.ports import (
     CacheBackend,
+    EmbeddingTransport,
     GateDecision,
     GateState,
     GateUpdate,
     Middleware,
+    OcrTransport,
     Permit,
     ProviderGate,
     RateLimiter,
@@ -69,7 +72,7 @@ class _DummyMw:
 
 
 class _DummyTransport:
-    async def complete(self, *, messages, source, stream, overlay, call_id):
+    async def complete(self, *, messages, source, stream, overlay, call_id, reasoning_effort):
         raise NotImplementedError
 
 
@@ -142,6 +145,34 @@ def test_protocols_are_runtime_checkable(impl, protocol):
     assert isinstance(impl, protocol)
 
 
+class TestReasoningTierIsOnlyOnTheChatPort:
+    """档位属于 chat 端口,且**只属于**它(Task 5b)。
+
+    `@runtime_checkable` 只查方法名不查签名,故协议签名本身必须被显式断言——
+    否则实现漏改一个参数,要到运行期调用才会以 `TypeError` 现形,而那时的现场
+    离根因已经很远。
+    """
+
+    def test_chat_transport_carries_the_per_call_tier(self):
+        params = inspect.signature(Transport.complete).parameters
+        assert "reasoning_effort" in params
+        # 不给默认值是有意的(与 TelemetryRecorder 同一既有约定): 库外无第三方
+        # 实现者,写全签名成本为零,而默认值会把"漏传"变成静默的"不表态"
+        assert params["reasoning_effort"].default is inspect.Parameter.empty
+
+    @pytest.mark.parametrize(
+        ("protocol", "method"),
+        [
+            (EmbeddingTransport, "embed"),
+            (OcrTransport, "recognize_text"),
+            (OcrTransport, "parse_layout"),
+        ],
+    )
+    def test_other_transports_have_no_reasoning_tier(self, protocol, method):
+        """embedding 与 OCR 没有推理语义,给它们加档位只会静默无效(issue #4 同款决策)。"""
+        assert "reasoning_effort" not in inspect.signature(getattr(protocol, method)).parameters
+
+
 class _DummyStatusProvider(_DummyRecorder):
     @property
     def telemetry_status(self) -> TelemetryStatus:

tests/unit/test_retry.py

@@ -24,6 +24,7 @@ from polygateway.types import (
     BackpressurePolicy,
     BreakerConfig,
     ChatRequest,
+    Effort,
     GlobalLimits,
     RetryPolicy,
     SourceConfig,
@@ -63,14 +64,21 @@ def _ok(content="ok"):
 
 
 class FakeTransport:
-    """按脚本逐次返回结果或抛异常;记录每次 (source_name, call_id)。"""
+    """按脚本逐次返回结果或抛异常;记录每次 (source_name, call_id) 与收到的档位。
+
+    `reasoning_effort` 刻意**不给默认值**,与 `Transport` 协议保持逐字一致:
+    `@runtime_checkable` 只查方法名不查签名,fake 上多一个默认值就会把"中间件漏传"
+    这类缺口伪装成"调用方没表态",而报错现场离根因很远。
+    """
 
     def __init__(self, script):
         self.script = list(script)
         self.calls = []
+        self.efforts = []
 
-    async def complete(self, *, messages, source, stream, overlay, call_id):
+    async def complete(self, *, messages, source, stream, overlay, call_id, reasoning_effort):
         self.calls.append((source.name, call_id))
+        self.efforts.append(reasoning_effort)
         action = self.script.pop(0)
         if isinstance(action, Exception):
             raise action
@@ -135,6 +143,37 @@ def _harness(
 _REQ = ChatRequest(messages=[{"role": "user", "content": "hi"}])
 
 
+class TestRequestTierReachesTransport:
+    """请求级档位必须一路穿过洋葱到达 transport(Task 5b)。
+
+    `ChatRequest` 上填了字段而中间件不搬运,是"看起来配了、实际没发出去"的静默
+    失效——正是 issue #20 里下游改用 extra_body 绕过治理的成因。
+    """
+
+    async def test_request_tier_reaches_transport(self):
+        mw, _, _, transport, *_ = _harness([_src("a")], [_ok()])
+        await mw(
+            ChatRequest(messages=[{"role": "user", "content": "hi"}], reasoning_effort=Effort.HIGH)
+        )
+        assert transport.efforts == [Effort.HIGH]
+
+    async def test_absent_tier_is_carried_as_none(self):
+        """不表态也要显式传下去: 漏传与"传了 None"在协议上必须区分不开才安全。"""
+        mw, _, _, transport, *_ = _harness([_src("a")], [_ok()])
+        await mw(_REQ)
+        assert transport.efforts == [None]
+
+    async def test_tier_is_carried_on_every_retry_attempt(self):
+        """换源重试时档位不得在第二次尝试上丢失。"""
+        mw, _, _, transport, *_ = _harness(
+            [_src("a"), _src("b")], [TransientError("boom", source_name="a"), _ok()]
+        )
+        await mw(
+            ChatRequest(messages=[{"role": "user", "content": "hi"}], reasoning_effort=Effort.LOW)
+        )
+        assert transport.efforts == [Effort.LOW, Effort.LOW]
+
+
 class TestSuccessPath:
     async def test_first_attempt_success_builds_response(self):
         mw, limiter, gate, transport, sleep, _ = _harness([_src("a")], [_ok("hello")])

tests/unit/test_usage_source_domain.py

@@ -134,6 +134,7 @@ async def test_salvage_override_stays_in_domain(usage):
         stream=True,
         overlay={},
         call_id="cid",
+        reasoning_effort=None,
     )
     assert result.usage_source in USAGE_SOURCES