chore(v0.9): wrap over-length lines and record spec rulings

Wrap the judge prompt and two docstrings past the 120-col convention;
record three implementation rulings in the v0.9 spec (exploration read
outlets, round-based sampling, findings carrying all ratings).

.scratch/v0.9/issues/04-judge-review.md

@@ -4,9 +4,9 @@
 
 **Blocked by:** 01（会话关闭事件与对话数据）
 
-**Status:** ready-for-agent
+**Status:** completed
 
-- [ ] 会话关闭后自动触发抽样复核，结果结构化落库
-- [ ] 抽样上限生效；无模型配置时静默跳过
-- [ ] 复核失败落错误，会话仍为 completed
-- [ ] 假客户端测试覆盖编排、解析失败、无模型分支（先例：v0.7 分析测试）
+- [x] 会话关闭后自动触发抽样复核，结果结构化落库
+- [x] 抽样上限生效；无模型配置时静默跳过
+- [x] 复核失败落错误，会话仍为 completed
+- [x] 假客户端测试覆盖编排、解析失败、无模型分支（先例：v0.7 分析测试）

.scratch/v0.9/issues/05-report-analysis-export.md

@@ -4,9 +4,9 @@
 
 **Blocked by:** 01（聚合主料是体验记录）；04 若已就位则复核结论一并纳入，未就位不阻塞
 
-**Status:** ready-for-agent
+**Status:** completed
 
-- [ ] 活动报告聚合含探索发现维度，达成率与问题清单口径正确
-- [ ] 分析输入含探索摘要且不含全量对话
-- [ ] Markdown 导出追加探索附录；无数据时逐字节不变（缺则无痕）
-- [ ] 纯函数渲染与聚合单测 + 导出集成测试（先例：report_render 单测、活动导出测试）
+- [x] 活动报告聚合含探索发现维度，达成率与问题清单口径正确
+- [x] 分析输入含探索摘要且不含全量对话
+- [x] Markdown 导出追加探索附录；无数据时逐字节不变（缺则无痕）
+- [x] 纯函数渲染与聚合单测 + 导出集成测试（先例：report_render 单测、活动导出测试）

.scratch/v0.9/issues/06-drawer-exploration-section.md

@@ -4,9 +4,9 @@
 
 **Blocked by:** 05（报告聚合与查询端点就绪）
 
-**Status:** ready-for-agent
+**Status:** completed（a12eb09 + 9abf572 回复文本修复；Playwright 浏览器验证通过）
 
-- [ ] 报告抽屉呈现探索发现区块，数据与后端口径一致
-- [ ] 会话详情展示完整对话（角色、内容、延迟）与体验记录
-- [ ] 无探索数据时区块隐藏
-- [ ] tsc 通过；人工在浏览器验证展开/下钻交互
+- [x] 报告抽屉呈现探索发现区块，数据与后端口径一致
+- [x] 会话详情展示完整对话（角色、内容、延迟）与体验记录
+- [x] 无探索数据时区块隐藏
+- [x] tsc 通过；浏览器验证展开/下钻交互（Playwright 在 t480 实测：区块渲染、问题清单、judge 复核、会话下钻气泡与体验卡片均正常，截图 /tmp/ticket06）

.scratch/v0.9/issues/07-finalize-and-line-gates.md

@@ -4,9 +4,9 @@
 
 **Blocked by:** 01（会话实体与状态机）
 
-**Status:** ready-for-agent
+**Status:** completed
 
-- [ ] 活动终态结算时 running 会话转 expired，后续消息 409
-- [ ] 加速线自动触发被拒并附原因；手动触发可用
-- [ ] 正式线两种触发来源均可用
-- [ ] 集成测试覆盖结算与档位分支（先例：活动自动触发类测试）
+- [x] 活动终态结算时 running 会话转 expired，后续消息 409
+- [x] 加速线自动触发被拒并附原因；手动触发可用
+- [x] 正式线两种触发来源均可用
+- [x] 集成测试覆盖结算与档位分支（先例：活动自动触发类测试）

.scratch/v0.9/issues/08-openclaw-patrol-skill.md

@@ -4,9 +4,9 @@
 
 **Blocked by:** 03（巡检 API 契约就绪才能写说明书）
 
-**Status:** ready-for-agent
+**Status:** completed（skill fc167a8；常驻作业 e6f19ac8 已注册并重启存活；E2E 活动 e1f3bdde 自动派发会话 3e46b6a3，体验记录落库，活动已收尾取消）
 
-- [ ] skill 文档覆盖巡检决策、会话驱动、体验记录提交、409 收敛与衍生留痕
-- [ ] 部署同步管线分发 skill（复用现有同步步骤）
-- [ ] 常驻巡检作业注册成功，重启后不丢（依赖 OpenClaw 调度持久化）
-- [ ] t480 端到端验证：正式线活动窗口内出现自动派发的探索会话并有体验记录落库
+- [x] skill 文档覆盖巡检决策、会话驱动、体验记录提交、409 收敛与衍生留痕
+- [x] 部署同步管线分发 skill（复用现有同步步骤）
+- [x] 常驻巡检作业注册成功，重启后不丢（依赖 OpenClaw 调度持久化）
+- [x] t480 端到端验证：正式线活动窗口内出现自动派发的探索会话并有体验记录落库

.scratch/v0.9/spec.md

@@ -51,7 +51,10 @@
 - `POST /api/exploration/sessions`：body = campaign_id、persona、goal、seed_ref（可空）；平台硬校验（活动存在且 running、正式线仅接受 auto 或 manual 触发；加速线仅接受 manual；预算余量）→ 创建 running 会话；超限 409 + 原因
 - `POST /api/exploration/sessions/{id}/messages`：body = content；平台转发到目标的通道（复用 ChannelFactory）、持久化双方轮次、返回回复与延迟；单会话轮数超限 409；会话非 running 状态 409
 - `POST /api/exploration/sessions/{id}/close`：body = 体验记录（goal_achieved: bool、blockers[]、misled[]、emotion、notes）；结构校验 + 非法值归一（沿 v0.7 白名单经验）；会话转 completed
-- `GET /api/campaigns/{id}/exploration`：会话列表 + 探索发现聚合（会话数、达成率、问题清单）
+- 探索发现读出口（实现裁决：不设独立 `/api/campaigns/{id}/exploration`，避免与报告链重复聚合）：
+  - `GET /api/campaigns/{id}/report` 的 `exploration` 键携带探索发现聚合（会话数、达成率、问题清单、judge 复核），无探索数据时缺则无痕
+  - `GET /api/exploration/campaigns/{id}/sessions`：会话列表（下钻用）
+  - `GET /api/exploration/sessions/{id}/messages`：单会话对话记录（下钻用）
 - 活动窗口 finalize 时：仍 running 的会话转 expired，不再接受消息（在 `resolve_finalize` 同处挂接）
 
 ### 护栏（平台硬执行，不信任客户端自律）
@@ -62,11 +65,11 @@
 ### 判定双证据线
 
 - 体验判定：close 接口收结构化自报，为第一手证据
-- judge 抽样复核：会话结束后平台对对话抽样（默认 ≤3 段，控 token），经 judge 岗位模型产出质量维度复核；ChatClient 可注入（沿 v0.7 分析 seam）；异步后台执行，失败落错误不阻塞
+- judge 抽样复核：会话结束后平台对对话抽样（实现裁决：按「轮」抽样，默认 ≤3 轮即最多 6 条消息、每条截断 500 字符控 token），经 judge 岗位模型产出质量维度复核；ChatClient 可注入（沿 v0.7 分析 seam）；异步后台执行，失败落错误不阻塞
 
 ### 报告 / 分析 / 导出
 
-- 活动报告聚合新增"探索发现"维度（会话数、目标达成率、问题清单来自体验记录聚合）；周期对比口径不变（探索数据不参与）
+- 活动报告聚合新增"探索发现"维度（会话数、目标达成率、问题清单来自体验记录聚合）；judge 复核结论一并纳入（实现裁决：findings 全量收各档发现、poor 档排前，而非只收 poor）；周期对比口径不变（探索数据不参与）
 - v0.7 分析输入追加探索**摘要**（问题清单 + 达成统计，非全量对话）
 - Markdown 导出：存在探索数据时追加「探索发现」附录（缺则无痕），附录顺序：智能分析 → 周期对比 → 探索发现
 

backend/agenteval/evaluation/analysis.py

@@ -14,6 +14,7 @@ from typing import Any, Awaitable, Callable, Optional
 from sqlmodel import Session
 
 from agenteval.evaluation.report import generate_campaign_report
+from agenteval.exploration.summary import summarize_campaign_exploration
 from agenteval.model_gateway import ModelGateway
 from agenteval.models import Campaign, ModelCapability, RunStatus
 from agenteval.services.model_configs import (
@@ -152,6 +153,7 @@ async def _synthesize(
     report: dict[str, Any],
     stage1: list[dict[str, Any]],
     chat_client: ChatClient,
+    exploration_summary: Optional[dict[str, Any]] = None,
 ) -> dict[str, Any]:
     """阶段二：汇总各场景产出，产总体结论 + 跨场景问题 + 优先级建议。"""
     system_prompt = (
@@ -163,17 +165,18 @@ async def _synthesize(
         '"suggestions": [{"priority": 1, "text": "可执行的改善建议"}]}'
         "；问题按严重度从高到低排列，建议按优先级排列。全部使用中文。"
     )
-    user_prompt = json.dumps(
-        {
-            "活动": {
-                "名称": campaign.name,
-                "窗口秒数": campaign.window_seconds,
-                "总体指标": report.get("summary", {}),
-            },
-            "各场景诊断": stage1,
+    payload: dict[str, Any] = {
+        "活动": {
+            "名称": campaign.name,
+            "窗口秒数": campaign.window_seconds,
+            "总体指标": report.get("summary", {}),
         },
-        ensure_ascii=False,
-    )
+        "各场景诊断": stage1,
+    }
+    if exploration_summary is not None:
+        # 探索式评测证据线：只给统计与问题清单，不含全量对话
+        payload["探索发现"] = exploration_summary
+    user_prompt = json.dumps(payload, ensure_ascii=False)
     parsed = _parse_stage(
         await chat_client([
             {"role": "system", "content": system_prompt},
@@ -194,6 +197,7 @@ async def analyze_campaign(
     failure_samples: dict[str, list[dict[str, str]]],
     valid_run_ids: set[str],
     chat_client: ChatClient,
+    exploration_summary: Optional[dict[str, Any]] = None,
 ) -> dict[str, Any]:
     """两阶段编排：阶段一按场景并行诊断，阶段二综合研判。
 
@@ -209,7 +213,7 @@ async def analyze_campaign(
         _analyze_scenario(entry, failure_samples.get(entry["scenario_id"], []), chat_client)
         for entry in capability
     ])
-    stage2 = await _synthesize(campaign, report, list(stage1), chat_client)
+    stage2 = await _synthesize(campaign, report, list(stage1), chat_client, exploration_summary=exploration_summary)
 
     valid_scenario_ids = {entry["scenario_id"] for entry in capability}
     problems = []
@@ -310,6 +314,7 @@ async def execute_campaign_analysis(
                 failure_samples=collect_failure_samples(campaign_id, session),
                 valid_run_ids={r.id for r in runs if r.id},
                 chat_client=client,
+                exploration_summary=summarize_campaign_exploration(session, campaign_id),
             )
         except Exception as exc:
             _logger.warning("活动 %s 智能分析失败: %s", campaign_id, exc)

backend/agenteval/evaluation/campaign_runner.py

@@ -30,6 +30,7 @@ from agenteval.evaluation.campaign_scheduler import (
     resolve_finalize,
 )
 from agenteval.evaluation.engine import EvalEngine
+from agenteval.exploration.settlement import settle_campaign_sessions
 from agenteval.models import (
     Campaign,
     CampaignStatus,
@@ -201,6 +202,18 @@ def _auto_start_analysis(campaign: Campaign, session: Session) -> None:
         _logger.warning("活动 %s 自动分析触发失败（已跳过）: %s", campaign.id, exc)
 
 
+def _settle_exploration(campaign_id: str, session: Session) -> None:
+    """Expire dangling exploration sessions on completion.
+
+    Never blocks completion: a settlement failure is logged and skipped, the
+    same non-blocking semantics as ``_auto_start_analysis``.
+    """
+    try:
+        settle_campaign_sessions(campaign_id, session)
+    except Exception as exc:
+        _logger.warning("活动 %s 探索会话结算失败（已跳过）: %s", campaign_id, exc)
+
+
 # ── durable scheduler loop ──────────────────────────────────────────────────
 
 
@@ -250,6 +263,7 @@ async def run_campaign_loop(
                     current.status = CampaignStatus.COMPLETED
                     current.completed_at = utc_now()
                     repo.update(current)
+                    _settle_exploration(campaign_id, session)
                     _auto_start_analysis(current, session)
                 return
 

backend/agenteval/evaluation/report_render.py

@@ -347,19 +347,63 @@ def _window_line(report: dict[str, Any]) -> str:
     return f"**窗口**: {_offset(window)}（加速调试线 ×{scale:g}{wall}）"
 
 
+def _render_exploration_lines(exploration: dict[str, Any]) -> list[str]:
+    """探索发现附录：会话统计 + 问题清单 + judge 复核发现（有才渲染）。"""
+    lines: list[str] = [
+        "",
+        "## 探索发现",
+        "",
+        "| 指标 | 数值 |",
+        "|------|------|",
+        f"| 探索会话数 | {exploration.get('session_count', 0)} |",
+        f"| 有体验记录会话数 | {exploration.get('sessions_with_experience', 0)} |",
+        f"| 目标达成率 | {_pct(exploration.get('goal_achievement_rate'))} |",
+        "",
+        "### 问题清单",
+        "",
+    ]
+    issues = exploration.get("issues") or []
+    misled = exploration.get("misled") or []
+    if not issues and not misled:
+        lines.append("无")
+    else:
+        for item in issues:
+            lines.append(f"- {item['issue']} ×{item['count']}")
+        for item in misled:
+            lines.append(f"- （被误导）{item['issue']} ×{item['count']}")
+    judge = exploration.get("judge_review")
+    if judge:
+        lines += [
+            "",
+            "### judge 复核",
+            "",
+            f"已复核 {judge.get('reviewed_sessions', 0)} 个会话：",
+        ]
+        findings = judge.get("findings") or []
+        if findings:
+            for item in findings:
+                lines.append(f"- [{item.get('dimension')}/{item.get('rating')}] {item.get('comment')}")
+        else:
+            lines.append("- 未发现问题")
+        for summary_text in judge.get("summaries") or []:
+            lines.append(f"- 复核结论：{summary_text}")
+    return lines
+
+
 def render_campaign_markdown(
     report: dict[str, Any],
     *,
     analysis: Optional[dict[str, Any]] = None,
     comparison: Optional[dict[str, Any]] = None,
+    exploration: Optional[dict[str, Any]] = None,
     target_name: Optional[str] = None,
     scenario_names: Optional[dict[str, str]] = None,
 ) -> str:
     """Render a dual-axis campaign report dict as Markdown.
 
     ``analysis`` and ``comparison`` are the stored 智能分析 / 周期对比 results
-    (completed only); when absent the corresponding appendix is omitted
-    entirely.
+    (completed only); ``exploration`` is the on-the-fly 探索发现 aggregate;
+    when absent the corresponding appendix is omitted entirely (缺则无痕).
     """
     s = report["summary"]
     status = _CAMPAIGN_STATUS_LABELS.get(report.get("status"), report.get("status") or "-")
@@ -409,4 +453,6 @@ def render_campaign_markdown(
         lines += _render_analysis_lines(analysis, scenario_names or {})
     if comparison:
         lines += _render_comparison_lines(comparison, scenario_names or {})
+    if exploration:
+        lines += _render_exploration_lines(exploration)
     return "\n".join(lines)

backend/agenteval/exploration/judge.py

@@ -0,0 +1,192 @@
+"""Judge sampling review for exploration sessions (judge 岗位抽样复核).
+
+会话关闭后，平台对该会话对话抽样（默认 ≤3 段，控 token），经 judge 岗位模型
+独立复核，产出质量维度结论（态度、专业性、幻觉）落入会话的 ``judge_review``。
+复核是异步后台执行：失败落错误不阻塞会话状态，也不影响体验记录这条第一手
+证据线；未配置模型时静默跳过。LLM 调用沿 v0.7 分析的 ``ChatClient`` 接缝
+注入，测试用假客户端覆盖。
+"""
+
+import asyncio
+import json
+import logging
+from typing import Any, Optional
+
+from sqlmodel import Session
+
+from agenteval.evaluation.analysis import (
+    ChatClient,
+    gateway_chat_client,
+    resolve_analysis_model,
+)
+from agenteval.exploration.models import ExplorationMessage, ExplorationSession
+from agenteval.models import Campaign
+from agenteval.services.model_configs import ModelRuntimeConfig
+from agenteval.storage.db import get_session, iso_utc, utc_now
+from agenteval.storage.repository import (
+    CampaignRepository,
+    ExplorationMessageRepository,
+    ExplorationSessionRepository,
+)
+from agenteval.utils.llm import parse_json_from_llm_text
+
+_logger = logging.getLogger("agenteval")
+
+MAX_JUDGE_SAMPLES = 3
+SAMPLE_TEXT_LIMIT = 500
+VALID_DIMENSIONS = ("attitude", "professionalism", "hallucination")
+VALID_RATINGS = ("good", "acceptable", "poor")
+
+JUDGE_SYSTEM_PROMPT = """你是评测平台的 judge 岗位模型，负责独立复核一段「虚拟用户」与被评对象的探索对话。
+只依据给定对话抽样判断，不要臆测抽样之外的内容。
+输出必须是合法 JSON，且仅包含以下结构：
+{
+  "dimensions": [
+    {
+      "dimension": "attitude | professionalism | hallucination",
+      "rating": "good | acceptable | poor",
+      "comment": "一句话依据"
+    }
+  ],
+  "summary": "一句话总体结论"
+}
+维度说明：attitude=服务态度；professionalism=专业性（流程与答复正确性）；
+hallucination=幻觉（编造事实、政策或能力），无幻觉时 rating 为 good。
+三个维度必须各出现一次。"""
+
+
+class JudgeReviewError(RuntimeError):
+    """judge 复核失败（模型输出无法解析或调用失败），不影响会话状态。"""
+
+
+def resolve_judge_model(campaign: Campaign, session: Session) -> Optional[ModelRuntimeConfig]:
+    """解析 judge 岗位模型：沿用活动分析模型口径（活动覆盖 ?? 全局分析默认）。"""
+    return resolve_analysis_model(campaign, session)
+
+
+def sample_round_indexes(rounds: list[int], limit: int = MAX_JUDGE_SAMPLES) -> list[int]:
+    """超限时均匀取样（含首尾），保证抽样可复现。"""
+    if len(rounds) <= limit:
+        return list(rounds)
+    span = len(rounds) - 1
+    return [rounds[round(i * span / (limit - 1))] for i in range(limit)]
+
+
+def build_judge_messages(
+    session_obj: ExplorationSession, samples: list[ExplorationMessage]
+) -> list[dict[str, str]]:
+    transcript = []
+    for message in samples:
+        speaker = "虚拟用户" if message.role == "user" else "被评对象"
+        content = str(message.content)[:SAMPLE_TEXT_LIMIT]
+        transcript.append(f"[第 {message.round_index} 轮] {speaker}: {content}")
+    payload = {
+        "persona": session_obj.persona,
+        "goal": session_obj.goal,
+        "transcript": transcript,
+    }
+    return [
+        {"role": "system", "content": JUDGE_SYSTEM_PROMPT},
+        {"role": "user", "content": json.dumps(payload, ensure_ascii=False)},
+    ]
+
+
+def normalize_judge_review(raw: Any) -> dict[str, Any]:
+    """白名单归一：未知维度丢弃，非法档位归 acceptable（沿 v0.7 白名单经验）。"""
+    dims = raw.get("dimensions") if isinstance(raw, dict) else None
+    dims = dims if isinstance(dims, list) else []
+    normalized = []
+    for item in dims:
+        if not isinstance(item, dict):
+            continue
+        name = item.get("dimension")
+        if name not in VALID_DIMENSIONS:
+            continue
+        rating = item.get("rating")
+        normalized.append(
+            {
+                "dimension": name,
+                "rating": rating if rating in VALID_RATINGS else "acceptable",
+                "comment": str(item.get("comment") or "")[:SAMPLE_TEXT_LIMIT],
+            }
+        )
+    summary = str(raw.get("summary") or "")[:SAMPLE_TEXT_LIMIT] if isinstance(raw, dict) else ""
+    return {"dimensions": normalized, "summary": summary}
+
+
+async def judge_conversation(
+    session_obj: ExplorationSession,
+    samples: list[ExplorationMessage],
+    *,
+    chat_client: ChatClient,
+) -> dict[str, Any]:
+    text = await chat_client(build_judge_messages(session_obj, samples))
+    try:
+        raw = parse_json_from_llm_text(text)
+    except Exception as exc:
+        raise JudgeReviewError(f"judge 输出解析失败: {exc}") from exc
+    if not isinstance(raw, dict):
+        raise JudgeReviewError("judge 输出解析失败: 输出不是 JSON 对象")
+    review = normalize_judge_review(raw)
+    review["sampled_rounds"] = sorted({m.round_index for m in samples})
+    return review
+
+
+async def execute_judge_review(
+    exploration_session_id: str,
+    *,
+    chat_client: Optional[ChatClient] = None,
+) -> None:
+    """后台执行体：抽样 → judge 复核 → judge_review 落库。
+
+    与 Runs 同款后台任务约定：自持 Session、try/finally 关闭、失败落 error
+    不阻塞会话状态；未配置模型时静默跳过。
+    """
+    session = get_session()
+    try:
+        repo = ExplorationSessionRepository(session)
+        session_obj = repo.get(exploration_session_id)
+        if not session_obj:
+            return
+        campaign = CampaignRepository(session).get(session_obj.campaign_id)
+        runtime = resolve_judge_model(campaign, session) if campaign else None
+        if runtime is None:
+            _logger.info("探索会话 %s 跳过 judge 复核：未配置模型", exploration_session_id)
+            return
+
+        messages = ExplorationMessageRepository(session).list_by_session(session_obj.id)
+        round_indexes = sorted({m.round_index for m in messages})
+        sampled = set(sample_round_indexes(round_indexes))
+        samples = [m for m in messages if m.round_index in sampled]
+        if not samples:
+            return
+
+        client = chat_client or gateway_chat_client(runtime)
+        reviewed_at = iso_utc(utc_now())
+        try:
+            review = await judge_conversation(session_obj, samples, chat_client=client)
+            review.update(status="completed", model_config_id=runtime.id, reviewed_at=reviewed_at)
+        except Exception as exc:
+            _logger.warning("探索会话 %s judge 复核失败: %s", exploration_session_id, exc)
+            review = {
+                "status": "failed",
+                "model_config_id": runtime.id,
+                "reviewed_at": reviewed_at,
+                "error": str(exc)[:SAMPLE_TEXT_LIMIT],
+            }
+        session_obj.judge_review = review
+        repo.update(session_obj)
+    finally:
+        session.close()
+
+
+_BACKGROUND_TASKS: set[asyncio.Task] = set()
+
+
+def start_judge_review(exploration_session_id: str) -> asyncio.Task:
+    """以后台任务启动抽样复核（fire-and-forget；结果经会话 judge_review 观测）。"""
+    task = asyncio.create_task(execute_judge_review(exploration_session_id))
+    # 强引用防止未持有引用的任务被 GC（asyncio 已知坑）
+    _BACKGROUND_TASKS.add(task)
+    task.add_done_callback(_BACKGROUND_TASKS.discard)
+    return task

backend/agenteval/exploration/settlement.py

@@ -0,0 +1,29 @@
+"""Exploration session settlement on campaign finalize (活动终态结算).
+
+活动窗口结束或被取消进入终态时，仍 running 的探索会话转 expired——
+不留悬挂会话。结算后消息端点按既有状态机拒收（409）。
+"""
+
+from sqlmodel import Session
+
+from agenteval.exploration.models import ExplorationSessionStatus
+from agenteval.storage.db import utc_now
+from agenteval.storage.repository import ExplorationSessionRepository
+
+
+def settle_campaign_sessions(campaign_id: str, session: Session) -> int:
+    """Expire every still-running exploration session of a finalized campaign.
+
+    Returns the number of sessions expired. Completed/failed sessions keep
+    their evidence untouched.
+    """
+    repo = ExplorationSessionRepository(session)
+    expired = 0
+    for session_obj in repo.list_by_campaign(campaign_id):
+        if session_obj.status != ExplorationSessionStatus.RUNNING:
+            continue
+        session_obj.status = ExplorationSessionStatus.EXPIRED
+        session_obj.closed_at = utc_now()
+        repo.update(session_obj)
+        expired += 1
+    return expired

