fix: give final acceptance a shared retry budget and diagnostics

artemis/services/llm.py

@@ -47,6 +47,10 @@ from artemis.config import (
 from artemis.context import ArtemisContext
 from artemis.data_engine.trace import CURRENT_TRACE_ID, DataEngineCallbackHandler
 from artemis.llm.diagnostics import ACTIVE_ATTEMPT, RequestAttempt, error_text, input_images
+from artemis.llm.final_review import (
+    ACTIVE_FINAL_REVIEW,
+    FinalReviewBudget,
+)
 from artemis.llm.cache_affinity import cache_affinity_options
 from artemis.llm.reliability import (
     CircuitBreaker,
@@ -353,8 +357,9 @@ async def _run_with_recovery[T](
       immediately: retrying cannot help and pausing would hang the task.
     - When retryable attempts are exhausted: if a fallback model is waiting
       (with_fallback), raise LLMExhaustedError so it takes over immediately;
-      otherwise pause the task (bounded by settings.LLM_PAUSE_TIMEOUT_SECONDS)
-      and retry from scratch on resume.
+      final acceptance also raises instead of pausing an already-finished task.
+      Other calls pause (bounded by settings.LLM_PAUSE_TIMEOUT_SECONDS) and
+      retry the same request on resume.
     """
     while True:
         request_token = _begin_llm_request(provider)
@@ -394,7 +399,7 @@ async def _run_with_recovery[T](
                     async with asyncio.timeout(request_timeout or settings.OPENAI_TIMEOUT_SECONDS):
                         result = await call_once()
                 except (KeyboardInterrupt, asyncio.CancelledError) as e:
-                    observation.finish("cancelled", error=e, category="cancelled")
+                    observation.cancel(e)
                     raise
                 except LLMCallError as e:
                     observation.finish("failed", error=e, category=e.failure.category.value)
@@ -458,7 +463,7 @@ async def _run_with_recovery[T](
                     except asyncio.CancelledError:
                         # Guidance may cancel a retry wait while the task continues.
                         # Close this attempt so it cannot look active next to the new request.
-                        observation.finish("cancelled")
+                        observation.cancel()
                         raise
                 else:
                     observation.finish("completed", response=result)
@@ -469,7 +474,7 @@ async def _run_with_recovery[T](
                     ACTIVE_ATTEMPT.reset(observation_token)
 
             # Retryable attempts exhausted.
-            if _FALLBACK_AVAILABLE.get():
+            if _FALLBACK_AVAILABLE.get() or ACTIVE_FINAL_REVIEW.get() is not None:
                 if failed_endpoints is not None and endpoint_key:
                     failed_endpoints[endpoint_key] = last_error
                 _record_llm_event(
@@ -478,7 +483,7 @@ async def _run_with_recovery[T](
                         "error": str(last_error)[:1000],
                         "category": last_failure.category.value,
                         "retryable": True,
-                        "handover": "fallback",
+                        "handover": "fallback" if _FALLBACK_AVAILABLE.get() else "final_review_end",
                     },
                     status="failed",
                 )
@@ -799,6 +804,16 @@ class RobustChatModelWrapper:
                     if emit_deltas:
                         self._emit_stream_delta(stream_exec_id, chunk)
         except (KeyboardInterrupt, asyncio.CancelledError):
+            if full_response is not None:
+                _record_llm_event(
+                    "llm_stream_reset",
+                    {
+                        "stream_exec_id": str(stream_exec_id),
+                        "action": "discard",
+                        "reason": "incomplete_response",
+                        "message": "本次输出未完成，已丢弃半截内容。",
+                    },
+                )
             raise
         except Exception as stream_error:
             if full_response is not None or isinstance(stream_error, ToolArgumentOverflow):
@@ -1085,6 +1100,20 @@ def _resolve_endpoint(
     )
 
 
+def final_review_budget(ctx: ArtemisContext, *, attempt_id: str) -> FinalReviewBudget:
+    """Use the same effective Checker endpoint as get_llm, without creating a model."""
+    endpoint = _resolve_endpoint(ctx, "checker")
+    return FinalReviewBudget(
+        ctx.data_engine,
+        attempt_id=attempt_id,
+        request_timeout=endpoint.timeout_seconds,
+        max_retries=endpoint.max_retries,
+        retry_delay=max(
+            settings.OPENAI_RETRY_INTERVAL_SECONDS, settings.OPENAI_RATE_LIMIT_RETRY_SECONDS
+        ),
+    )
+
+
 @overload
 def get_llm(
     ctx: ArtemisContext,