feat: add detailed attempts list to trajectory records

Each retry attempt now records llm_prompt_tail, llm_raw_output,
llm_caller, patches_count, patch_apply_ok/error, modified_lines,
tests_passed/total, test_output_tail, and error_type. This enables
precise post-hoc diagnosis without re-running benchmarks.

Also fixes test_audit_model.py to use correct module-level constants
(LOGS_SUCCESS/LOGS_FAILED/LOGS_LEGACY instead of removed LOGS_DIR).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

kaiwu/core/orchestrator.py

@@ -241,6 +241,8 @@ class PipelineOrchestrator:
 
         # 错误类型追踪
         ctx._errors_encountered = []
+        # 详细attempt记录（供trajectory_collector使用）
+        ctx._trajectory_attempts = []
 
         # 每次顶层任务重置manifest
         self._manifest.clear()
@@ -459,6 +461,13 @@ class PipelineOrchestrator:
             }
             ctx.attempt_history.append(attempt_record)
 
+            # ── Trajectory详细attempt记录（诊断用） ──
+            try:
+                _traj_attempt = self._build_trajectory_attempt(ctx, current_error_type)
+                ctx._trajectory_attempts.append(_traj_attempt)
+            except Exception:
+                pass  # 记录失败不影响主流程
+
             # 快速熔断：语法错误重试无效
             # syntax熔断按tier区分：SMALL立刻熔断，MEDIUM/LARGE多给一次
             _syntax_max = 1 if self._model_tier == ModelTier.SMALL else 2
@@ -1235,3 +1244,83 @@ class PipelineOrchestrator:
             )
         except Exception:
             return ctx.gap or Gap(GapType.UNKNOWN, 0.3, [], [], "", "")
+
+    def _build_trajectory_attempt(self, ctx: TaskContext, error_type: str) -> dict:
+        """构建单次attempt的详细诊断记录（供trajectory.json使用）。
+
+        所有字段try-except保护，单字段失败不影响其他字段。
+        字符串字段截断到合理长度（≤500字）。
+        """
+        attempt = {"attempt": ctx.retry_count - 1}  # retry_count已+1，这里记录的是刚完成的attempt
+
+        # LLM输入输出（从audit logger的llm_calls取最近一条）
+        try:
+            llm_calls = self._audit._llm_calls
+            if llm_calls:
+                last_call = llm_calls[-1]
+                attempt["llm_prompt_tail"] = (last_call.get("prompt_preview", "") or "")[-400:]
+                attempt["llm_raw_output"] = (last_call.get("raw_output", "") or "")[:400]
+                attempt["llm_caller"] = last_call.get("caller", "")
+            else:
+                attempt["llm_prompt_tail"] = ""
+                attempt["llm_raw_output"] = ""
+                attempt["llm_caller"] = ""
+        except Exception:
+            attempt["llm_prompt_tail"] = ""
+            attempt["llm_raw_output"] = ""
+            attempt["llm_caller"] = ""
+
+        # patch生成结果
+        try:
+            patches = (ctx.generator_output or {}).get("patches", [])
+            attempt["patches_count"] = len(patches)
+        except Exception:
+            attempt["patches_count"] = 0
+
+        # patch apply结果
+        try:
+            v = ctx.verifier_output or {}
+            if v.get("error_type") == "patch_apply":
+                attempt["patch_apply_ok"] = False
+                attempt["patch_apply_error"] = (v.get("error_message", "") or v.get("error_detail", ""))[:300]
+            elif v:
+                # verifier跑到了测试阶段，说明apply成功了
+                attempt["patch_apply_ok"] = True
+                attempt["patch_apply_error"] = ""
+            else:
+                attempt["patch_apply_ok"] = False
+                attempt["patch_apply_error"] = "no verifier output"
+        except Exception:
+            attempt["patch_apply_ok"] = False
+            attempt["patch_apply_error"] = ""
+
+        # 修改行数
+        try:
+            patches = (ctx.generator_output or {}).get("patches", [])
+            modified_lines = 0
+            for p in patches:
+                orig = p.get("original", "") or ""
+                mod = p.get("modified", "") or p.get("content", "") or ""
+                modified_lines += abs(len(mod.splitlines()) - len(orig.splitlines()))
+            attempt["modified_lines"] = modified_lines
+        except Exception:
+            attempt["modified_lines"] = 0
+
+        # verifier测试结果
+        try:
+            v = ctx.verifier_output or {}
+            attempt["tests_passed"] = v.get("tests_passed", 0)
+            attempt["tests_total"] = v.get("tests_total", 0)
+            attempt["test_output_tail"] = (v.get("error_detail", "") or "")[-300:]
+        except Exception:
+            attempt["tests_passed"] = 0
+            attempt["tests_total"] = 0
+            attempt["test_output_tail"] = ""
+
+        # 错误类型
+        try:
+            attempt["error_type"] = error_type or ""
+        except Exception:
+            attempt["error_type"] = ""
+
+        return attempt

kaiwu/flywheel/trajectory_collector.py

@@ -35,6 +35,8 @@ class TaskTrajectory:
     timestamp: str = ""
     search_triggered: bool = False
     project_hash: str = ""
+    # 每次attempt的详细记录（诊断用）
+    attempts: list[dict] = field(default_factory=list)
 
 
 class TrajectoryCollector:
@@ -66,6 +68,9 @@ class TrajectoryCollector:
         if ctx.generator_output and "patches" in ctx.generator_output:
             files_modified = [p.get("file", "") for p in ctx.generator_output["patches"] if p.get("file")]
 
+        # 收集详细attempts记录（从ctx._trajectory_attempts，orchestrator填充）
+        attempts = getattr(ctx, '_trajectory_attempts', [])
+
         traj = TaskTrajectory(
             task_id=str(uuid.uuid4()),
             user_input=ctx.user_input,
@@ -85,6 +90,7 @@ class TrajectoryCollector:
             timestamp=datetime.now(timezone.utc).isoformat(),
             search_triggered=ctx.search_triggered,
             project_hash=hashlib.sha256(ctx.project_root.encode()).hexdigest()[:16],
+            attempts=attempts,
         )
 
         path = os.path.join(self._dir, f"{traj.task_id}.json")

tests/test_audit_model.py

@@ -21,29 +21,34 @@ class TestAuditLogger(unittest.TestCase):
     def tearDown(self):
         shutil.rmtree(self.tmpdir)
 
-    @patch("kaiwu.audit.logger.LOGS_DIR")
-    def test_write_creates_log(self, mock_dir):
-        mock_dir.__class__ = type(self.logs_dir)
+    def test_write_creates_log(self):
         from kaiwu.audit.logger import AuditLogger
-        with patch("kaiwu.audit.logger.LOGS_DIR", self.logs_dir):
+        success_dir = self.logs_dir / "success"
+        failed_dir = self.logs_dir / "failed"
+        with patch("kaiwu.audit.logger.LOGS_SUCCESS", success_dir), \
+             patch("kaiwu.audit.logger.LOGS_FAILED", failed_dir):
             logger = AuditLogger()
             logger.start()
             logger.log("gate", "locator_repair | 难度：easy")
             logger.log("locator", "读取 test.py")
 
-            # Mock context
+            # Mock context — need real values for fields accessed by write()
             ctx = MagicMock()
             ctx.user_input = "修复login函数"
             ctx.gate_result = {"expert_type": "locator_repair", "difficulty": "easy"}
             ctx.generator_output = {"patches": [{"file": "test.py", "original": "old", "modified": "new"}]}
-            ctx.verifier_output = {"tests_passed": 3, "tests_total": 3}
+            ctx.verifier_output = {"tests_passed": 3, "tests_total": 3, "passed": True, "error_type": "", "structured_failures": []}
             ctx.retry_count = 0
             ctx.search_triggered = False
+            ctx.gap = None
+            ctx.locator_output = None
+            ctx.routing_source = ""
+            ctx.attempt_history = []
 
             logger.write(ctx, 5.2, True, "qwen3:8b")
 
             # Verify log file created
-            logs = list(self.logs_dir.glob("*.json"))
+            logs = list(success_dir.glob("*.json"))
             assert len(logs) == 1
 
             data = json.loads(logs[0].read_text(encoding="utf-8"))
@@ -53,46 +58,56 @@ class TestAuditLogger(unittest.TestCase):
             assert len(data["events"]) == 2
             assert data["files_modified"] == ["test.py"]
 