backend/agenteval/exploration/summary.py

@@ -0,0 +1,84 @@
+"""Exploration findings aggregation (探索发现聚合).
+
+把探索会话的体验记录聚合成探索摘要：会话数、目标达成率、问题清单
+（blockers / misled 按出现次数降序）、judge 复核结论（已就位才纳入；
+findings 全量收各档发现，poor 档排前）。这是体验记录这条第一手证据线
+进入报告 / 分析 / 导出三个出口前的唯一聚合口径；只含统计与问题清单，
+不含全量对话。
+"""
+
+from collections import Counter
+from typing import Any, Optional
+
+from sqlmodel import Session
+
+from agenteval.exploration.models import ExplorationSession
+
+_RATING_SEVERITY = {"poor": 0, "acceptable": 1, "good": 2}
+
+
+def _issue_list(counter: Counter) -> list[dict[str, Any]]:
+    return [{"issue": issue, "count": count} for issue, count in counter.most_common()]
+
+
+def summarize_exploration(sessions: list[ExplorationSession]) -> Optional[dict[str, Any]]:
+    """聚合一组探索会话为探索摘要；无会话返回 None（缺则无痕）。"""
+    if not sessions:
+        return None
+
+    with_experience = [s for s in sessions if s.experience is not None]
+    blockers: Counter = Counter()
+    misled: Counter = Counter()
+    achieved = 0
+    for session_obj in with_experience:
+        experience = session_obj.experience
+        if experience.get("goal_achieved"):
+            achieved += 1
+        blockers.update(experience.get("blockers") or [])
+        misled.update(experience.get("misled") or [])
+
+    judge = _summarize_judge_reviews(sessions)
+
+    return {
+        "session_count": len(sessions),
+        "sessions_with_experience": len(with_experience),
+        "goal_achieved_count": achieved,
+        "goal_achievement_rate": round(achieved / len(with_experience), 4) if with_experience else None,
+        "issues": _issue_list(blockers),
+        "misled": _issue_list(misled),
+        "judge_review": judge,
+    }
+
+
+def summarize_campaign_exploration(db_session: Session, campaign_id: str) -> Optional[dict[str, Any]]:
+    """取数 + 聚合一步完成：报告 / 分析 / 导出三个出口共用的探索摘要取法。"""
+    from agenteval.storage.repository import ExplorationSessionRepository
+
+    return summarize_exploration(ExplorationSessionRepository(db_session).list_by_campaign(campaign_id))
+
+
+def _summarize_judge_reviews(sessions: list[ExplorationSession]) -> Optional[dict[str, Any]]:
+    """只纳入复核完成的会话；findings 全量收各档发现（poor 档排前），
+    summaries 收复核总体结论。"""
+    findings: list[dict[str, Any]] = []
+    summaries: list[str] = []
+    reviewed = 0
+    for session_obj in sessions:
+        review = session_obj.judge_review
+        if not review or review.get("status") != "completed":
+            continue
+        reviewed += 1
+        if review.get("summary"):
+            summaries.append(str(review["summary"]))
+        for item in review.get("dimensions") or []:
+            findings.append(
+                {
+                    "dimension": item.get("dimension"),
+                    "rating": item.get("rating"),
+                    "comment": item.get("comment") or "",
+                }
+            )
+    if reviewed == 0:
+        return None
+    findings.sort(key=lambda f: _RATING_SEVERITY.get(f["rating"], len(_RATING_SEVERITY)))
+    return {"reviewed_sessions": reviewed, "findings": findings, "summaries": summaries}

backend/agenteval/web/routers/campaigns.py

@@ -26,6 +26,8 @@ from agenteval.evaluation.report import (
     summarize_campaign_progress,
 )
 from agenteval.evaluation.report_render import render_campaign_markdown
+from agenteval.exploration.settlement import settle_campaign_sessions
+from agenteval.exploration.summary import summarize_campaign_exploration
 from agenteval.models import Campaign, CampaignPlanEntry, CampaignStatus, ExplorationBudgetConfig, ExplorationSeeds
 from agenteval.storage.db import iso_utc, utc_now
 from agenteval.storage.model_config_repository import ModelConfigRepository
@@ -120,6 +122,7 @@ async def cancel_campaign(campaign_id: str, session: Session = Depends(get_db))
     campaign.completed_at = utc_now()
     repo.update(campaign)
     request_cancel(campaign_id)
+    settle_campaign_sessions(campaign_id, session)
     return campaign.model_dump()
 
 
@@ -130,7 +133,11 @@ async def get_campaign_report(campaign_id: str, session: Session = Depends(get_d
         raise HTTPException(status_code=404, detail="campaign not found")
     runs = RunRepository(session).list_by_campaign(campaign_id)
     scenario_names = {s.id: s.name for s in ScenarioRepository(session).list_all()}
-    return generate_campaign_report(campaign, runs, scenario_names=scenario_names)
+    report = generate_campaign_report(campaign, runs, scenario_names=scenario_names)
+    exploration = summarize_campaign_exploration(session, campaign_id)
+    if exploration is not None:
+        report["exploration"] = exploration
+    return report
 
 
 @router.get("/{campaign_id}/report/markdown")
@@ -174,6 +181,7 @@ async def get_campaign_report_markdown(campaign_id: str, session: Session = Depe
         generate_campaign_report(campaign, runs, scenario_names=scenario_names),
         analysis=analysis,
         comparison=comparison,
+        exploration=summarize_campaign_exploration(session, campaign_id),
         target_name=target_name,
         scenario_names=scenario_names,
     )

backend/agenteval/web/routers/exploration.py

@@ -10,6 +10,7 @@ with 409 plus a readable reason, so the rejection itself is feedback to the
 resident agent.
 """
 
+import json
 from datetime import datetime
 from typing import Any
 
@@ -20,6 +21,7 @@ from sqlmodel import Session
 from agenteval.channels.factory import ChannelFactory
 from agenteval.config import get_settings
 from agenteval.evaluation.report import generate_campaign_report
+from agenteval.exploration.judge import start_judge_review
 from agenteval.exploration.models import (
     ExplorationBudget,
     ExplorationMessage,
@@ -60,6 +62,21 @@ class CloseSessionRequest(BaseModel):
     experience: dict[str, Any]
 
 
+def _coerce_reply_text(content: Any) -> str:
+    """Flatten a reply payload to text; tutu returns msgBody as a parsed object,
+    and str(dict) would leak a Python repr into the view."""
+    if isinstance(content, str):
+        return content
+    if isinstance(content, dict):
+        for key in ("content", "text", "message"):
+            value = content.get(key)
+            if isinstance(value, str) and value:
+                return value
+    if content is None:
+        return ""
+    return json.dumps(content, ensure_ascii=False)
+
+
 def _check_creation_guardrails(
     campaign: Campaign,
     triggered_by: ExplorationTrigger,
@@ -200,6 +217,28 @@ async def create_session(
     return repo.create(session_obj).model_dump(mode="json")
 
 
+@router.get("/campaigns/{campaign_id}/sessions")
+async def list_campaign_sessions(
+    campaign_id: str,
+    session: Session = Depends(get_db),
+) -> dict:
+    if not CampaignRepository(session).get(campaign_id):
+        raise HTTPException(status_code=404, detail="campaign not found")
+    sessions = ExplorationSessionRepository(session).list_by_campaign(campaign_id)
+    return {"sessions": [s.model_dump(mode="json") for s in sessions]}
+
+
+@router.get("/sessions/{session_id}/messages")
+async def list_session_messages(
+    session_id: str,
+    session: Session = Depends(get_db),
+) -> dict:
+    if not ExplorationSessionRepository(session).get(session_id):
+        raise HTTPException(status_code=404, detail="exploration session not found")
+    messages = ExplorationMessageRepository(session).list_by_session(session_id)
+    return {"messages": [m.model_dump(mode="json") for m in messages]}
+
+
 @router.post("/sessions/{session_id}/messages")
 async def send_session_message(
     session_id: str,
@@ -260,7 +299,7 @@ async def send_session_message(
 
     received_at = utc_now()
     latency_ms = int((received_at - sent_at).total_seconds() * 1000)
-    reply_text = str(reply.content)
+    reply_text = _coerce_reply_text(reply.content)
     message_repo.save_message(
         ExplorationMessage(
             session_id=session_obj.id,
@@ -290,4 +329,6 @@ async def close_session(
     session_obj.experience = normalize_experience(request.experience)
     session_obj.status = ExplorationSessionStatus.COMPLETED
     session_obj.closed_at = utc_now()
-    return repo.update(session_obj).model_dump(mode="json")
+    updated = repo.update(session_obj)
+    start_judge_review(session_obj.id)
+    return updated.model_dump(mode="json")

backend/plugins/openclaw/README.md

@@ -32,7 +32,17 @@
 部署脚本（`scripts/deploy-t480.sh`）会自动把它同步到 OpenClaw 工作区：
 `data/openclaw/workspace/skills/agenteval-run/SKILL.md`。
 
-### 方式二：Python Skill 封装
+### 方式二：agenteval-patrol 巡检技能（探索式评测，v0.9）
+
+`skills/agenteval-patrol/SKILL.md` 指导 AI 助手作为常驻巡检代理完成探索闭环：
+调巡检 API（`GET /api/exploration/patrol`）→ 研判新结果与预算余量 → 决定是否派发
+探索会话 → 以虚拟用户身份驱动会话对话 → 提交体验记录关闭会话。说明书内含收到
+409 时的收敛行为与衍生变体的种子回溯（`seed_ref`）要求。
+
+部署脚本会把 `backend/plugins/openclaw/skills/` 下的每个技能同步到
+`data/openclaw/workspace/skills/<skill-name>/SKILL.md`。
+
+### 方式三：Python Skill 封装
 
 参考 `agenteval_skill.py`（httpx 调 API），在 OpenClaw 中注册技能时传入配置：
 
@@ -60,6 +70,21 @@ config:
   report_format: "json"
 ```
 
+### 常驻巡检作业（探索式评测）
+
+巡检代理需注册为全局常驻作业，默认节拍 1 小时；调度状态由 OpenClaw 自身持久化，
+重启后不丢：
+
+```yaml
+skill: agenteval-patrol
+schedule: "0 * * * *"  # 每小时整点巡检一次
+config:
+  api_base_url: "http://agenteval:8000"
+```
+
+巡检 API 自带水位（`last_patrolled_at`），每个节拍只报增量结果，重复触发不会
+重复上报；预算与间隔护栏在服务端强制执行，代理收到 409 即收敛。
+
 ## API 速查
 
 | 操作 | 请求 |
@@ -67,6 +92,10 @@ config:
 | 启动评测 | `POST /api/runs` body: `{"target_id", "scenario_id", "triggered_by": "ai_assistant"}` |
 | 查询状态 | `GET /api/runs/{run_id}`（轮询至 completed/failed） |
 | 获取报告 | `GET /api/reports/{run_id}` |
+| 巡检探索 | `GET /api/exploration/patrol`（无参数，报增量新结果与预算余量） |
+| 派发探索会话 | `POST /api/exploration/sessions` body: `{"campaign_id", "persona", "goal", "seed_ref?", "triggered_by"}` |
+| 驱动会话对话 | `POST /api/exploration/sessions/{session_id}/messages` body: `{"content"}` |
+| 提交体验记录 | `POST /api/exploration/sessions/{session_id}/close` body: `{"experience"}` |
 
 如果平台设置了 `AGENTEVAL_API_KEY`，所有请求需带 `X-API-Key` 头。
 

backend/plugins/openclaw/skills/agenteval-patrol/SKILL.md

@@ -0,0 +1,110 @@
+---
+name: agenteval-patrol
+description: 常驻巡检：巡检 AgentEvalTool 评测活动，研判新结果与预算余量，自主派发并驱动探索式评测会话
+---
+
+你是平台的常驻巡检代理。本 skill 指导你按「巡检 → 研判 → 派发 → 驱动 → 提交」的闭环行动。
+所有操作必须走 AgentEvalTool 标准 HTTP API（禁止直接调 CLI 或操作数据库）。
+
+平台可能启用了 API Key 鉴权。每次执行命令前先读取密钥（文件不存在则为空，不影响未启用鉴权的环境）：
+
+```bash
+KEY=$(cat ~/.openclaw/agenteval-api-key 2>/dev/null)
+```
+
+以下所有 curl 命令都必须带 `-H "X-API-Key: $KEY"`。
+
+## 巡检（闭环第一步）
+
+```bash
+curl -s -H "X-API-Key: $KEY" http://agenteval:8000/api/exploration/patrol | python3 -m json.tool
+```
+
+返回所有参与探索的进行中正式线活动。每个活动包含：
+
+- `new_results`：自上次巡检以来的新评测结果（null 表示无新增）。有值时先消化它——通过率下滑的能力项是本次探索的优先方向。
+- `budget`：探索预算台账。`remaining_sessions` 为剩余可派发会话数；`seconds_since_last_session` 与 `min_interval_seconds` 用于判断会话间隔是否已满足。
+- 该接口自带水位：每次调用只报增量，无需自己记录上次结果。
+
+## 研判与派发决策
+
+对每个活动，按以下规则决定是否派发探索会话：
+
+1. `remaining_sessions` 为 0 → 不派发（预算耗尽）。
+2. `seconds_since_last_session` 非 null 且小于 `min_interval_seconds` → 不派发（间隔未到，等下个节拍）。
+3. 其余情况可派发一个会话。persona 与 goal 必须取自活动的种子集：
+
+```bash
+curl -s -H "X-API-Key: $KEY" http://agenteval:8000/api/campaigns/<campaign_id> | python3 -m json.tool
+```
+
+读取响应中的 `exploration_seeds.personas` 与 `exploration_seeds.goals`，各取其一组合。允许在种子上做衍生变体（如把「查账单」衍生为「查账单并开发票」），但必须在 `seed_ref` 里回溯来源。
+
+## 派发探索会话
+
+```bash
+curl -s -X POST http://agenteval:8000/api/exploration/sessions \
+  -H "X-API-Key: $KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+    "campaign_id": "<campaign_id>",
+    "persona": {"name": "<人设名>", "traits": ["<特质1>", "<特质2>"]},
+    "goal": "<探索目标>",
+    "seed_ref": {"persona_seed": "<所用 persona 种子原文>", "goal_seed": "<所用 goal 种子原文>", "derived": false},
+    "triggered_by": "auto"
+  }'
+```
+
+- 直接使用种子时 `derived` 为 false；衍生变体时为 true，并追加 `"derived_from": "<原种子>"` 与 `"variant": "<变异点说明>"`——留痕是探索数据可回溯的底线。
+- 响应中的 `id` 即 session_id，后续所有步骤都要用它。
+
+## 驱动会话对话
+
+以 persona 的身份向目标推进 goal，每轮发送一条用户消息并同步收到回复：
+
+```bash
+curl -s -X POST http://agenteval:8000/api/exploration/sessions/<session_id>/messages \
+  -H "X-API-Key: $KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"content": "<本轮用户消息>"}'
+```
+
+- 根据 `reply` 决定下一轮：追问、纠正、换路径，直到目标达成或确认走不通。
+- 单会话默认最多 12 轮，以 patrol 返回的 `budget.max_turns` 为准，接近上限时尽快收尾。
+- 502 表示通道异常；连续两次 502 就结束会话，把异常写进体验记录备注。
+
+## 提交体验记录（闭环收尾）
+
+会话必须显式关闭，提交第一人称体验记录——这是探索证据链的源头，务必如实：
+
+```bash
+curl -s -X POST http://agenteval:8000/api/exploration/sessions/<session_id>/close \
+  -H "X-API-Key: $KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+    "experience": {
+      "goal_achieved": true,
+      "blockers": ["<遇到的障碍，逐条列出>"],
+      "misled": ["<被误导的经历：回答把你带偏的地方>"],
+      "emotion": "positive",
+      "notes": "<整体体验简述>"
+    }
+  }'
+```
+
+- `emotion` 只能是 positive / neutral / confused / frustrated 之一，按真实体感选择。
+- `blockers` 与 `misled` 写具体事实（如「让我重复提供订单号三次」），不写空泛评价。
+
+## 收到 409 的收敛行为
+
+409 是平台账本的明确拒绝，一律停止当前动作、不重试：
+
+- 创建会话 409（预算耗尽 / 间隔不足 / 活动非进行中）→ 本活动本轮放弃派发，记录原因，等下个巡检节拍。
+- 发消息 409（会话不在进行中 / 轮数超限）→ 立即走关闭流程提交已有体验记录；若关闭也返回 409，说明会话已被平台收口（如活动终态结算），放弃并记录原因。
+- 404（活动或会话不存在）→ 对象已消失，跳过。
+
+## 汇报要求
+
+每个巡检节拍结束后向用户汇报：巡检到的活动数、派发的会话数（含被 409 拦截的原因）、关闭的会话及其目标达成情况与关键问题。
+
+请将 <campaign_id>、<session_id> 等占位符替换为实际值。

frontend/web/src/api.ts

@@ -324,6 +324,69 @@ export interface ExplorationBudgetConfig {
   min_interval_seconds?: number | null
 }
 
+export interface ExplorationIssueCount {
+  issue: string
+  count: number
+}
+
+export interface ExplorationJudgeReview {
+  reviewed_sessions: number
+  findings: { dimension: string; rating: string; comment: string }[]
+  summaries: string[]
+}
+
+export interface ExplorationSummary {
+  session_count: number
+  sessions_with_experience: number
+  goal_achieved_count: number
+  goal_achievement_rate: number | null
+  issues: ExplorationIssueCount[]
+  misled: ExplorationIssueCount[]
+  judge_review: ExplorationJudgeReview | null
+}
+
+export interface ExplorationExperience {
+  goal_achieved: boolean
+  blockers: string[]
+  misled: string[]
+  emotion: string
+  notes: string
+}
+
+export interface ExplorationSession {
+  id: string
+  campaign_id: string
+  target_id: string
+  persona: Record<string, unknown>
+  goal: string
+  seed_ref: Record<string, unknown> | null
+  status: string
+  triggered_by: string
+  experience: ExplorationExperience | null
+  judge_review: Record<string, unknown> | null
+  turn_count: number
+  error: string | null
+  created_at: string | null
+  closed_at: string | null
+}
+
+export interface ExplorationMessage {
+  id: string
+  session_id: string
+  round_index: number
+  role: string
+  content: string
+  latency_ms: number | null
+  created_at: string | null
+}
+
+export const explorationApi = {
+  listSessions: (campaignId: string) =>
+    api.get<{ sessions: ExplorationSession[] }>(`/exploration/campaigns/${campaignId}/sessions`),
+  listMessages: (sessionId: string) =>
+    api.get<{ messages: ExplorationMessage[] }>(`/exploration/sessions/${sessionId}/messages`),
+}
+
 export interface CampaignProgress {
   current_offset_seconds: number
   spawned_runs: number
@@ -394,6 +457,7 @@ export interface CampaignReport {
   }
   time_trend: CampaignTrendBucket[]
   capability_summary: CampaignCapability[]
+  exploration?: ExplorationSummary | null
 }
 
 export interface CampaignTimelineEntry {

frontend/web/src/components/ChatBubble.tsx

@@ -7,9 +7,11 @@ interface ChatBubbleProps {
   content: ReactNode
   meta?: ReactNode
   pending?: boolean
+  /** 镜像布局：探索下钻里虚拟用户居左、被评对象居右（与常规聊天相反） */
+  mirrored?: boolean
 }
 
-export default function ChatBubble({ role, content, meta, pending }: ChatBubbleProps) {
+export default function ChatBubble({ role, content, meta, pending, mirrored }: ChatBubbleProps) {
   const isUser = role === 'user'
   const isAgent = role === 'agent'
   const isSystem = role === 'system'
@@ -17,8 +19,9 @@ export default function ChatBubble({ role, content, meta, pending }: ChatBubbleP
   const bg = isUser ? colors.chatUser : isAgent ? colors.chatAgent : colors.chatSystemError
   const border = isSystem ? '#ffccc7' : 'transparent'
   const textColor = isSystem ? '#cf1322' : colors.text
-  const align = isUser ? 'flex-end' : 'flex-start'
-  const radius = isUser ? '10px 10px 2px 10px' : '10px 10px 10px 2px'
+  const alignRight = isSystem ? false : isUser ? !mirrored : !!mirrored
+  const align = alignRight ? 'flex-end' : 'flex-start'
+  const radius = alignRight ? '10px 10px 2px 10px' : '10px 10px 10px 2px'
 
   const icon = isUser
     ? <UserOutlined style={{ color: colors.primary }} />
@@ -29,7 +32,7 @@ export default function ChatBubble({ role, content, meta, pending }: ChatBubbleP
 
   return (
     <div style={{ display: 'flex', justifyContent: align, marginBottom: 6 }}>
-      <div style={{ maxWidth: '82%', display: 'flex', flexDirection: 'column', alignItems: isUser ? 'flex-end' : 'flex-start', gap: 2 }}>
+      <div style={{ maxWidth: '82%', display: 'flex', flexDirection: 'column', alignItems: alignRight ? 'flex-end' : 'flex-start', gap: 2 }}>
         <div style={{ display: 'flex', gap: 4, alignItems: 'center', fontSize: 11, color: colors.textMuted }}>
           {!isUser && icon}
           <span>{label}</span>

frontend/web/src/components/ExplorationSection.tsx

@@ -0,0 +1,227 @@
+import { useEffect, useState } from 'react'
+import { Collapse, Empty, Space, Spin, Tag } from 'antd'
+import ChatBubble from './ChatBubble'
+import {
+  explorationApi,
+  type ExplorationMessage,
+  type ExplorationSession,
+  type ExplorationSummary,
+} from '../api'
+import { colors } from '../tokens'
+import { shortDateTime } from '../utils/date'
+
+const SESSION_STATUS: Record<string, { label: string; color: string }> = {
+  running: { label: '进行中', color: 'processing' },
+  completed: { label: '已完成', color: 'success' },
+  failed: { label: '失败', color: 'error' },
+  expired: { label: '已过期', color: 'default' },
+}
+
+const EMOTION_LABELS: Record<string, string> = {
+  positive: '满意',
+  neutral: '平静',
+  confused: '困惑',
+  frustrated: '沮丧',
+}
+
+const DIMENSION_LABELS: Record<string, string> = {
+  attitude: '服务态度',
+  professionalism: '专业度',
+  hallucination: '幻觉',
+}
+
+const RATING_TAG: Record<string, { label: string; color: string }> = {
+  poor: { label: '差', color: 'red' },
+  acceptable: { label: '一般', color: 'orange' },
+  good: { label: '优', color: 'green' },
+}
+
+function pct(rate: number | null): string {
+  return rate == null ? '—' : `${(rate * 100).toFixed(1)}%`
+}
+
+function IssueList({ label, items, misled }: {
+  label: string
+  items: { issue: string; count: number }[]
+  misled?: boolean
+}) {
+  if (items.length === 0) return null
+  return (
+    <div style={{ marginBottom: 10 }}>
+      <div style={{ fontSize: 12, color: colors.textSecondary, marginBottom: 6 }}>{label}</div>
+      <Space size={8} wrap>
+        {items.map((item, i) => (
+          <Tag key={i} color={misled ? 'orange' : 'red'}>
+            {misled ? `（被误导）${item.issue}` : item.issue} ×{item.count}
+          </Tag>
+        ))}
+      </Space>
+    </div>
+  )
+}
+
+function SessionDetail({ session, messages }: {
+  session: ExplorationSession
+  messages: ExplorationMessage[] | null
+}) {
+  const experience = session.experience
+  return (
+    <div>
+      <div style={{ fontSize: 12, color: colors.textSecondary, marginBottom: 10 }}>
+        目标：{session.goal}
+        {session.created_at && <> · 开始于 {shortDateTime(session.created_at)}</>}
+        {session.closed_at && <> · 结束于 {shortDateTime(session.closed_at)}</>}
+      </div>
+
+      {!messages && <Spin size="small" />}
+      {messages && messages.length === 0 && (
+        <Empty description="暂无对话记录" image={Empty.PRESENTED_IMAGE_SIMPLE} />
+      )}
+      {messages && messages.map((m) => (
+        <ChatBubble
+          key={m.id}
+          role={m.role === 'user' ? 'user' : 'agent'}
+          mirrored
+          content={m.content}
+          meta={m.role !== 'user' && m.latency_ms != null ? (
+            <span style={{ color: colors.textMuted, fontSize: 11 }}>{m.latency_ms}ms</span>
+          ) : undefined}
+        />
+      ))}
+
+      {experience && (
+        <div style={{
+          marginTop: 8, border: `1px solid ${colors.border}`, borderRadius: 8,
+          background: colors.bgSubtle, padding: '10px 12px', fontSize: 13, lineHeight: 1.8,
+        }}>
+          <div style={{ fontWeight: 500, marginBottom: 4 }}>体验记录</div>
+          <Space size={8} wrap>
+            {experience.goal_achieved
+              ? <Tag color="success">目标达成</Tag>
+              : <Tag color="error">目标未达成</Tag>}
+            <Tag>{EMOTION_LABELS[experience.emotion] ?? experience.emotion}</Tag>
+          </Space>
+          {experience.blockers.length > 0 && (
+            <div style={{ color: colors.textSecondary }}>障碍：{experience.blockers.join('；')}</div>
+          )}
+          {experience.misled.length > 0 && (
+            <div style={{ color: colors.warning }}>被误导：{experience.misled.join('；')}</div>
+          )}
+          {experience.notes && (
+            <div style={{ color: colors.textSecondary }}>备注：{experience.notes}</div>
+          )}
+        </div>
+      )}
+      {session.error && (
+        <div style={{ marginTop: 8, fontSize: 12, color: colors.textMuted }}>会话异常：{session.error}</div>
+      )}
+    </div>
+  )
+}
+
+export default function ExplorationSection({ campaignId, summary }: {
+  campaignId: string
+  summary: ExplorationSummary
+}) {
+  const [sessions, setSessions] = useState<ExplorationSession[] | null>(null)
+  const [loading, setLoading] = useState(false)
+  const [messageCache, setMessageCache] = useState<Record<string, ExplorationMessage[]>>({})
+
+  useEffect(() => {
+    let cancelled = false
+    setLoading(true)
+    explorationApi.listSessions(campaignId)
+      .then((res) => { if (!cancelled) setSessions(res.data.sessions) })
+      .catch(() => undefined) // 拦截器已弹错；避免未处理 rejection
+      .finally(() => { if (!cancelled) setLoading(false) })
+    return () => { cancelled = true }
+  }, [campaignId])
+
+  const onExpand = (keys: string | string[]) => {
+    for (const key of Array.isArray(keys) ? keys : [keys]) {
+      if (key in messageCache) continue
+      void explorationApi.listMessages(key)
+        .then((res) => { setMessageCache((prev) => ({ ...prev, [key]: res.data.messages })) })
+        .catch(() => undefined) // 拦截器已弹错；收起再展开可重试
+    }
+  }
+
+  const judge = summary.judge_review
+  return (
+    <div style={{
+      border: `1px solid ${colors.border}`, borderRadius: 8,
+      padding: '12px 14px', marginBottom: 16,
+    }}>
+      <div style={{ fontSize: 12, color: colors.textSecondary, marginBottom: 10 }}>
+        探索会话 {summary.session_count} 个 · 有体验记录 {summary.sessions_with_experience} 个
+        {' · '}目标达成率 {pct(summary.goal_achievement_rate)}（{summary.goal_achieved_count}/{summary.sessions_with_experience}）
+      </div>
+
+      <IssueList label="问题清单（体验记录）" items={summary.issues} />
+      <IssueList label="被误导清单（体验记录）" items={summary.misled} misled />
+      {summary.issues.length === 0 && summary.misled.length === 0 && (
+        <div style={{ fontSize: 12, color: colors.textMuted, marginBottom: 10 }}>体验记录未报告问题</div>
+      )}
+
+      {judge && (
+        <div style={{ marginBottom: 10 }}>
+          <div style={{ fontSize: 12, color: colors.textSecondary, marginBottom: 6 }}>
+            judge 复核（{judge.reviewed_sessions} 个会话抽样）
+          </div>
+          {judge.findings.length === 0
+            ? <div style={{ fontSize: 13, color: colors.textSecondary }}>无复核发现</div>
+            : (
+              <Space direction="vertical" size={4} style={{ width: '100%' }}>
+                {judge.findings.map((f, i) => (
+                  <div key={i} style={{ fontSize: 13 }}>
+                    <Tag color={RATING_TAG[f.rating]?.color ?? 'default'} style={{ margin: 0 }}>
+                      {DIMENSION_LABELS[f.dimension] ?? f.dimension}
+                      {' · '}
+                      {RATING_TAG[f.rating]?.label ?? f.rating}
+                    </Tag>
+                    {' '}{f.comment}
+                  </div>
+                ))}
+              </Space>
+            )}
+          {judge.summaries.map((s, i) => (
+            <div key={i} style={{ fontSize: 13, color: colors.textSecondary, marginTop: 4 }}>
+              复核结论：{s}
+            </div>
+          ))}
+        </div>
+      )}
+
+      {loading && <Spin size="small" />}
+      {!loading && sessions && sessions.length === 0 && (
+        <Empty description="暂无探索会话" image={Empty.PRESENTED_IMAGE_SIMPLE} />
+      )}
+      {sessions && sessions.length > 0 && (
+        <Collapse
+          size="small"
+          onChange={onExpand}
+          items={sessions.map((s) => {
+            const meta = SESSION_STATUS[s.status] ?? SESSION_STATUS.running
+            const personaName = typeof s.persona?.name === 'string' ? s.persona.name : s.id.slice(0, 8)
+            return {
+              key: s.id,
+              label: (
+                <Space size={8} wrap>
+                  <span style={{ fontWeight: 500 }}>{personaName}</span>
+                  <Tag color={meta.color}>{meta.label}</Tag>
+                  <Tag>{s.turn_count} 轮</Tag>
+                  {s.experience?.goal_achieved != null && (
+                    s.experience.goal_achieved
+                      ? <Tag color="success">目标达成</Tag>
+                      : <Tag color="error">目标未达成</Tag>
+                  )}
+                </Space>
+              ),
+              children: <SessionDetail session={s} messages={messageCache[s.id] ?? null} />,
+            }
+          })}
+        />
+      )}
+    </div>
+  )
+}

frontend/web/src/pages/Campaigns.tsx

@@ -8,7 +8,7 @@ import {
   PlusOutlined, ReloadOutlined, StopOutlined, BarChartOutlined,
   FileMarkdownOutlined, MinusCircleOutlined, QuestionCircleOutlined,
   RocketOutlined, CheckCircleOutlined, SafetyOutlined, ClockCircleOutlined,
-  RobotOutlined, SwapOutlined,
+  RobotOutlined, SwapOutlined, CompassOutlined,
 } from '@ant-design/icons'
 import { Line, Bar } from '@ant-design/charts'
 import PageWrapper from '../components/PageWrapper'
@@ -25,6 +25,7 @@ import { deriveTimeScale, acceleratedDuration, formatScale } from '../utils/camp
 import WindowTimeline, { type TimelineMarker } from '../components/WindowTimeline'
 import CampaignRunTimeline from '../components/CampaignRunTimeline'
 import PeriodComparisonSection from '../components/PeriodComparisonSection'
+import ExplorationSection from '../components/ExplorationSection'
 import { useResource } from '../hooks/useResource'
 import { usePolling } from '../hooks/usePolling'
 import { useTabStore } from '../stores/tabStore'
@@ -1055,6 +1056,17 @@ export default function CampaignsPage() {
                 onGenerate={generateComparison}
               />
 
+              {report.exploration && (
+                <>
+                  <SectionTitle><CompassOutlined /> 探索发现</SectionTitle>
+                  <ExplorationSection
+                    key={report.campaign_id}
+                    campaignId={report.campaign_id}
+                    summary={report.exploration}
+                  />
+                </>
+              )}
+
               <SectionTitle>过程时间轴</SectionTitle>
               <div style={{
                 border: `1px solid ${colors.border}`, borderRadius: 8,

scripts/deploy-t480.sh

@@ -93,14 +93,16 @@ run rsync -az --delete \
     ./ "$HOST:$REMOTE_DIR/"
 
 # ── OpenClaw skill sync ────────────────────────────────────────────────
-# The agenteval-run skill lives in the openclaw workspace volume (data/),
-# which rsync excludes. Sync the versioned copy explicitly so the AI
-# assistant always follows the standard HTTP API flow.
-SKILL_SRC="backend/plugins/openclaw/skills/agenteval-run/SKILL.md"
-SKILL_DST="$REMOTE_DIR/data/openclaw/workspace/skills/agenteval-run/SKILL.md"
-log "sync OpenClaw skill → $SKILL_DST"
-run ssh "$HOST" "mkdir -p $(dirname "$SKILL_DST")"
-run rsync -az "$SKILL_SRC" "$HOST:$SKILL_DST"
+# OpenClaw skills live in the openclaw workspace volume (data/), which
+# rsync excludes. Sync every versioned skill explicitly so the AI
+# assistant always follows the standard HTTP API flows.
+for SKILL_SRC in backend/plugins/openclaw/skills/*/SKILL.md; do
+    SKILL_NAME=$(basename "$(dirname "$SKILL_SRC")")
+    SKILL_DST="$REMOTE_DIR/data/openclaw/workspace/skills/$SKILL_NAME/SKILL.md"
+    log "sync OpenClaw skill → $SKILL_DST"
+    run ssh "$HOST" "mkdir -p $(dirname "$SKILL_DST")"
+    run rsync -az "$SKILL_SRC" "$HOST:$SKILL_DST"
+done
 
 # Provision the API key file the skill reads (~/.openclaw/agenteval-api-key in
 # the openclaw container = data/openclaw/agenteval-api-key on the host). Sourced

tests/integration/test_campaigns_api.py

@@ -302,6 +302,61 @@ async def test_campaign_report_markdown_export(client, seeded_db):
     assert "## 能力汇总" in resp.text
 
 
+# ── exploration findings in report/export (v0.9 ticket 05) ───────────────────
+
+
+def _seed_exploration_session(seeded_db, campaign_id: str) -> None:
+    from agenteval.exploration.models import ExplorationSession
+    from agenteval.storage.repository import ExplorationSessionRepository
+
+    repo = ExplorationSessionRepository(seeded_db)
+    session_obj = repo.create(
+        ExplorationSession(campaign_id=campaign_id, target_id="t-1", persona={"name": "x"}, goal="缴费")
+    )
+    session_obj.experience = {
+        "goal_achieved": True,
+        "blockers": ["缴费入口难找"],
+        "misled": [],
+        "emotion": "neutral",
+        "notes": "",
+    }
+    repo.update(session_obj)
+
+
+async def test_campaign_report_includes_exploration_findings(client, seeded_db):
+    campaign_id = (await client.post("/api/campaigns", json=_valid_payload())).json()["id"]
+    _seed_exploration_session(seeded_db, campaign_id)
+
+    report = (await client.get(f"/api/campaigns/{campaign_id}/report")).json()
+    exploration = report["exploration"]
+    assert exploration["session_count"] == 1
+    assert exploration["goal_achievement_rate"] == 1.0
+    assert exploration["issues"] == [{"issue": "缴费入口难找", "count": 1}]
+
+
+async def test_campaign_report_without_exploration_omits_key(client, seeded_db):
+    campaign_id = (await client.post("/api/campaigns", json=_valid_payload())).json()["id"]
+    report = (await client.get(f"/api/campaigns/{campaign_id}/report")).json()
+    assert "exploration" not in report
+
+
+async def test_markdown_export_appends_exploration_appendix(client, seeded_db):
+    campaign_id = (await client.post("/api/campaigns", json=_valid_payload())).json()["id"]
+    _seed_exploration_session(seeded_db, campaign_id)
+
+    resp = await client.get(f"/api/campaigns/{campaign_id}/report/markdown")
+    assert resp.status_code == 200
+    assert "## 探索发现" in resp.text
+    assert "缴费入口难找 ×1" in resp.text
+
+
+async def test_markdown_export_without_exploration_leaves_no_trace(client, seeded_db):
+    campaign_id = (await client.post("/api/campaigns", json=_valid_payload())).json()["id"]
+    resp = await client.get(f"/api/campaigns/{campaign_id}/report/markdown")
+    assert resp.status_code == 200
+    assert "探索发现" not in resp.text
+
+
 # ── campaign timeline endpoint (ticket 06) ───────────────────────────────────
 
 async def test_campaign_timeline_structure(client, seeded_db):

tests/integration/test_exploration_api.py

@@ -54,6 +54,11 @@ def seeded_db(db_session, monkeypatch):
 
     app.dependency_overrides[get_db] = _test_get_db
 
+    # 隔离后台 judge 复核任务：API 测试不真正派发后台任务
+    from agenteval.web.routers import exploration as exploration_router
+
+    monkeypatch.setattr(exploration_router, "start_judge_review", lambda session_id: None)
+
     target = EvalTarget(
         id="t-1",
         name="mock-target",
@@ -170,6 +175,37 @@ async def test_full_lifecycle_create_message_close(seeded_db, mock_channel, clie
     assert closed["closed_at"] is not None
 
 
+async def test_dict_reply_content_is_flattened_to_text(seeded_db, monkeypatch, client):
+    """通道回复 content 为对象（如 tutu msgBody）时应提取文本而非存 str(dict)。"""
+    from agenteval.channels.base import Reply
+    from tests.unit.mock_channel import MockChannel
+
+    class _DictReplyChannel(MockChannel):
+        async def poll_reply(self, question_msg_id, timeout=30.0, poll_interval=1.0):
+            return Reply(
+                question_msg_id=question_msg_id,
+                content={"content": "您好，我是客服"},
+                raw_message={},
+            )
+
+    _stub_channel_factory(monkeypatch, _DictReplyChannel(reply_text="unused"))
+
+    resp = await _create_session(client)
+    assert resp.status_code == 200, resp.text
+    session_id = resp.json()["id"]
+
+    resp = await client.post(
+        f"/api/exploration/sessions/{session_id}/messages",
+        json={"content": "你好"},
+    )
+    assert resp.status_code == 200, resp.text
+    assert resp.json()["reply"] == "您好，我是客服"
+
+    messages_resp = await client.get(f"/api/exploration/sessions/{session_id}/messages")
+    assistant = [m for m in messages_resp.json()["messages"] if m["role"] == "assistant"]
+    assert assistant[0]["content"] == "您好，我是客服"
+
+
 async def test_create_requires_existing_running_campaign(seeded_db, client):
     resp = await _create_session(client, campaign_id="nope")
     assert resp.status_code == 404
@@ -330,6 +366,80 @@ async def test_close_twice_rejected(seeded_db, mock_channel, client):
     assert resp.status_code == 409
 
 
+async def test_close_triggers_judge_review(seeded_db, mock_channel, client, monkeypatch):
+    from agenteval.web.routers import exploration as exploration_router
+
+    started: list[str] = []
+    monkeypatch.setattr(exploration_router, "start_judge_review", started.append)
+
+    session_id = (await _create_session(client)).json()["id"]
+    resp = await client.post(
+        f"/api/exploration/sessions/{session_id}/close", json={"experience": {"goal_achieved": True}}
+    )
+    assert resp.status_code == 200, resp.text
+    assert started == [session_id]
+
+
+# ---------------------------------------------------------------- read endpoints (ticket 06)
+
+
+async def test_list_campaign_sessions_returns_lifecycle_fields(seeded_db, mock_channel, client):
+    resp = await _create_session(client)
+    assert resp.status_code == 200, resp.text
+    session_id = resp.json()["id"]
+
+    await client.post(f"/api/exploration/sessions/{session_id}/messages", json={"content": "查账单"})
+    resp = await client.post(
+        f"/api/exploration/sessions/{session_id}/close",
+        json={"experience": {"goal_achieved": True, "blockers": [], "misled": ["跳转误导"], "notes": "绕了三圈"}},
+    )
+    assert resp.status_code == 200, resp.text
+
+    resp = await client.get("/api/exploration/campaigns/c-1/sessions")
+    assert resp.status_code == 200, resp.text
+    sessions = resp.json()["sessions"]
+    assert len(sessions) == 1
+    entry = sessions[0]
+    assert entry["id"] == session_id
+    assert entry["status"] == "completed"
+    assert entry["turn_count"] == 1
+    assert entry["goal"] == "查询本月账单并完成缴费"
+    assert entry["persona"]["name"] == "急性子用户"
+    assert entry["experience"]["goal_achieved"] is True
+    assert entry["experience"]["misled"] == ["跳转误导"]
+    assert entry["closed_at"] is not None
+
+
+async def test_list_sessions_empty_campaign(seeded_db, client):
+    resp = await client.get("/api/exploration/campaigns/c-1/sessions")
+    assert resp.status_code == 200, resp.text
+    assert resp.json() == {"sessions": []}
+
+
+async def test_list_session_messages_returns_conversation(seeded_db, mock_channel, client):
+    session_id = (await _create_session(client)).json()["id"]
+    await client.post(f"/api/exploration/sessions/{session_id}/messages", json={"content": "我要查账单"})
+
+    resp = await client.get(f"/api/exploration/sessions/{session_id}/messages")
+    assert resp.status_code == 200, resp.text
+    messages = resp.json()["messages"]
+    assert len(messages) == 2
+    user_msg, agent_msg = messages
+    assert user_msg["role"] == "user"
+    assert user_msg["content"] == "我要查账单"
+    assert user_msg["round_index"] == 1
+    assert agent_msg["role"] == "assistant"
+    assert agent_msg["content"] == "您好，请问有什么可以帮您？"
+    assert agent_msg["round_index"] == 1
+    assert isinstance(agent_msg["latency_ms"], int)
+    assert user_msg["created_at"] is not None
+
+
+async def test_list_messages_unknown_session_returns_404(seeded_db, client):
+    resp = await client.get("/api/exploration/sessions/nope/messages")
+    assert resp.status_code == 404
+
+
 # ---------------------------------------------------------------- migration
 
 

tests/integration/test_exploration_settlement.py

@@ -0,0 +1,189 @@
+"""Integration tests for exploration settlement on campaign finalize (v0.9 票据 07).
+
+Covers both terminal transitions: window completion (durable loop COMPLETE)
+and mid-flight cancellation. Running sessions must turn expired and reject
+further messages (409), leaving no dangling sessions.
+"""
+
+import asyncio
+
+import pytest
+from agenteval.evaluation import campaign_runner as runner_module
+from agenteval.evaluation.campaign_runner import run_campaign_loop
+from agenteval.exploration.models import ExplorationSessionStatus
+from agenteval.models import (
+    Campaign,
+    CampaignPlanEntry,
+    Case,
+    CaseType,
+    ChannelType,
+    EvalTarget,
+    PlatformType,
+    Scenario,
+    TargetStatus,
+)
+from agenteval.storage.db import utc_now
+from agenteval.storage.repository import (
+    CampaignRepository,
+    ExplorationSessionRepository,
+    ScenarioRepository,
+    TargetRepository,
+)
+from agenteval.web.app import app
+from httpx import ASGITransport, AsyncClient
+
+
+def _session_payload(campaign_id: str) -> dict:
+    return {
+        "campaign_id": campaign_id,
+        "persona": {"name": "巡检用户", "traits": ["耐心"]},
+        "goal": "查询账单",
+        "triggered_by": "manual",
+    }
+
+
+@pytest.fixture()
+def seeded_db(db_session, monkeypatch):
+    """Point every get_session consumer at the test session and seed the
+    target + scenario needed by both the loop path and the cancel path."""
+    from agenteval.channels import factory as factory_module
+    from agenteval.evaluation import engine as engine_module
+    from agenteval.storage import db as db_module
+    from agenteval.storage import repository as repo_module
+    from agenteval.web import app as app_module
+    from agenteval.web.routers import exploration as exploration_router
+
+    monkeypatch.setattr(app_module, "init_db", lambda: None)
+
+    def _test_get_session():
+        return db_session
+
+    monkeypatch.setattr(db_module, "get_session", _test_get_session)
+    monkeypatch.setattr(repo_module, "get_session", _test_get_session)
+    monkeypatch.setattr(engine_module, "get_session", _test_get_session)
+    monkeypatch.setattr(runner_module, "get_session", _test_get_session)
+    monkeypatch.setattr(runner_module, "_auto_start_analysis", lambda campaign, session: None)
+
+    from tests.unit.mock_channel import MockChannel
+
+    channel = MockChannel(reply_delay=0.0)
+    monkeypatch.setattr(factory_module.ChannelFactory, "create", lambda target: channel)
+    monkeypatch.setattr(exploration_router, "start_judge_review", lambda session_id: None)
+
+    from agenteval.web.deps import get_db
+
+    def _test_get_db():
+        try:
+            yield db_session
+        finally:
+            pass
+
+    app.dependency_overrides[get_db] = _test_get_db
+
+    target = EvalTarget(
+        id="t-1",
+        name="mock-target",
+        platform=PlatformType.AI_DIGITAL_EMPLOYEE,
+        channel_type=ChannelType.TUTU_API,
+        channel_config={"base_url": "http://mock", "token": "x"},
+        status=TargetStatus.ACTIVE,
+    )
+    TargetRepository(db_session).create(target)
+    scenario = Scenario(
+        id="s-1",
+        name="mock-scenario",
+        cases=[Case(id="c1", type=CaseType.SINGLE, messages=["hi"])],
+    )
+    ScenarioRepository(db_session).create(scenario)
+
+    yield db_session
+    app.dependency_overrides.clear()
+
+
+def _make_campaign(db_session, campaign_id: str, *, time_scale: float) -> Campaign:
+    return CampaignRepository(db_session).create(
+        Campaign(
+            id=campaign_id,
+            name=f"campaign-{campaign_id}",
+            target_id="t-1",
+            window_seconds=3600,
+            time_scale=time_scale,
+            plan=[CampaignPlanEntry(scenario_id="s-1", offset_seconds=0, count=1)],
+            status="running",
+            started_at=utc_now(),
+        )
+    )
+
+
+async def _create_running_session(db_session, campaign_id: str) -> str:
+    async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as c:
+        resp = await c.post("/api/exploration/sessions", json=_session_payload(campaign_id))
+    assert resp.status_code == 200, resp.text
+    session_id = resp.json()["id"]
+    assert ExplorationSessionRepository(db_session).get(session_id).status == (ExplorationSessionStatus.RUNNING)
+    return session_id
+
+
+async def test_window_completion_expires_running_sessions(seeded_db):
+    campaign = _make_campaign(seeded_db, "c-done", time_scale=3600.0)
+    session_id = await _create_running_session(seeded_db, "c-done")
+
+    # time_scale 3600 → 1 real second covers the whole window; the loop
+    # completes on its first tick and settles exploration sessions.
+    cancel = asyncio.Event()
+    await asyncio.wait_for(run_campaign_loop(campaign.id, cancel, tick_seconds=0.05), timeout=5)
+
+    assert CampaignRepository(seeded_db).get("c-done").status.value == "completed"
+    settled = ExplorationSessionRepository(seeded_db).get(session_id)
+    assert settled.status == ExplorationSessionStatus.EXPIRED
+    assert settled.closed_at is not None
+
+
+async def test_cancel_expires_running_sessions_and_rejects_messages(seeded_db):
+    _make_campaign(seeded_db, "c-1", time_scale=1.0)
+    session_id = await _create_running_session(seeded_db, "c-1")
+
+    async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as client:
+        resp = await client.post("/api/campaigns/c-1/cancel")
+        assert resp.status_code == 200, resp.text
+        assert resp.json()["status"] == "cancelled"
+
+        settled = ExplorationSessionRepository(seeded_db).get(session_id)
+        assert settled.status == ExplorationSessionStatus.EXPIRED
+        assert settled.closed_at is not None
+
+        resp = await client.post(f"/api/exploration/sessions/{session_id}/messages", json={"content": "还在吗"})
+        assert resp.status_code == 409
+
+
+async def test_completed_session_survives_settlement(seeded_db):
+    _make_campaign(seeded_db, "c-1", time_scale=1.0)
+    async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as client:
+        resp = await client.post("/api/exploration/sessions", json=_session_payload("c-1"))
+        session_id = resp.json()["id"]
+        resp = await client.post(
+            f"/api/exploration/sessions/{session_id}/close",
+            json={"experience": {"goal_achieved": True}},
+        )
+        assert resp.status_code == 200
+
+        await client.post("/api/campaigns/c-1/cancel")
+
+    assert ExplorationSessionRepository(seeded_db).get(session_id).status == (ExplorationSessionStatus.COMPLETED)
+
+
+async def test_production_line_manual_trigger_also_settles(seeded_db):
+    _make_campaign(seeded_db, "c-1", time_scale=1.0)
+    async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as client:
+        resp = await client.post(
+            "/api/exploration/sessions",
+            json={**_session_payload("c-1"), "triggered_by": "manual"},
+        )
+        assert resp.status_code == 200, resp.text
+        session_id = resp.json()["id"]
+        assert resp.json()["triggered_by"] == "manual"
+
+        # 正式线手动会话同样纳入终态结算，触发来源不影响收口。
+        await client.post("/api/campaigns/c-1/cancel")
+
+    assert ExplorationSessionRepository(seeded_db).get(session_id).status == (ExplorationSessionStatus.EXPIRED)

tests/unit/test_campaign_analysis.py

@@ -152,6 +152,45 @@ async def test_scenario_without_failure_samples_still_gets_narrative():
     assert len(result["scenario_narratives"]) == 2
 
 
+async def test_analysis_input_includes_exploration_summary():
+    """v0.9 票据 05：阶段二输入追加探索摘要（问题清单 + 达成统计，非全量对话）。"""
+    exploration_summary = {
+        "session_count": 2,
+        "sessions_with_experience": 2,
+        "goal_achieved_count": 1,
+        "goal_achievement_rate": 0.5,
+        "issues": [{"issue": "缴费入口难找", "count": 2}],
+        "misled": [],
+        "judge_review": None,
+    }
+    client = FakeChatClient(STAGE1_A, STAGE1_B, STAGE2)
+    await analyze_campaign(
+        campaign=_campaign(),
+        report=_report(),
+        failure_samples={},
+        valid_run_ids=set(),
+        chat_client=client,
+        exploration_summary=exploration_summary,
+    )
+    stage2_prompt = json.dumps(client.calls[2], ensure_ascii=False)
+    assert "探索发现" in stage2_prompt
+    assert "缴费入口难找" in stage2_prompt
+    assert "goal_achievement_rate" in stage2_prompt or "0.5" in stage2_prompt
+
+
+async def test_analysis_without_exploration_omits_section():
+    client = FakeChatClient(STAGE1_A, STAGE1_B, STAGE2)
+    await analyze_campaign(
+        campaign=_campaign(),
+        report=_report(),
+        failure_samples={},
+        valid_run_ids=set(),
+        chat_client=client,
+    )
+    stage2_prompt = json.dumps(client.calls[2], ensure_ascii=False)
+    assert "探索发现" not in stage2_prompt
+
+
 # ── 分析模型解析 ─────────────────────────────────────────────────────────
 
 def _seed_config(session, config_id: str, *, analysis_default: bool = False, enabled: bool = True) -> None:

tests/unit/test_campaign_report_format.py

@@ -0,0 +1,235 @@
+"""Report aggregation logic & output format validation (报告聚合与格式校验).
+
+补充 test_campaign_report.py（数值口径）与 test_metrics.py（单点聚合）之外：
+- 聚合边界：用户取消不计分母（ADR-0004）、窗口外/窗口前运行钳制、
+  SQLite naive 时间戳与 aware 等价；
+- 格式校验：报告 dict 的键集合、类型、取值范围、时间序列连续性、
+  能力面排序与 ISO-UTC 时间戳格式。generate_campaign_report 是纯函数，
+  直接构造模型对象，不落库。
+"""
+
+from datetime import datetime, timedelta, timezone
+
+from agenteval.evaluation.metrics import aggregate_runs
+from agenteval.evaluation.report import generate_campaign_report
+from agenteval.models import Campaign, CampaignPlanEntry, CampaignStatus, EvalRun, RunStatus
+
+T0 = datetime(2026, 1, 1, 0, 0, 0, tzinfo=timezone.utc)
+
+
+def _campaign(*, window_seconds: int = 12, time_scale: float = 1.0, started_at=T0, completed_at=None) -> Campaign:
+    return Campaign(
+        id="camp-fmt",
+        name="格式校验活动",
+        target_id="t-1",
+        window_seconds=window_seconds,
+        time_scale=time_scale,
+        plan=[CampaignPlanEntry(scenario_id="s-a", offset_seconds=0, count=1)],
+        status=CampaignStatus.RUNNING,
+        started_at=started_at,
+        completed_at=completed_at,
+    )
+
+
+def _run(
+    scenario_id: str = "s-a",
+    *,
+    status: RunStatus = RunStatus.COMPLETED,
+    offset_seconds: float = 0.0,
+    pass_rate: float | None = None,
+    latency: float | None = None,
+    error: dict | None = None,
+    started_at: datetime | None = None,
+) -> EvalRun:
+    summary = None
+    if pass_rate is not None or latency is not None or error is not None:
+        summary = {}
+        if pass_rate is not None:
+            summary["pass_rate"] = pass_rate
+        if latency is not None:
+            summary["avg_latency_ms"] = latency
+        if error is not None:
+            summary["error"] = error
+    return EvalRun(
+        target_id="t-1",
+        scenario_id=scenario_id,
+        campaign_id="camp-fmt",
+        status=status,
+        started_at=started_at if started_at is not None else T0 + timedelta(seconds=offset_seconds),
+        summary=summary,
+    )
+
+
+def _report(runs, **kwargs):
+    return generate_campaign_report(_campaign(), runs, **kwargs)
+
+
+# ── 格式校验：顶层结构 ────────────────────────────────────────────────
+
+
+def test_top_level_schema_keys_and_types():
+    report = _report([_run(pass_rate=1.0, latency=100)])
+    assert set(report) == {
+        "campaign_id",
+        "name",
+        "target_id",
+        "status",
+        "window_seconds",
+        "time_scale",
+        "started_at",
+        "completed_at",
+        "summary",
+        "time_trend",
+        "capability_summary",
+    }
+    assert report["campaign_id"] == "camp-fmt"
+    assert isinstance(report["window_seconds"], int) and report["window_seconds"] > 0
+    assert isinstance(report["time_scale"], float)
+    assert report["status"] == CampaignStatus.RUNNING.value
+    assert isinstance(report["time_trend"], list)
+    assert isinstance(report["capability_summary"], list)
+
+
+def test_summary_schema_types_and_bounds():
+    report = _report(
+        [
+            _run(pass_rate=1.0, latency=100),
+            _run(scenario_id="s-b", status=RunStatus.FAILED),
+        ]
+    )
+    summary = report["summary"]
+    assert set(summary) == {
+        "total_runs",
+        "completed_runs",
+        "overall_pass_rate",
+        "overall_availability",
+        "avg_latency_ms",
+    }
+    assert isinstance(summary["total_runs"], int) and summary["total_runs"] == 2
+    assert isinstance(summary["completed_runs"], int) and summary["completed_runs"] == 1
+    for key in ("overall_pass_rate", "overall_availability"):
+        value = summary[key]
+        assert value is None or (isinstance(value, float) and 0.0 <= value <= 1.0), key
+    assert summary["avg_latency_ms"] is None or summary["avg_latency_ms"] >= 0
+
+
+def test_summary_matches_single_aggregation_seam():
+    """summary 数值必须与单点聚合口径 aggregate_runs 完全一致（ADR-0004）。"""
+    runs = [
+        _run(pass_rate=1.0, latency=100),
+        _run(pass_rate=0.5, latency=300),
+        _run(status=RunStatus.FAILED),
+    ]
+    report = _report(runs)
+    agg = aggregate_runs(runs)
+    summary = report["summary"]
+    assert summary["overall_pass_rate"] == agg["pass_rate"]
+    assert summary["overall_availability"] == agg["availability"]
+    assert summary["avg_latency_ms"] == agg["avg_latency_ms"]
+
+
+# ── 格式校验：时间趋势序列 ────────────────────────────────────────────
+
+
+def test_time_trend_contiguous_buckets_with_custom_count():
+    report = generate_campaign_report(_campaign(window_seconds=12), [_run(pass_rate=1.0)], bucket_count=4)
+    trend = report["time_trend"]
+    assert len(trend) == 4
+    for idx, bucket in enumerate(trend):
+        assert bucket["bucket_index"] == idx
+        assert bucket["start_seconds"] == idx * 3.0  # window / bucket_count
+        assert bucket["end_seconds"] == (idx + 1) * 3.0
+        assert {"run_count", "pass_rate", "availability", "avg_latency_ms"} <= set(bucket)
+    # 相邻桶首尾相接
+    for prev, cur in zip(trend, trend[1:]):
+        assert prev["end_seconds"] == cur["start_seconds"]
+
+
+def test_time_trend_bucket_types_and_bounds():
+    report = _report([_run(pass_rate=0.8, latency=120)])
+    for bucket in report["time_trend"]:
+        assert isinstance(bucket["run_count"], int) and bucket["run_count"] >= 0
+        if bucket["run_count"] == 0:
+            assert bucket["pass_rate"] is None and bucket["availability"] is None
+        else:
+            assert 0.0 <= bucket["pass_rate"] <= 1.0
+            assert 0.0 <= bucket["availability"] <= 1.0
+
+
+# ── 格式校验：能力面汇总 ──────────────────────────────────────────────
+
+
+def test_capability_summary_schema_and_sorted_by_run_count_desc():
+    runs = [
+        _run("s-a", pass_rate=1.0),
+        _run("s-a", pass_rate=0.5),
+        _run("s-a", pass_rate=0.0),
+        _run("s-b", pass_rate=1.0),
+    ]
+    report = _report(runs, scenario_names={"s-a": "夜间问诊"})
+    caps = report["capability_summary"]
+    assert [c["scenario_id"] for c in caps] == ["s-a", "s-b"]  # run_count 降序
+    for cap in caps:
+        assert {"scenario_id", "scenario_name", "run_count", "pass_rate", "availability", "avg_latency_ms"} <= set(cap)
+        assert cap["run_count"] >= 1
+    assert caps[0]["scenario_name"] == "夜间问诊"
+    assert caps[1]["scenario_name"] == "s-b"[:8]  # 无映射时回落 id 前缀
+
+
+# ── 格式校验：时间戳 ISO-UTC ──────────────────────────────────────────
+
+
+def test_timestamps_serialize_as_iso_utc_or_none():
+    report = generate_campaign_report(_campaign(completed_at=T0 + timedelta(hours=1)), [_run()])
+    assert report["started_at"].endswith("Z")
+    assert report["completed_at"].endswith("Z")
+    assert "T" in report["started_at"]
+
+    unfinished = generate_campaign_report(_campaign(), [_run()])
+    assert unfinished["completed_at"] is None
+
+
+def test_naive_campaign_timestamps_still_serialize_utc():
+    naive_start = datetime(2026, 1, 1, 0, 0, 0)  # SQLite round-trip drops tzinfo
+    report = generate_campaign_report(_campaign(started_at=naive_start), [_run()])
+    assert report["started_at"].endswith("Z")
+
+
+# ── 聚合边界：取消 / 钳制 / naive 时间 ────────────────────────────────
+
+
+def test_cancelled_run_excluded_from_summary_but_counted():
+    runs = [
+        _run(pass_rate=1.0, latency=100),
+        _run(status=RunStatus.FAILED, error={"code": "cancelled_by_user", "message": "stop"}),
+    ]
+    summary = _report(runs)["summary"]
+    assert summary["total_runs"] == 2  # 发生过的事仍然可见
+    assert summary["completed_runs"] == 1
+    assert summary["overall_pass_rate"] == 1.0  # 取消不进分母（ADR-0004）
+    assert summary["overall_availability"] == 1.0
+
+
+def test_run_started_before_window_clamped_to_first_bucket():
+    report = _report([_run(pass_rate=1.0, started_at=T0 - timedelta(seconds=10))])
+    assert report["time_trend"][0]["run_count"] == 1
+
+
+def test_run_started_after_window_clamped_to_last_bucket():
+    report = _report([_run(pass_rate=1.0, started_at=T0 + timedelta(seconds=100))])
+    assert report["time_trend"][-1]["run_count"] == 1
+    assert sum(b["run_count"] for b in report["time_trend"]) == 1
+
+
+def test_naive_run_timestamps_bucket_same_as_aware():
+    aware = _run(pass_rate=1.0, started_at=T0 + timedelta(seconds=6))
+    naive = _run(pass_rate=1.0, started_at=datetime(2026, 1, 1, 0, 0, 6))  # SQLite 读回的 naive UTC
+    trend_aware = _report([aware])["time_trend"]
+    trend_naive = _report([naive])["time_trend"]
+    assert trend_aware == trend_naive
+    assert trend_naive[6]["run_count"] == 1  # window 12s / 12 buckets → offset 6 落第 6 桶
+
+
+def test_unstarted_run_sits_at_offset_zero():
+    report = _report([_run(pass_rate=1.0, started_at=None)])
+    assert report["time_trend"][0]["run_count"] == 1

tests/unit/test_exploration_judge.py

@@ -0,0 +1,239 @@
+"""Judge sampling review for exploration sessions (v0.9 票据 04).
+
+会话关闭后，平台抽样对话（≤3 段）经 judge 岗位模型独立复核，质量维度结论
+结构化落入会话的 judge_review。复核是异步后台执行：失败落错误不阻塞会话
+状态；无模型配置时静默跳过。LLM 调用经可注入 ChatClient 接缝（沿 v0.7
+分析 seam），测试用假客户端覆盖（先例：test_campaign_analysis.py）。
+"""
+
+import json
+
+import pytest
+from agenteval.exploration.judge import (
+    MAX_JUDGE_SAMPLES,
+    execute_judge_review,
+    normalize_judge_review,
+    sample_round_indexes,
+)
+from agenteval.exploration.models import (
+    ExplorationMessage,
+    ExplorationSession,
+    ExplorationSessionStatus,
+)
+from agenteval.models import Campaign, CampaignPlanEntry
+from agenteval.storage.db import ModelConfigDB
+from agenteval.storage.model_config_repository import ModelConfigRepository
+from agenteval.storage.repository import (
+    CampaignRepository,
+    ExplorationMessageRepository,
+    ExplorationSessionRepository,
+)
+
+
+class FakeChatClient:
+    """Queued-response fake for the judge LLM seam."""
+
+    def __init__(self, *responses):
+        self._responses = list(responses)
+        self.calls: list[list[dict]] = []
+
+    async def __call__(self, messages: list[dict]) -> str:
+        self.calls.append(messages)
+        if not self._responses:
+            raise AssertionError("unexpected extra LLM call")
+        item = self._responses.pop(0)
+        if isinstance(item, Exception):
+            raise item
+        return item
+
+
+REVIEW_JSON = json.dumps(
+    {
+        "dimensions": [
+            {"dimension": "attitude", "rating": "good", "comment": "态度友好"},
+            {"dimension": "professionalism", "rating": "acceptable", "comment": "流程基本正确"},
+            {"dimension": "hallucination", "rating": "poor", "comment": "编造了不存在的政策"},
+        ],
+        "summary": "服务态度好但存在幻觉",
+    },
+    ensure_ascii=False,
+)
+
+
+def _campaign() -> Campaign:
+    return Campaign(
+        id="camp-1",
+        name="24h 正式线",
+        target_id="t-1",
+        window_seconds=86400,
+        time_scale=1.0,
+        plan=[CampaignPlanEntry(scenario_id="s-1", offset_seconds=0, count=1)],
+    )
+
+
+def _seed_config(db_session, config_id: str = "mc-default") -> None:
+    ModelConfigRepository(db_session).create(
+        ModelConfigDB(
+            id=config_id,
+            name=f"cfg-{config_id}",
+            provider="openai_compatible",
+            capability="chat",
+            endpoint_url="https://models.example.com/v1/chat/completions",
+            model_name="m",
+            enabled=True,
+            is_analysis_default=True,
+        )
+    )
+
+
+def _seed_session_with_messages(db_session, rounds: int = 5) -> str:
+    CampaignRepository(db_session).create(_campaign())
+    repo = ExplorationSessionRepository(db_session)
+    session_obj = repo.create(
+        ExplorationSession(
+            campaign_id="camp-1",
+            target_id="t-1",
+            persona={"name": "急性子用户"},
+            goal="查询账单并缴费",
+        )
+    )
+    session_obj.status = ExplorationSessionStatus.COMPLETED
+    repo.update(session_obj)
+
+    message_repo = ExplorationMessageRepository(db_session)
+    for i in range(1, rounds + 1):
+        message_repo.save_message(
+            ExplorationMessage(session_id=session_obj.id, round_index=i, role="user", content=f"用户消息-{i}")
+        )
+        message_repo.save_message(
+            ExplorationMessage(session_id=session_obj.id, round_index=i, role="assistant", content=f"回复内容-{i}")
+        )
+    return session_obj.id
+
+
+@pytest.fixture()
+def judge_env(db_session, monkeypatch):
+    from agenteval.exploration import judge as judge_module
+
+    monkeypatch.setattr(judge_module, "get_session", lambda: db_session)
+    return db_session
+
+
+# ── 抽样 ──────────────────────────────────────────────────────────────
+
+
+def test_sample_round_indexes_returns_all_within_cap():
+    assert sample_round_indexes([1, 2]) == [1, 2]
+    assert sample_round_indexes([1, 2, 3]) == [1, 2, 3]
+
+
+def test_sample_round_indexes_spreads_evenly_over_cap():
+    assert sample_round_indexes([1, 2, 3, 4, 5]) == [1, 3, 5]
+    assert len(sample_round_indexes(list(range(1, 11)))) == MAX_JUDGE_SAMPLES
+
+
+# ── 归一化 ────────────────────────────────────────────────────────────
+
+
+def test_normalize_judge_review_applies_whitelists():
+    raw = {
+        "dimensions": [
+            {"dimension": "attitude", "rating": "good", "comment": "ok"},
+            {"dimension": "神秘维度", "rating": "good", "comment": "drop me"},
+            {"dimension": "hallucination", "rating": "离谱", "comment": "bad rating"},
+        ],
+        "summary": "结论",
+    }
+    review = normalize_judge_review(raw)
+    dims = {d["dimension"]: d for d in review["dimensions"]}
+    assert set(dims) == {"attitude", "hallucination"}
+    assert dims["attitude"]["rating"] == "good"
+    assert dims["hallucination"]["rating"] == "acceptable"  # 非法档位归一
+    assert review["summary"] == "结论"
+
+
+def test_normalize_judge_review_rejects_non_list_dimensions():
+    review = normalize_judge_review({"dimensions": "不是列表", "summary": 123})
+    assert review["dimensions"] == []
+    assert review["summary"] == "123"
+
+
+# ── 后台执行编排 ──────────────────────────────────────────────────────
+
+
+async def test_execute_persists_structured_review(judge_env):
+    session_id = _seed_session_with_messages(judge_env, rounds=5)
+    _seed_config(judge_env)
+
+    client = FakeChatClient(REVIEW_JSON)
+    await execute_judge_review(session_id, chat_client=client)
+
+    session_obj = ExplorationSessionRepository(judge_env).get(session_id)
+    assert session_obj.status == ExplorationSessionStatus.COMPLETED
+    review = session_obj.judge_review
+    assert review["status"] == "completed"
+    assert review["model_config_id"] == "mc-default"
+    assert review["sampled_rounds"] == [1, 3, 5]
+    assert len(review["dimensions"]) == 3
+    assert review["summary"] == "服务态度好但存在幻觉"
+
+
+async def test_execute_prompt_carries_only_sampled_rounds(judge_env):
+    session_id = _seed_session_with_messages(judge_env, rounds=5)
+    _seed_config(judge_env)
+
+    client = FakeChatClient(REVIEW_JSON)
+    await execute_judge_review(session_id, chat_client=client)
+
+    payload = json.dumps(client.calls[0], ensure_ascii=False)
+    assert "用户消息-1" in payload and "用户消息-3" in payload and "用户消息-5" in payload
+    assert "用户消息-2" not in payload and "用户消息-4" not in payload
+
+
+async def test_execute_records_error_on_unparseable_output(judge_env):
+    session_id = _seed_session_with_messages(judge_env)
+    _seed_config(judge_env)
+
+    await execute_judge_review(session_id, chat_client=FakeChatClient("这不是 JSON"))
+
+    session_obj = ExplorationSessionRepository(judge_env).get(session_id)
+    assert session_obj.status == ExplorationSessionStatus.COMPLETED  # 复核失败不阻塞会话
+    review = session_obj.judge_review
+    assert review["status"] == "failed"
+    assert review["error"]
+    assert review["model_config_id"] == "mc-default"
+
+
+async def test_execute_records_error_on_client_exception(judge_env):
+    session_id = _seed_session_with_messages(judge_env)
+    _seed_config(judge_env)
+
+    await execute_judge_review(session_id, chat_client=FakeChatClient(RuntimeError("模型网关超时")))
+
+    review = ExplorationSessionRepository(judge_env).get(session_id).judge_review
+    assert review["status"] == "failed"
+    assert "模型网关超时" in review["error"]
+
+
+async def test_execute_skips_silently_without_model_config(judge_env):
+    session_id = _seed_session_with_messages(judge_env)
+
+    client = FakeChatClient(REVIEW_JSON)
+    await execute_judge_review(session_id, chat_client=client)
+
+    assert client.calls == []
+    assert ExplorationSessionRepository(judge_env).get(session_id).judge_review is None
+
+
+async def test_execute_skips_session_without_messages(judge_env):
+    CampaignRepository(judge_env).create(_campaign())
+    session_obj = ExplorationSessionRepository(judge_env).create(
+        ExplorationSession(campaign_id="camp-1", target_id="t-1", persona={}, goal="g")
+    )
+    _seed_config(judge_env)
+
+    client = FakeChatClient(REVIEW_JSON)
+    await execute_judge_review(session_obj.id, chat_client=client)
+
+    assert client.calls == []
+    assert ExplorationSessionRepository(judge_env).get(session_obj.id).judge_review is None

tests/unit/test_exploration_summary.py

@@ -0,0 +1,106 @@
+"""Exploration summary aggregation (v0.9 票据 05).
+
+体验记录聚合成探索摘要：会话数、目标达成率、问题清单（blockers/misled
+按出现次数降序）、judge 复核结论（已就位才纳入；findings 全量收各档发现，
+poor 档排前）。缺则无痕：无会话返回 None。纯函数，不落库。
+"""
+
+from agenteval.exploration.models import ExplorationSession, ExplorationSessionStatus
+from agenteval.exploration.summary import summarize_exploration
+
+
+def _session(experience=None, judge_review=None, status=ExplorationSessionStatus.COMPLETED) -> ExplorationSession:
+    return ExplorationSession(
+        campaign_id="c-1",
+        target_id="t-1",
+        persona={"name": "x"},
+        goal="查询账单",
+        status=status,
+        experience=experience,
+        judge_review=judge_review,
+    )
+
+
+def _experience(goal_achieved: bool, blockers=None, misled=None) -> dict:
+    return {
+        "goal_achieved": goal_achieved,
+        "blockers": blockers or [],
+        "misled": misled or [],
+        "emotion": "neutral",
+        "notes": "",
+    }
+
+
+def test_no_sessions_returns_none():
+    assert summarize_exploration([]) is None
+
+
+def test_aggregates_counts_rate_and_issue_lists():
+    sessions = [
+        _session(_experience(True, blockers=["缴费入口难找"])),
+        _session(_experience(False, blockers=["缴费入口难找", "验证码收不到"], misled=["被误导选了错误套餐"])),
+        _session(_experience(False, blockers=["验证码收不到"])),
+    ]
+    summary = summarize_exploration(sessions)
+    assert summary["session_count"] == 3
+    assert summary["sessions_with_experience"] == 3
+    assert summary["goal_achieved_count"] == 1
+    assert summary["goal_achievement_rate"] == round(1 / 3, 4)
+    # 问题清单按出现次数降序（同频按首次出现顺序）
+    assert summary["issues"] == [
+        {"issue": "缴费入口难找", "count": 2},
+        {"issue": "验证码收不到", "count": 2},
+    ]
+    assert summary["misled"] == [{"issue": "被误导选了错误套餐", "count": 1}]
+
+
+def test_sessions_without_experience_excluded_from_rate_denominator():
+    sessions = [
+        _session(_experience(True)),
+        _session(),  # 未关闭/无体验记录
+    ]
+    summary = summarize_exploration(sessions)
+    assert summary["session_count"] == 2
+    assert summary["sessions_with_experience"] == 1
+    assert summary["goal_achievement_rate"] == 1.0
+    assert summary["issues"] == []
+
+
+def test_no_experience_at_all_rate_is_none():
+    summary = summarize_exploration([_session(), _session()])
+    assert summary["goal_achievement_rate"] is None
+    assert summary["issues"] == [] and summary["misled"] == []
+
+
+def test_judge_review_included_only_when_completed():
+    good_review = {
+        "status": "completed",
+        "dimensions": [
+            {"dimension": "attitude", "rating": "good", "comment": "友好"},
+            {"dimension": "professionalism", "rating": "acceptable", "comment": "基本准确"},
+            {"dimension": "hallucination", "rating": "poor", "comment": "编造了政策"},
+        ],
+        "summary": "存在幻觉",
+    }
+    failed_review = {"status": "failed", "error": "解析失败"}
+    sessions = [
+        _session(_experience(True), judge_review=good_review),
+        _session(_experience(False), judge_review=failed_review),
+        _session(_experience(True)),  # 未复核
+    ]
+    summary = summarize_exploration(sessions)
+    judge = summary["judge_review"]
+    assert judge["reviewed_sessions"] == 1
+    # 复核发现全量纳入，poor 档排前（票据 05：复核结论一并纳入）
+    assert judge["findings"] == [
+        {"dimension": "hallucination", "rating": "poor", "comment": "编造了政策"},
+        {"dimension": "professionalism", "rating": "acceptable", "comment": "基本准确"},
+        {"dimension": "attitude", "rating": "good", "comment": "友好"},
+    ]
+    # 复核总体结论一并纳入
+    assert judge["summaries"] == ["存在幻觉"]
+
+
+def test_judge_review_absent_when_none_reviewed():
+    summary = summarize_exploration([_session(_experience(True))])
+    assert summary["judge_review"] is None

tests/unit/test_report_render.py

@@ -346,3 +346,70 @@ def test_render_campaign_markdown_comparison_comes_after_analysis():
 def test_render_campaign_markdown_without_comparison_unchanged():
     md = render_campaign_markdown(_campaign_report())
     assert "周期对比" not in md
+
+
+# ── render_campaign_markdown: 探索发现附录（v0.9 票据 05）────────────────────
+
+
+def _exploration() -> dict:
+    return {
+        "session_count": 3,
+        "sessions_with_experience": 3,
+        "goal_achieved_count": 1,
+        "goal_achievement_rate": 0.3333,
+        "issues": [
+            {"issue": "缴费入口难找", "count": 2},
+            {"issue": "验证码收不到", "count": 1},
+        ],
+        "misled": [{"issue": "被误导选了错误套餐", "count": 1}],
+        "judge_review": {
+            "reviewed_sessions": 1,
+            "findings": [{"dimension": "hallucination", "rating": "poor", "comment": "编造了不存在的政策"}],
+            "summaries": ["服务态度好但存在幻觉"],
+        },
+    }
+
+
+def test_render_campaign_markdown_appends_exploration_appendix():
+    md = render_campaign_markdown(_campaign_report(), exploration=_exploration())
+    assert "## 探索发现" in md
+    assert "| 探索会话数 | 3 |" in md
+    assert "| 目标达成率 | 33.3% |" in md
+    assert "缴费入口难找 ×2" in md
+    assert "验证码收不到 ×1" in md
+    assert "被误导选了错误套餐 ×1" in md
+    assert "编造了不存在的政策" in md
+    assert "复核结论：服务态度好但存在幻觉" in md
+
+
+def test_render_campaign_markdown_exploration_without_judge_or_issues():
+    md = render_campaign_markdown(
+        _campaign_report(),
+        exploration={
+            "session_count": 1,
+            "sessions_with_experience": 1,
+            "goal_achieved_count": 1,
+            "goal_achievement_rate": 1.0,
+            "issues": [],
+            "misled": [],
+            "judge_review": None,
+        },
+    )
+    assert "## 探索发现" in md
+    assert "无" in md  # 空问题清单回落文案
+
+
+def test_render_campaign_markdown_exploration_comes_after_comparison():
+    md = render_campaign_markdown(
+        _campaign_report(), analysis=_analysis(), comparison=_comparison(), exploration=_exploration()
+    )
+    assert md.index("## 智能分析") < md.index("## 周期对比") < md.index("## 探索发现")
+
+
+def test_render_campaign_markdown_without_exploration_byte_identical():
+    base = render_campaign_markdown(_campaign_report(), analysis=_analysis(), comparison=_comparison())
+    with_none = render_campaign_markdown(
+        _campaign_report(), analysis=_analysis(), comparison=_comparison(), exploration=None
+    )
+    assert base == with_none
+    assert "探索发现" not in base