merge

Author: elecvoid243

astrbot/core/agent/runners/tool_loop_agent_runner.py

@@ -16,11 +16,18 @@
     TextContent,
     TextResourceContents,
 )
+from tenacity import (
+    AsyncRetrying,
+    retry_if_exception_type,
+    stop_after_attempt,
+    wait_exponential,
+)
 
 from astrbot import logger
 from astrbot.core.agent.message import ImageURLPart, TextPart, ThinkPart
 from astrbot.core.agent.tool import ToolSet
 from astrbot.core.agent.tool_image_cache import tool_image_cache
+from astrbot.core.exceptions import EmptyModelOutputError
 from astrbot.core.message.components import Json
 from astrbot.core.message.message_event_result import (
     MessageChain,
@@ -150,6 +157,32 @@ def _get_persona_custom_error_message(self) -> str | None:
         event = getattr(self.run_context.context, "event", None)
         return extract_persona_custom_error_message_from_event(event)
 
+    async def _complete_with_assistant_response(self, llm_resp: LLMResponse) -> None:
+        """Finalize the current step as a plain assistant response with no tool calls."""
+        self.final_llm_resp = llm_resp
+        self._transition_state(AgentState.DONE)
+        self.stats.end_time = time.time()
+
+        parts = []
+        if llm_resp.reasoning_content or llm_resp.reasoning_signature:
+            parts.append(
+                ThinkPart(
+                    think=llm_resp.reasoning_content,
+                    encrypted=llm_resp.reasoning_signature,
+                )
+            )
+        if llm_resp.completion_text:
+            parts.append(TextPart(text=llm_resp.completion_text))
+        if len(parts) == 0:
+            logger.warning("LLM returned empty assistant message with no tool calls.")
+        self.run_context.messages.append(Message(role="assistant", content=parts))
+
+        try:
+            await self.agent_hooks.on_agent_done(self.run_context, llm_resp)
+        except Exception as e:
+            logger.error(f"Error in on_agent_done hook: {e}", exc_info=True)
+        self._resolve_unconsumed_follow_ups()
+
     @override
     async def reset(
         self,
@@ -305,31 +338,61 @@ async def _iter_llm_responses_with_fallback(
                     candidate_id,
                 )
             self.provider = candidate
-            has_stream_output = False
             try:
-                async for resp in self._iter_llm_responses(include_model=idx == 0):
-                    if resp.is_chunk:
-                        has_stream_output = True
-                        yield resp
-                        continue
-
-                    if (
-                        resp.role == "err"
-                        and not has_stream_output
-                        and (not is_last_candidate)
-                    ):
-                        last_err_response = resp
-                        logger.warning(
-                            "Chat Model %s returns error response, trying fallback to next provider.",
-                            candidate_id,
-                        )
-                        break
-
-                    yield resp
-                    return
+                retrying = AsyncRetrying(
+                    retry=retry_if_exception_type(EmptyModelOutputError),
+                    stop=stop_after_attempt(self.EMPTY_OUTPUT_RETRY_ATTEMPTS),
+                    wait=wait_exponential(
+                        multiplier=1,
+                        min=self.EMPTY_OUTPUT_RETRY_WAIT_MIN_S,
+                        max=self.EMPTY_OUTPUT_RETRY_WAIT_MAX_S,
+                    ),
+                    reraise=True,
+                )
 
-                if has_stream_output:
-                    return
+                async for attempt in retrying:
+                    has_stream_output = False
+                    with attempt:
+                        try:
+                            async for resp in self._iter_llm_responses(
+                                include_model=idx == 0
+                            ):
+                                if resp.is_chunk:
+                                    has_stream_output = True
+                                    yield resp
+                                    continue
+
+                                if (
+                                    resp.role == "err"
+                                    and not has_stream_output
+                                    and (not is_last_candidate)
+                                ):
+                                    last_err_response = resp
+                                    logger.warning(
+                                        "Chat Model %s returns error response, trying fallback to next provider.",
+                                        candidate_id,
+                                    )
+                                    break
+
+                                yield resp
+                                return
+
+                            if has_stream_output:
+                                return
+                        except EmptyModelOutputError:
+                            if has_stream_output:
+                                logger.warning(
+                                    "Chat Model %s returned empty output after streaming started; skipping empty-output retry.",
+                                    candidate_id,
+                                )
+                            else:
+                                logger.warning(
+                                    "Chat Model %s returned empty output on attempt %s/%s.",
+                                    candidate_id,
+                                    attempt.retry_state.attempt_number,
+                                    self.EMPTY_OUTPUT_RETRY_ATTEMPTS,
+                                )
+                            raise
             except Exception as exc:  # noqa: BLE001
                 last_exception = exc
                 logger.warning(
@@ -540,35 +603,7 @@ async def step(self):
             return
 
