fix: give final acceptance a shared retry budget and diagnostics

artemis/agents/checker/checker.py

@@ -51,6 +51,7 @@ from artemis.context import ArtemisContext
 from artemis.data_engine.context_vars import CURRENT_TRACE_ID
 from artemis.data_engine.trace import trace, trace_langchain_tool
 from artemis.llm.parser import extract_thinking_tags
+from artemis.llm.final_review import final_review_stage
 from artemis.llm.structured import content_to_text
 from artemis.services.llm import acomplete, get_llm, invoke_llm_with_timeout_message
 from artemis.tools.history import get_history_tools
@@ -499,7 +500,9 @@ def _completed_text_report(response, check_items) -> CheckReport | None:
 
 async def _structured_report(llm, messages, check_items=()) -> CheckReport:
     if messages and (report := _completed_text_report(messages[-1], check_items)) is not None:
+        final_review_stage("validating_verdict")
         return report
+    final_review_stage("structuring_verdict")
     structured_llm = llm.with_structured_output(CheckReport)
     # A conversation must end with a user turn: Gemini rejects requests whose
     # last message is the model's own answer ("Requests ending with a model
@@ -575,6 +578,7 @@ async def _run_check_loop(
 
     report: CheckReport | None = None
     for i in range(max_iterations):
+        final_review_stage("reviewing_evidence", iteration=i + 1)
         if i == max_iterations - 1:
             messages.append(
                 HumanMessage(
@@ -594,6 +598,7 @@ async def _run_check_loop(
             break
 
         messages.append(response)
+        final_review_stage("reading_evidence")
         for tc in response.tool_calls:
             tool_name = tc["name"].split(":")[-1] if ":" in tc["name"] else tc["name"]
             args = dict(tc["args"])

artemis/agents/checker/flash.py

@@ -1,10 +1,10 @@
 """Flash's exit-only acceptance check uses the same read-only Checker as Pro."""
 
-import asyncio
 from pathlib import Path
 
 from artemis.agents.checker.checker import run_final_check
 from artemis.context import ArtemisContext
+from artemis.llm.diagnostics import final_review_error_text
 from artemis.graph.checkpoints import (
     RunOutcome,
     TestSummary,
@@ -14,6 +14,7 @@ from artemis.graph.checkpoints import (
     write_run_outcome,
 )
 from artemis.runtime.case_result import CaseTestResult, resolve_case_result
+from artemis.services.llm import final_review_budget
 from artemis.utils.logger import get_logger
 from artemis.utils.notes import save_verified_note
 
@@ -50,7 +51,6 @@ async def settle_flash_outcome(
                 logger.warning(f"Flash 最终答复保存失败：{exc}")
                 delivery = {"ok": False, "verified": False, "error": str(exc)}
         if final_check_enabled(ctx):
-            timeout = float(getattr(ctx.execution_setup, "checkpoint_timeout", 180.0))
             publish_checker_event(
                 ctx,
                 {
@@ -63,8 +63,8 @@ async def settle_flash_outcome(
                 },
             )
             try:
-                report = await asyncio.wait_for(
-                    run_final_check(
+                report = await final_review_budget(ctx, attempt_id="final#1").run(
+                    lambda: run_final_check(
                         ctx,
                         goal=goal,
                         plan_text="",
@@ -74,7 +74,6 @@ async def settle_flash_outcome(
                         execution_report=execution_report,
                         report_delivery=delivery,
                     ),
-                    timeout=timeout,
                 )
                 result = report.test_result
                 if report.report_corrections and directory:
@@ -101,11 +100,7 @@ async def settle_flash_outcome(
                     },
                 )
             except Exception as exc:
-                detail = (
-                    f"超过 {timeout:g} 秒"
-                    if isinstance(exc, TimeoutError)
-                    else f"{type(exc).__name__}: {exc}"
-                )
+                detail = final_review_error_text(exc)
                 logger.warning(f"Flash 测试验收未完成：{detail}")
                 result = CaseTestResult(
                     status="inconclusive", reason=f"操作已结束，但测试验收未完成：{detail}。"