Replace temperature with reasoning effort and retry in-stream errors

- Providers get a default reasoning effort (low..max, empty = not sent);
  chat can override it per message. OpenAI sends reasoning_effort,
  Claude sends output_config.effort with adaptive thinking, and thinking
  blocks are replayed during the tool loop.
- Show a thinking indicator while waiting for model output.
- Surface in-stream upstream errors (e.g. overloaded) instead of
  breaking the SSE stream, and retry them up to 3 times when nothing
  has been emitted yet.

backend/src/alphapulse/api/chat.py

@@ -53,6 +53,8 @@ class SessionOut(BaseModel):
 class SendIn(BaseModel):
     text: Annotated[str, Field(min_length=1, max_length=20000)]
     provider_id: int | None = None
+    # 本轮思考强度；为空时用模型配置里的默认值
+    reasoning_effort: llm.Effort | None = None
 
 
 class ResetIn(BaseModel):
@@ -121,7 +123,7 @@ async def get_chat_session(session: Annotated[AsyncSession, Depends(get_session)
 
 @router.post("/messages", response_class=EventSourceResponse)
 async def send_message(body: SendIn) -> AsyncIterator[ServerSentEvent]:
-    """事件：prompt、delta、tool_call、tool_result、skill_draft、context_updated、usage、error、done"""
+    """事件：prompt、thinking、delta、tool_call、tool_result、skill_draft、context_updated、usage、error、done"""
     if generating.locked():
         yield event("error", {"message": "上一轮还在生成，请等待完成或先停止"})
         return
@@ -148,14 +150,17 @@ async def send_message(body: SendIn) -> AsyncIterator[ServerSentEvent]:
         await save_messages(ChatMessage(session_id=current.id, role="user", content={"text": body.text}))
 
         specs = tool_specs()
+        effort = body.reasoning_effort or provider.reasoning_effort
         for round_no in range(MAX_TOOL_ROUNDS + 1):
             text = ""
             completion: llm.Completion | None = None
             try:
-                async for item in llm.stream(provider, system, messages, specs if round_no < MAX_TOOL_ROUNDS else []):
+                async for item in llm.stream(provider, system, messages, specs if round_no < MAX_TOOL_ROUNDS else [], effort):
                     if isinstance(item, llm.TextDelta):
                         text += item.text
                         yield event("delta", {"text": item.text})
+                    elif isinstance(item, llm.ThinkingStarted):
+                        yield event("thinking", {})
                     else:
                         completion = item
             except llm.LlmError as e:
@@ -170,6 +175,8 @@ async def send_message(body: SendIn) -> AsyncIterator[ServerSentEvent]:
                 raise
 
             assistant = {"text": completion.text, "tool_calls": completion.tool_calls}
+            if completion.thinking_blocks:
+                assistant["thinking_blocks"] = completion.thinking_blocks
             await save_messages(ChatMessage(session_id=current.id, role="assistant", content=assistant, provider_id=provider.id, usage=completion.usage))
             messages.append({"role": "assistant", **assistant})
             yield event("usage", completion.usage)

backend/src/alphapulse/api/settings.py

@@ -30,7 +30,7 @@ class ProviderOut(BaseModel):
     model: str
     max_tokens: int
     context_window: int
-    temperature: float | None
+    reasoning_effort: llm.Effort | None
     is_default: bool
     updated_at: datetime
 
@@ -43,7 +43,7 @@ class ProviderIn(BaseModel):
     model: Annotated[str, Field(min_length=1, max_length=128)]
     max_tokens: Annotated[int, Field(ge=256, le=128000)] = 8192
     context_window: Annotated[int, Field(ge=4096, le=2000000)] = 200000
-    temperature: Annotated[float | None, Field(ge=0, le=2)] = None
+    reasoning_effort: llm.Effort | None = None
     is_default: bool = False
 
 
@@ -56,7 +56,8 @@ class ProviderPatch(BaseModel):
     model: Annotated[str | None, Field(min_length=1, max_length=128)] = None
     max_tokens: Annotated[int | None, Field(ge=256, le=128000)] = None
     context_window: Annotated[int | None, Field(ge=4096, le=2000000)] = None
-    temperature: Annotated[float | None, Field(ge=0, le=2)] = None
+    # 显式传 null 表示不传思考强度
+    reasoning_effort: llm.Effort | None = None
     is_default: bool | None = None
 
 
@@ -111,9 +112,9 @@ async def create_provider(body: ProviderIn, session: Annotated[AsyncSession, Dep
 async def patch_provider(provider_id: int, body: ProviderPatch, session: Annotated[AsyncSession, Depends(get_session)]) -> ProviderOut:
     provider = await get_provider(session, provider_id)
     values = body.model_dump(exclude_unset=True, exclude={"api_key", "base_url", "is_default"})
-    # temperature 显式传 null 表示清除，其他字段为 null 时忽略
+    # reasoning_effort 显式传 null 表示清除，其他字段为 null 时忽略
     for field, value in values.items():
-        if value is not None or field == "temperature":
+        if value is not None or field == "reasoning_effort":
             setattr(provider, field, value)
     if body.base_url is not None:
         provider.base_url = str(body.base_url)
@@ -147,7 +148,7 @@ async def test_provider(provider_id: int, session: Annotated[AsyncSession, Depen
     started = time.monotonic()
     try:
         completion = None
-        async for event in llm.stream(provider, "只回复 OK。", [{"role": "user", "text": "ping"}], []):
+        async for event in llm.stream(provider, "只回复 OK。", [{"role": "user", "text": "ping"}], [], provider.reasoning_effort):
             if isinstance(event, llm.Completion):
                 completion = event
     except llm.LlmError as e: