diff --git a/README.md b/README.md
index f4dc193..bbf6a47 100644
--- a/README.md
+++ b/README.md
@@ -18,7 +18,9 @@
 | 版本比较 | baseline、`scan --compare`、golden CI | 检查发版后失败分布是否退化 |
 | 发布协作 | 可读报告、修复边界、intervention ledger | 支持 review/hold 与后续复盘 |
 
-**当前阶段：** 适合本地或 CI 的低成本回归门禁，不是完整 APM、云 tracing 或自动修复系统；真实团队接入仍需脱敏、权限、时序存储和外部复现验证。
+**当前阶段：** 适合本地或 CI 的低成本回归门禁。已在独立 GitHub 沙箱中复现验收失败，输出
+`acceptance_failed` 并交给评测引擎形成 `hold`；这属于 `external_real_sandbox`，不是生产团队接入。
+项目仍不是完整 APM、云 tracing 或自动修复系统；真实团队接入仍需脱敏、权限和时序存储。
 
 ---
 
@@ -100,7 +102,10 @@ python -m pytest tests/test_failure_golden.py   # CI 同款
 | 干预 ledger（Learning Capture） | [docs/intervention_ledger.json](docs/intervention_ledger.json) |
 | 业务证明自评 ~65% | [docs/VALUE.md](docs/VALUE.md) |
 
-仍缺：真人秒表、非模拟 PR hold、外部团队复现。
+外部沙箱证据：[`agent-delivery-sandbox`](https://github.com/weihuaguo270-ops/agent-delivery-sandbox)
+已完成非模拟 PR 的接受、拒绝、回滚，以及 `acceptance_failed -> hold` 故障回流。
+
+仍缺：真人执行耗时基线、生产团队接入复现、长期时序数据和告警闭环。
 
 Golden CI：[docs/golden_evidence_baseline.md](docs/golden_evidence_baseline.md)
 
diff --git a/pyproject.toml b/pyproject.toml
index 1fdcbc6..e374b04 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -5,7 +5,7 @@ build-backend = "setuptools.build_meta"
 [project]
 name = "trace-debugger"
 version = "0.5.0"
-description = "Agent 回归测试与失败治理门禁 — Format B 轨迹、scan/compare baseline、7 类启发式检测、golden CI"
+description = "Agent 回归测试与失败治理门禁 — Format B 轨迹、scan/compare baseline、8 类启发式检测、golden CI"
 readme = "README.md"
 requires-python = ">=3.10"
 
diff --git a/trace_debugger/analyzer.py b/trace_debugger/analyzer.py
index cd16b2f..0499a3f 100644
--- a/trace_debugger/analyzer.py
+++ b/trace_debugger/analyzer.py
@@ -86,6 +86,7 @@ def content_tokens(text: str) -> set[str]:
 
 
 def looks_like_overflow_text(text: str) -> bool:
+    """按已知中英文服务错误文案识别上下文溢出信号。"""
     if not text:
         return False
     low = text.lower()
@@ -93,6 +94,7 @@ def looks_like_overflow_text(text: str) -> bool:
 
 
 def is_search_tool(name: str, *, substrings: tuple[str, ...] = ("search",), extra_names: tuple[str, ...] = ()) -> bool:
+    """按可配置名称规则判断工具是否属于搜索类。"""
     n = (name or "").lower()
     if any(sub in n for sub in substrings):
         return True
@@ -102,6 +104,7 @@ def is_search_tool(name: str, *, substrings: tuple[str, ...] = ("search",), extr
 
 
 def is_final_thought(thought: str, markers: tuple[str, ...]) -> bool:
+    """按调用方提供的大小写不敏感标记识别终答思考。"""
     upper = (thought or "").upper()
     return any(m.upper() in upper for m in markers if m)
 
@@ -195,9 +198,11 @@ def __init__(
         self.search_tool_names = search_tool_names
 
     def step_is_final(self, step: Step) -> bool:
+        """按当前分析器终答标记判断步骤。"""
         return is_final_thought(step.thought, self.final_answer_markers)
 
     def tool_is_search(self, name: str) -> bool:
+        """按当前分析器搜索工具配置判断名称。"""
         return is_search_tool(
             name,
             substrings=self.search_tool_substrings,
diff --git a/trace_debugger/episode.py b/trace_debugger/episode.py
index 3187f30..b0d508d 100644
--- a/trace_debugger/episode.py
+++ b/trace_debugger/episode.py
@@ -14,6 +14,8 @@
 
 @dataclass(frozen=True)
 class ImportedEpisode:
+    """从跨框架 EvaluationEpisode 提取的只读调试证据。"""
+
     episode_id: str
     framework: str
     agent_version: str
diff --git a/trace_debugger/golden.py b/trace_debugger/golden.py
index 61486c2..cc641f2 100644
--- a/trace_debugger/golden.py
+++ b/trace_debugger/golden.py
@@ -14,6 +14,8 @@
 
 @dataclass
 class GoldenCase:
+    """一个轨迹夹具及其必须命中、不得命中的失败断言。"""
+
     id: str
     file: str
     split: str
@@ -25,6 +27,7 @@ class GoldenCase:
 
     @classmethod
     def from_dict(cls, data: dict[str, Any]) -> "GoldenCase":
+        """从 manifest 条目构建用例，缺省值保持旧清单兼容。"""
         return cls(
             id=data["id"],
             file=data["file"],
@@ -39,12 +42,15 @@ def from_dict(cls, data: dict[str, Any]) -> "GoldenCase":
 
 @dataclass
 class GoldenManifest:
+    """带 schema 版本的失败回归集清单。"""
+
     schema_version: str
     description: str
     cases: list[GoldenCase]
 
     @classmethod
     def load(cls, path: Optional[Path] = None) -> "GoldenManifest":
+        """从指定路径或内置夹具目录加载清单。"""
         root = path or (DEFAULT_GOLDEN_DIR / "manifest.json")
         with open(root, encoding="utf-8") as f:
             data = json.load(f)
@@ -56,11 +62,13 @@ def load(cls, path: Optional[Path] = None) -> "GoldenManifest":
 
 
 def load_manifest(manifest_path: Optional[str] = None) -> GoldenManifest:
+    """加载默认或显式指定的 Golden 清单。"""
     path = Path(manifest_path) if manifest_path else DEFAULT_GOLDEN_DIR / "manifest.json"
     return GoldenManifest.load(path)
 
 
 def analyze_case(case: GoldenCase, *, golden_dir: Optional[Path] = None) -> TrajectoryAnalysis:
+    """加载一个 Golden 轨迹并运行默认分析器。"""
     base = golden_dir or DEFAULT_GOLDEN_DIR
     return Analyzer().analyze(load(str(base / case.file)))
 
diff --git a/trace_debugger/harness.py b/trace_debugger/harness.py
index 1ada249..e2f166a 100644
--- a/trace_debugger/harness.py
+++ b/trace_debugger/harness.py
@@ -51,6 +51,7 @@ class StepEvent:
     error_message: str = ""
 
     def tool_args_str(self) -> str:
+        """将工具参数序列化为 Format B 使用的字符串。"""
         return normalize_tool_input(self.tool_input)
 
 
@@ -210,6 +211,7 @@ def finish(
         total_duration: float = 0.0,
         metadata: Optional[dict[str, Any]] = None,
     ) -> TrajectoryAnalysis:
+        """结束运行，记录路径级失败并返回完整轨迹分析。"""
         return self._watcher.on_finish(
             final_answer=final_answer,
             total_duration=total_duration,
@@ -217,10 +219,12 @@ def finish(
         )
 
     def trajectory_dict(self) -> dict[str, Any]:
+        """返回当前运行的 Format B 轨迹快照。"""
         data = self._watcher.to_trajectory_dict()
         data.setdefault("timestamp", self.context.timestamp)
         return data
 
     @property
     def record_path(self) -> str:
+        """返回 FailureHarness 实际使用的失败事件文件。"""
         return self._watcher.record_path
diff --git a/trace_debugger/reader.py b/trace_debugger/reader.py
index 17fd5fa..90a471a 100644
--- a/trace_debugger/reader.py
+++ b/trace_debugger/reader.py
@@ -24,18 +24,22 @@ class Step:
 
     @property
     def is_action(self) -> bool:
+        """返回该步是否包含工具动作。"""
         return bool(self.action_name)
 
     @property
     def is_final(self) -> bool:
+        """按兼容格式中的 FINAL ANSWER 标记识别终答步骤。"""
         return "FINAL ANSWER" in self.thought.upper()
 
     @property
     def is_thought(self) -> bool:
+        """返回该步是否为非终答的有效思考。"""
         return bool(self.thought.strip()) and not self.is_final
 
     @property
     def summary(self) -> str:
+        """生成用于终端和报告的截断单行摘要。"""
         if self.is_final:
             return f"输出答案: {self.thought[:80]}"
         if self.is_action:
@@ -59,18 +63,22 @@ class Path:
 
     @property
     def num_steps(self) -> int:
+        """返回路径中的步骤数。"""
         return len(self.steps)
 
     @property
     def tools_used(self) -> list[str]:
+        """按调用顺序返回工具名，保留重复调用。"""
         return [s.action_name for s in self.steps if s.action_name]
 
