perf: support extended thinking for Anthropic, DeepSeek reasoning mode, and Gemini text part thought signatures to improve multi-turn reasoning performance.

astrbot/builtin_stars/astrbot/main.py

@@ -100,16 +100,8 @@ async def decorate_llm_req(self, event: AstrMessageEvent, req: ProviderRequest):
                 logger.error(f"ltm: {e}")
 
     @filter.on_llm_response()
-    async def inject_reasoning(self, event: AstrMessageEvent, resp: LLMResponse):
-        """在 LLM 响应后基于配置注入思考过程文本 / 在 LLM 响应后记录对话"""
-        umo = event.unified_msg_origin
-        cfg = self.context.get_config(umo).get("provider_settings", {})
-        show_reasoning = cfg.get("display_reasoning_text", False)
-        if show_reasoning and resp.reasoning_content:
-            resp.completion_text = (
-                f"🤔 思考: {resp.reasoning_content}\n\n{resp.completion_text}"
-            )
-
+    async def record_llm_resp_to_ltm(self, event: AstrMessageEvent, resp: LLMResponse):
+        """在 LLM 响应后记录对话"""
         if self.ltm and self.ltm_enabled(event):
             try:
                 await self.ltm.after_req_llm(event, resp)

astrbot/core/agent/message.py

@@ -12,7 +12,7 @@ class ContentPart(BaseModel):
 
     __content_part_registry: ClassVar[dict[str, type["ContentPart"]]] = {}
 
-    type: str
+    type: Literal["text", "think", "image_url", "audio_url"]
 
     def __init_subclass__(cls, **kwargs: Any) -> None:
         super().__init_subclass__(**kwargs)
@@ -63,6 +63,28 @@ class TextPart(ContentPart):
     text: str
 
 
+class ThinkPart(ContentPart):
+    """
+    >>> ThinkPart(think="I think I need to think about this.").model_dump()
+    {'type': 'think', 'think': 'I think I need to think about this.', 'encrypted': None}
+    """
+
+    type: str = "think"
+    think: str
+    encrypted: str | None = None
+    """Encrypted thinking content, or signature."""
+
+    def merge_in_place(self, other: Any) -> bool:
+        if not isinstance(other, ThinkPart):
+            return False
+        if self.encrypted:
+            return False
+        self.think += other.think
+        if other.encrypted:
+            self.encrypted = other.encrypted
+        return True
+
+
 class ImageURLPart(ContentPart):
     """
     >>> ImageURLPart(image_url="http://example.com/image.jpg").model_dump()

astrbot/core/agent/runners/tool_loop_agent_runner.py

@@ -13,6 +13,7 @@
 )
 
 from astrbot import logger
+from astrbot.core.agent.message import TextPart, ThinkPart
 from astrbot.core.message.components import Json
 from astrbot.core.message.message_event_result import (
     MessageChain,
@@ -169,13 +170,20 @@ async def step(self):
             self.final_llm_resp = llm_resp
             self._transition_state(AgentState.DONE)
             self.stats.end_time = time.time()
+
             # record the final assistant message
-            self.run_context.messages.append(
-                Message(
-                    role="assistant",
-                    content=llm_resp.completion_text or "*No response*",
-                ),
-            )
+            parts = []
+            if llm_resp.reasoning_content or llm_resp.reasoning_signature:
+                parts.append(
+                    ThinkPart(
+                        think=llm_resp.reasoning_content,
+                        encrypted=llm_resp.reasoning_signature,
+                    )
+                )
+            parts.append(TextPart(text=llm_resp.completion_text or "*No response*"))
+            self.run_context.messages.append(Message(role="assistant", content=parts))
+
+            # call the on_agent_done hook
             try:
                 await self.agent_hooks.on_agent_done(self.run_context, llm_resp)
             except Exception as e:
@@ -214,10 +222,19 @@ async def step(self):
                         data=AgentResponseData(chain=result),
                     )
             # 将结果添加到上下文中
+            parts = []
+            if llm_resp.reasoning_content or llm_resp.reasoning_signature:
+                parts.append(
+                    ThinkPart(
+                        think=llm_resp.reasoning_content,
+                        encrypted=llm_resp.reasoning_signature,
+                    )
+                )
+            parts.append(TextPart(text=llm_resp.completion_text or "*No response*"))
             tool_calls_result = ToolCallsResult(
                 tool_calls_info=AssistantMessageSegment(
                     tool_calls=llm_resp.to_openai_to_calls_model(),
-                    content=llm_resp.completion_text,
+                    content=parts,
                 ),
                 tool_calls_result=tool_call_result_blocks,
             )

astrbot/core/astr_agent_hooks.py

@@ -13,6 +13,12 @@
 class MainAgentHooks(BaseAgentRunHooks[AstrAgentContext]):
     async def on_agent_done(self, run_context, llm_response):
         # 执行事件钩子
+        if llm_response and llm_response.reasoning_content:
+            # we will use this in result_decorate stage to inject reasoning content to chain
+            run_context.context.event.set_extra(
+                "_llm_reasoning_content", llm_response.reasoning_content
+            )
+
         await call_event_hook(
             run_context.context.event,
             EventType.OnLLMResponseEvent,

astrbot/core/config/default.py

@@ -905,6 +905,7 @@ class ChatProviderTemplate(TypedDict):
                         "key": [],
                         "api_base": "https://api.anthropic.com/v1",
                         "timeout": 120,
+                        "anth_thinking_config": {"budget": 0},
                     },
                     "Moonshot": {
                         "id": "moonshot",
@@ -920,7 +921,7 @@ class ChatProviderTemplate(TypedDict):
                     "xAI": {
                         "id": "xai",
                         "provider": "xai",
-                        "type": "openai_chat_completion",
+                        "type": "xai_chat_completion",
                         "provider_type": "chat_completion",
                         "enable": True,
                         "key": [],
@@ -1787,6 +1788,17 @@ class ChatProviderTemplate(TypedDict):
                             },
                         },
                     },
+                    "anth_thinking_config": {
+                        "description": "Thinking Config",
+                        "type": "object",
+                        "items": {
+                            "budget": {
+                                "description": "Thinking Budget",
+                                "type": "int",
+                                "hint": "Anthropic thinking.budget_tokens param. Must >= 1024. See: https://platform.claude.com/docs/en/build-with-claude/extended-thinking",
+                            },
+                        },
+                    },
                     "minimax-group-id": {
                         "type": "string",
                         "description": "用户组",