fix(cli): 统一 CLI handler pending usage 的 api_format 来源

CLI stream/sync mixin 中创建 pending usage 记录时，api_format 取值
来源不一致（部分用 FORMAT_ID，部分用 allowed_api_formats[0]）。
新增 primary_api_format 属性并统一使用，确保 pending 记录的格式
与实际客户端请求格式一致。

src/api/handlers/base/cli_protocol.py

@@ -49,6 +49,7 @@ class CliHandlerProtocol(Protocol):
     user_agent: str
     start_time: float
     allowed_api_formats: list[str]
+    primary_api_format: str
     redis: Redis  # type: ignore[type-arg]
     telemetry: MessageTelemetry
     perf_metrics: dict[str, Any] | None

src/api/handlers/base/cli_stream_mixin.py

@@ -99,13 +99,14 @@ class CliStreamMixin:
         # 使用子类实现的方法提取 model（不同 API 格式的 model 位置不同）
         # 注意：使用 original_request_body，因为整流只修改 messages，不影响 model 字段
         model = self.extract_model_from_request(original_request_body, path_params)
+        client_api_format = self.primary_api_format
 
         # 提前创建 pending 记录，让前端可以立即看到"处理中"
         self._create_pending_usage(
             model=model,
             is_stream=True,
             request_type="chat",
-            api_format=self.FORMAT_ID,
+            api_format=client_api_format,
             request_headers=original_headers,
             request_body=original_request_body,
         )
@@ -113,7 +114,7 @@ class CliStreamMixin:
         # 创建流上下文
         ctx = StreamContext(
             model=model,
-            api_format=self.allowed_api_formats[0],
+            api_format=client_api_format,
             api_family=self.api_family,
             endpoint_kind=self.endpoint_kind,
             request_id=self.request_id,

src/api/handlers/base/cli_sync_mixin.py

@@ -78,7 +78,7 @@ class CliSyncMixin:
 
         # 使用子类实现的方法提取 model（不同 API 格式的 model 位置不同）
         model = self.extract_model_from_request(original_request_body, path_params)
-        api_format = self.allowed_api_formats[0]
+        api_format = self.primary_api_format
         sync_start_time = time.time()
 
         # 提前创建 pending 记录，让前端可以立即看到"处理中"
@@ -86,7 +86,7 @@ class CliSyncMixin:
             model=model,
             is_stream=False,
             request_type="chat",
-            api_format=self.FORMAT_ID,
+            api_format=api_format,
             request_headers=original_headers,
             request_body=original_request_body,
         )

tests/api/handlers/base/test_cli_pending_usage_format.py

@@ -0,0 +1,74 @@
+from __future__ import annotations
+
+import pytest
+
+from src.api.handlers.base.cli_stream_mixin import CliStreamMixin
+from src.api.handlers.base.cli_sync_mixin import CliSyncMixin
+
+
+class _StopExecution(Exception):
+    pass
+
+
+class _DummySyncHandler(CliSyncMixin):
+    FORMAT_ID = "openai:cli"
+
+    def __init__(self) -> None:
+        self.allowed_api_formats = ["openai:compact"]
+        self.primary_api_format = "openai:compact"
+        self.pending_calls: list[dict[str, object]] = []
+
+    def extract_model_from_request(
+        self, request_body: dict[str, object], path_params: dict[str, object] | None
+    ) -> str:
+        return str(request_body.get("model") or "unknown")
+
+    def _create_pending_usage(self, **kwargs: object) -> None:
+        self.pending_calls.append(kwargs)
+        raise _StopExecution()
+
+
+class _DummyStreamHandler(CliStreamMixin):
+    FORMAT_ID = "openai:cli"
+
+    def __init__(self) -> None:
+        self.allowed_api_formats = ["openai:compact"]
+        self.primary_api_format = "openai:compact"
+        self.pending_calls: list[dict[str, object]] = []
+
+    def extract_model_from_request(
+        self, request_body: dict[str, object], path_params: dict[str, object] | None
+    ) -> str:
+        return str(request_body.get("model") or "unknown")
+
+    def _create_pending_usage(self, **kwargs: object) -> None:
+        self.pending_calls.append(kwargs)
+        raise _StopExecution()
+
+
+@pytest.mark.asyncio
+async def test_sync_pending_usage_uses_primary_api_format() -> None:
+    handler = _DummySyncHandler()
+
+    with pytest.raises(_StopExecution):
+        await handler.process_sync(  # type: ignore[misc]
+            original_request_body={"model": "gpt-5.3-codex"},
+            original_headers={},
+        )
+
+    assert handler.pending_calls
+    assert handler.pending_calls[0]["api_format"] == "openai:compact"
+
+
+@pytest.mark.asyncio
+async def test_stream_pending_usage_uses_primary_api_format() -> None:
+    handler = _DummyStreamHandler()
+
+    with pytest.raises(_StopExecution):
+        await handler.process_stream(  # type: ignore[misc]
+            original_request_body={"model": "gpt-5.3-codex"},
+            original_headers={},
+        )
+
+    assert handler.pending_calls
+    assert handler.pending_calls[0]["api_format"] == "openai:compact"