     @property
     def has_errors(self) -> bool:
+        """返回路径是否含解析器识别出的工具错误。"""
         return any(s.has_error for s in self.steps)
 
     @property
     def error_summary(self) -> list[str]:
+        """返回路径内各错误步骤的截断消息。"""
         return [s.error_message for s in self.steps if s.has_error]
 
 
@@ -89,14 +97,17 @@ class Trajectory:
 
     @property
     def num_steps(self) -> int:
+        """返回顶层轨迹步骤数。"""
         return len(self.steps)
 
     @property
     def num_paths(self) -> int:
+        """返回解析后的执行路径数。"""
         return len(self.paths)
 
     @property
     def main_path(self) -> Optional[Path]:
+        """返回显式主路径；旧数据未标记时回退到最后一条。"""
         for p in self.paths:
             if p.is_main_path:
                 return p
@@ -104,6 +115,7 @@ def main_path(self) -> Optional[Path]:
 
     @property
     def failed_paths(self) -> list[Path]:
+        """返回失败的非主路径，避免把最终输出路径重复计为失败分支。"""
         return [p for p in self.paths if not p.success and not p.is_main_path]
 
 
diff --git a/trace_debugger/record.py b/trace_debugger/record.py
index d850d38..0b5597b 100644
--- a/trace_debugger/record.py
+++ b/trace_debugger/record.py
@@ -50,10 +50,12 @@ def _preview(text: str, limit: int = 160) -> str:
 
 
 def failure_label(failure_type: str) -> str:
+    """返回稳定失败类型对应的展示标签。"""
     return FailureType.LABELS.get(failure_type, failure_type)
 
 
 def failure_severity(failure_type: str, *, event_type: str = "step_failure") -> str:
+    """将失败类型映射为报告严重级别，不改变分类结果。"""
     if failure_type in (FailureType.NO_FINAL_ANSWER, FailureType.LLM_OFFTRACK):
         return "fail"
     if failure_type in (FailureType.CONTEXT_OVERFLOW, FailureType.SEARCH_TIMEOUT):
@@ -70,6 +72,7 @@ def build_failure_summary(
     action: str = "",
     event_type: str = "step_failure",
 ) -> str:
+    """生成包含失败类型、步骤和动作的单行摘要。"""
     label = failure_label(failure_type)
     if event_type == "path_failure":
         return f"路径级 · {label}"
@@ -88,6 +91,7 @@ def build_failure_context(
     observation: str = "",
     duration: float = 0.0,
 ) -> dict[str, Any]:
+    """提取用于定位失败的最小轨迹上下文。"""
     ctx: dict[str, Any] = {}
     if action:
         ctx["action"] = action
@@ -163,11 +167,13 @@ def format_event_readable(ev: dict[str, Any]) -> str:
 
 
 def readable_log_path(record_path: str) -> Path:
+    """返回机器可读 JSONL 对应的人类可读日志路径。"""
     p = Path(record_path)
     return p.with_name(p.stem + ".log")
 
 
 def session_summary_path(record_path: str, session_id: str) -> Path:
+    """返回指定会话的独立摘要文件路径。"""
     base = Path(record_path).parent / "sessions"
     safe = "".join(c if c.isalnum() or c in "-_" else "_" for c in session_id)
     return base / f"{safe}.md"
@@ -249,6 +255,7 @@ def write_session_summary(
 
 
 def load_failure_events(record_path: str, *, session_id: Optional[str] = None) -> list[dict[str, Any]]:
+    """按写入顺序读取有效事件，并可限定到单个会话。"""
     path = Path(record_path)
     if not path.is_file():
         return []
@@ -606,6 +613,7 @@ def build_scan_snapshot(
 
 
 def load_snapshot(path: str) -> dict[str, Any]:
+    """加载一次扫描快照，供回归对比使用。"""
     with open(path, encoding="utf-8") as f:
         return json.load(f)
 
diff --git a/trace_debugger/validate.py b/trace_debugger/validate.py
index 75845fe..e0c9b05 100644
--- a/trace_debugger/validate.py
+++ b/trace_debugger/validate.py
@@ -48,6 +48,7 @@ def validate_trajectory_file(
     use_schema: bool = False,
     schema_path: Optional[Path] = None,
 ) -> list[str]:
+    """加载并校验一个轨迹文件，读取或格式错误均作为消息返回。"""
     p = Path(path)
     if not p.is_file():
         return [f"file not found: {path}"]
@@ -62,6 +63,7 @@ def validate_trajectory_file(
 
 
 def format_validation_report(errors: list[str], *, path: str = "") -> str:
+    """将校验错误格式化为稳定的终端报告。"""
     if not errors:
         prefix = f"{path}: " if path else ""
         return f"{prefix}OK — Format B validation passed"