         if not llm_resp.tools_call_name:
-            # 如果没有工具调用，转换到完成状态
-            self.final_llm_resp = llm_resp
-            self._transition_state(AgentState.DONE)
-            self.stats.end_time = time.time()
-
-            # record the final assistant message
-            parts = []
-
-            if llm_resp.reasoning_content or llm_resp.reasoning_signature:
-                parts.append(
-                    ThinkPart(
-                        think=llm_resp.reasoning_content,
-                        encrypted=llm_resp.reasoning_signature,
-                    )
-                )
-            if llm_resp.completion_text:
-                parts.append(TextPart(text=llm_resp.completion_text))
-            if len(parts) == 0:
-                logger.warning(
-                    "LLM returned empty assistant message with no tool calls."
-                )
-            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:
-                logger.error(f"Error in on_agent_done hook: {e}", exc_info=True)
-            self._resolve_unconsumed_follow_ups()
+            await self._complete_with_assistant_response(llm_resp)
 
         # 返回 LLM 结果
         if llm_resp.result_chain:
@@ -588,6 +623,24 @@ async def step(self):
         if llm_resp.tools_call_name:
             if self.tool_schema_mode == "skills_like":
                 llm_resp, _ = await self._resolve_tool_exec(llm_resp)
+                if not llm_resp.tools_call_name:
+                    logger.warning(
+                        "skills_like tool re-query returned no tool calls; fallback to assistant response."
+                    )
+                    if llm_resp.result_chain:
+                        yield AgentResponse(
+                            type="llm_result",
+                            data=AgentResponseData(chain=llm_resp.result_chain),
+                        )
+                    elif llm_resp.completion_text:
+                        yield AgentResponse(
+                            type="llm_result",
+                            data=AgentResponseData(
+                                chain=MessageChain().message(llm_resp.completion_text),
+                            ),
+                        )
+                    await self._complete_with_assistant_response(llm_resp)
+                    return
 
             tool_call_result_blocks = []
             cached_images = []  # Collect cached images for LLM visibility
@@ -1040,7 +1093,9 @@ def _append_tool_call_result(tool_call_id: str, content: str) -> None:
             )
 
     def _build_tool_requery_context(
-        self, tool_names: list[str]
+        self,
+        tool_names: list[str],
+        extra_instruction: str | None = None,
     ) -> list[dict[str, T.Any]]:
         """Build contexts for re-querying LLM with param-only tool schemas."""
         contexts: list[dict[str, T.Any]] = []
@@ -1052,13 +1107,20 @@ def _build_tool_requery_context(
         instruction = self.SKILLS_LIKE_REQUERY_INSTRUCTION_TEMPLATE.format(
             tool_names=", ".join(tool_names)
         )
+        if extra_instruction:
+            instruction = f"{instruction}\n{extra_instruction}"
         if contexts and contexts[0].get("role") == "system":
             content = contexts[0].get("content") or ""
             contexts[0]["content"] = f"{content}\n{instruction}"
         else:
             contexts.insert(0, {"role": "system", "content": instruction})
         return contexts
 
+    @staticmethod
+    def _has_meaningful_assistant_reply(llm_resp: LLMResponse) -> bool:
+        text = (llm_resp.completion_text or "").strip()
+        return bool(text)
+
     def _build_tool_subset(self, tool_set: ToolSet, tool_names: list[str]) -> ToolSet:
         """Build a subset of tools from the given tool set based on tool names."""
         subset = ToolSet()
@@ -1096,6 +1158,7 @@ async def _resolve_tool_exec(
                     model=self.req.model,
                     session_id=self.req.session_id,
                     extra_user_content_parts=self.req.extra_user_content_parts,
+                    tool_choice="required",
                     abort_signal=self._abort_signal,
                 )
                 if requery_resp:

astrbot/core/astr_agent_tool_exec.py

@@ -1,5 +1,4 @@
 import asyncio
-import datetime
 import inspect
 import json
 import time
@@ -352,6 +351,7 @@ async def _execute_handoff(
 
         # 构建子代理的 system_prompt，添加 skills 提示词和公共上下文
         subagent_system_prompt = tool.agent.instructions or ""
+        subagent_system_prompt = f"# Role\nYour name is {agent_name}(used for tool calling)\n{subagent_system_prompt}\n"
         if agent_name:
             try:
                 from astrbot.core.dynamic_subagent_manager import DynamicSubAgentManager
@@ -362,7 +362,7 @@ async def _execute_handoff(
                     umo, agent_name, runtime
                 )
                 if skills_prompt:
-                    subagent_system_prompt += f"\n\n# Available Skills\n{skills_prompt}"
+                    subagent_system_prompt += f"{skills_prompt}" + "\n"
                     logger.info(f"[SubAgentSkills] Injected skills for {agent_name}")
 
                 # 注入公共上下文
@@ -376,10 +376,18 @@ async def _execute_handoff(
                     )
 
                 # 注入时间信息
-                current_time = (
-                    datetime.datetime.now().astimezone().strftime("%Y-%m-%d %H:%M (%Z)")
+                time_prompt = DynamicSubAgentManager.build_time_prompt(umo)
+                subagent_system_prompt += time_prompt
+
+                # 注入工作目录
+                workdir_prompt = DynamicSubAgentManager.build_workdir_prompt(
+                    umo, agent_name
                 )
-                subagent_system_prompt += f"Current datetime: {current_time}"
+                subagent_system_prompt += workdir_prompt
+
+                # 注入行为规范
+                rule_prompt = DynamicSubAgentManager.build_rule_prompt(umo)
+                subagent_system_prompt += rule_prompt
 
             except Exception:
                 pass

astrbot/core/astr_main_agent.py

@@ -1042,6 +1042,7 @@ def _apply_enhanced_subagent_tools(
         dynamic_subagent_prompt = DynamicSubAgentManager.build_dynamic_subagent_prompt(
             session_id
         )
+
         req.system_prompt = f"{req.system_prompt or ''}\n{dynamic_subagent_prompt}\n"
         # Register existing handoff tools from config
         plugin_context = getattr(event, "_plugin_context", None)

astrbot/core/computer/booters/local.py

@@ -159,10 +159,16 @@ def _run() -> dict[str, Any]:
                     [os.environ.get("PYTHON", sys.executable), "-c", code],
                     timeout=timeout,
                     capture_output=True,
-                    text=True,
+                    # text=True,
+                )
+                # stdout = "" if silent else result.stdout
+                # stderr = result.stderr if result.returncode != 0 else ""
+                stdout = "" if silent else _decode_shell_output(result.stdout)
+                stderr = (
+                    _decode_shell_output(result.stderr)
+                    if result.returncode != 0
+                    else ""
                 )
-                stdout = "" if silent else result.stdout
-                stderr = result.stderr if result.returncode != 0 else ""
                 return {
                     "data": {
                         "output": {"text": stdout, "images": []},

astrbot/core/config/default.py

@@ -195,7 +195,6 @@
         ),
         "agents": [],
     },
-    # 增强版动态SubAgent配置（独立于subagent_orchestrator）
     "enhanced_subagent": {
         "enabled": False,
         "log_level": "debug",
@@ -2537,17 +2536,17 @@ class ChatProviderTemplate(TypedDict):
                     "mimo-tts-style-prompt": {
                         "description": "风格提示词",
                         "type": "string",
-                        "hint": "用于控制生成语音的说话风格、语气或情绪，例如温柔、活泼、沉稳等。可留空。",
+                        "hint": "会以 <style>...</style> 标签形式添加到待合成文本开头，用于控制语速、情绪、角色或风格，例如 开心、变快、孙悟空、悄悄话。可留空。",
                     },
                     "mimo-tts-dialect": {
                         "description": "方言",
                         "type": "string",
-                        "hint": "指定生成语音时使用的方言或口音，例如四川话、粤语口音等。可留空。",
+                        "hint": "会与风格提示词一起写入开头的 <style>...</style> 标签中，例如 东北话、四川话、河南话、粤语。可留空。",
                     },
                     "mimo-tts-seed-text": {
                         "description": "种子文本",
                         "type": "string",
-                        "hint": "用于引导音色和说话方式的参考文本，会影响生成语音的表达风格。",
+                        "hint": "作为可选的 user 消息发送，用于辅助调节语气和风格，不会拼接到待合成文本中。",
                     },
                     "fishaudio-tts-character": {
                         "description": "character",