fix: prevent late LLM responses after stop

astrbot/builtin_stars/builtin_commands/commands/conversation.py

@@ -11,6 +11,7 @@
 )
 from astrbot.core.agent.runners.deerflow.deerflow_api_client import DeerFlowAPIClient
 from astrbot.core.db.po import ProviderStat
+from astrbot.core.pipeline.process_stage.follow_up import request_active_runner_stop
 from astrbot.core.utils.active_event_registry import active_event_registry
 
 from .utils.rst_scene import RstScene
@@ -206,6 +207,8 @@ async def stop(self, message: AstrMessageEvent) -> None:
                 umo,
                 exclude=message,
             )
+            if request_active_runner_stop(umo) and stopped_count == 0:
+                stopped_count = 1
 
         if stopped_count > 0:
             message.set_result(

astrbot/core/agent/stop_policy.py

@@ -0,0 +1,31 @@
+from typing import Any
+
+AGENT_OUTPUT_DELIVERY_CONFIRMED_KEY = "_agent_output_delivery_confirmed"
+
+
+class AgentOutputStopped(Exception):
+    """Unwind streaming adapters without triggering broad error handlers."""
+
+
+def event_requests_agent_stop(event: Any) -> bool:
+    """Return whether an event requests the current agent run to stop.
+
+    Args:
+        event: Event-like object that may expose stop state and extras.
+
+    Returns:
+        Whether hard stop, soft stop, or user-aborted state is active.
+    """
+    if event is None:
+        return False
+
+    is_stopped = getattr(event, "is_stopped", None)
+    if callable(is_stopped) and is_stopped():
+        return True
+
+    get_extra = getattr(event, "get_extra", None)
+    if not callable(get_extra):
+        return False
+    return bool(get_extra("agent_stop_requested")) or bool(
+        get_extra("agent_user_aborted")
+    )

astrbot/core/astr_agent_hooks.py

@@ -4,9 +4,10 @@
 
 from astrbot.core.agent.hooks import BaseAgentRunHooks
 from astrbot.core.agent.run_context import ContextWrapper
+from astrbot.core.agent.stop_policy import event_requests_agent_stop
 from astrbot.core.agent.tool import FunctionTool
 from astrbot.core.astr_agent_context import AstrAgentContext
-from astrbot.core.pipeline.context_utils import call_event_hook
+from astrbot.core.pipeline.context_utils import call_agent_done_hook, call_event_hook
 from astrbot.core.star.star_handler import EventType
 
 
@@ -22,20 +23,29 @@ async def on_agent_begin(
 
     async def on_agent_done(self, run_context, llm_response) -> None:
         # 执行事件钩子
-        if llm_response and llm_response.reasoning_content:
+        event = run_context.context.event
+        suppress_llm_response = event_requests_agent_stop(event)
+
+        if (
+            not suppress_llm_response
+            and 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
+            event.set_extra("_llm_reasoning_content", llm_response.reasoning_content)
+
+        if not suppress_llm_response:
+            await call_event_hook(
+                event,
+                EventType.OnLLMResponseEvent,
+                llm_response,
             )
 
-        await call_event_hook(
-            run_context.context.event,
-            EventType.OnLLMResponseEvent,
-            llm_response,
-        )
-        await call_event_hook(
-            run_context.context.event,
-            EventType.OnAgentDoneEvent,
+        if event_requests_agent_stop(event):
+            event.set_extra("_llm_reasoning_content", None)
+
+        await call_agent_done_hook(
+            event,
             run_context,
             llm_response,
         )