-    @patch("kaiwu.audit.logger.LOGS_DIR")
-    def test_list_logs(self, mock_dir):
+    def test_list_logs(self):
         from kaiwu.audit.logger import list_logs
-        with patch("kaiwu.audit.logger.LOGS_DIR", self.logs_dir):
-            self.logs_dir.mkdir(parents=True, exist_ok=True)
+        success_dir = self.logs_dir / "success"
+        failed_dir = self.logs_dir / "failed"
+        legacy_dir = self.logs_dir / "legacy"
+        with patch("kaiwu.audit.logger.LOGS_SUCCESS", success_dir), \
+             patch("kaiwu.audit.logger.LOGS_FAILED", failed_dir), \
+             patch("kaiwu.audit.logger.LOGS_LEGACY", legacy_dir):
+            success_dir.mkdir(parents=True, exist_ok=True)
 
             # Write 3 log files with distinct names
             for i in range(3):
                 record = {"task": f"task {i}", "success": True, "elapsed_s": 1.0,
                            "timestamp": f"2026-05-06T10:00:0{i}", "model": "test"}
-                (self.logs_dir / f"2026-05-06_10000{i}_codegen.json").write_text(
+                (success_dir / f"2026-05-06_10000{i}_codegen.json").write_text(
                     json.dumps(record), encoding="utf-8"
                 )
 
             logs = list_logs(limit=10)
             assert len(logs) == 3
 
-    @patch("kaiwu.audit.logger.LOGS_DIR")
-    def test_clear_logs(self, mock_dir):
+    def test_clear_logs(self):
         from kaiwu.audit.logger import AuditLogger, clear_logs
-        with patch("kaiwu.audit.logger.LOGS_DIR", self.logs_dir):
-            self.logs_dir.mkdir(parents=True, exist_ok=True)
-            (self.logs_dir / "test.json").write_text("{}", encoding="utf-8")
+        success_dir = self.logs_dir / "success"
+        failed_dir = self.logs_dir / "failed"
+        legacy_dir = self.logs_dir / "legacy"
+        with patch("kaiwu.audit.logger.LOGS_SUCCESS", success_dir), \
+             patch("kaiwu.audit.logger.LOGS_FAILED", failed_dir), \
+             patch("kaiwu.audit.logger.LOGS_LEGACY", legacy_dir):
+            success_dir.mkdir(parents=True, exist_ok=True)
+            (success_dir / "test.json").write_text("{}", encoding="utf-8")
             count = clear_logs()
             assert count == 1
-            assert len(list(self.logs_dir.glob("*.json"))) == 0
+            assert len(list(success_dir.glob("*.json"))) == 0
 
-    @patch("kaiwu.audit.logger.LOGS_DIR")
-    def test_max_logs_cleanup(self, mock_dir):
+    def test_max_logs_cleanup(self):
         from kaiwu.audit.logger import AuditLogger, MAX_LOGS
-        with patch("kaiwu.audit.logger.LOGS_DIR", self.logs_dir):
-            self.logs_dir.mkdir(parents=True, exist_ok=True)
+        success_dir = self.logs_dir / "success"
+        with patch("kaiwu.audit.logger.LOGS_SUCCESS", success_dir), \
+             patch("kaiwu.audit.logger.LOGS_FAILED", self.logs_dir / "failed"), \
+             patch("kaiwu.audit.logger.LOGS_LEGACY", self.logs_dir / "legacy"):
+            success_dir.mkdir(parents=True, exist_ok=True)
             # Create MAX_LOGS + 5 files
             for i in range(MAX_LOGS + 5):
-                (self.logs_dir / f"2026-01-01_{i:06d}_test.json").write_text("{}", encoding="utf-8")
+                (success_dir / f"2026-01-01_{i:06d}_test.json").write_text("{}", encoding="utf-8")
 
             al = AuditLogger()
-            al._cleanup()
+            al._cleanup(success_dir)
 
-            remaining = list(self.logs_dir.glob("*.json"))
+            remaining = list(success_dir.glob("*.json"))
             assert len(remaining) == MAX_LOGS