v0.3-s3: Webhook + OpenClaw HTTP Skill + Markdown/对比报告

## Webhook 通知（S3-1）
- settings.py: 增加 AGENTEVAL_WEBHOOK_URL / AGENTEVAL_WEBHOOK_SECRET
- utils/webhook.py: send_run_webhook()，非阻断，任何异常仅 warning log
- runs.py: run 完成后自动触发 webhook（payload 含 run_id/status/summary/report_url）
- .env.example: 新增 webhook 配置示例

## OpenClaw Skill HTTP 改造（S3-2）
- plugins/openclaw/agenteval_skill.py: 完全重写
  - 改用 HTTP API（POST /api/runs + GET /api/runs/{id} 轮询 + GET /api/reports/{id}）
  - 移除 subprocess + CLI 依赖
  - 轮询等待至 completed/failed，支持配置 poll_interval / timeout
  - 返回结构化中文摘要（summary_text），直接可用于 OpenClaw 对话展示

## Markdown 报告导出（S3-3）
- report.py: render_markdown_report() — 完整的 Markdown 表格 + 对话展示
- save_report: 支持 fmt="markdown"，输出 .md 文件
- reports.py: GET /api/reports/{run_id}/markdown，Content-Disposition 附件下载
- api.ts: reportsApi.markdownUrl()
- Reports.tsx: 「导出 MD」按钮

## 对比报告（S3-4）
- report.py: generate_compare_report(run_id_1, run_id_2)
  - run_a / run_b 汇总 + delta（pass_rate / passed_cases / passed_rules）
  - case-level diff，标记 changed 用例
- reports.py: GET /api/reports/compare?run1=&run2=
- api.ts: reportsApi.compare()
- Reports.tsx: 完整对比视图
  - Segmented 切换「单次报告」/「对比报告」
  - 双 Select（报告 A vs B）+ 对比按钮
  - 汇总 delta card（pass_rate 变化 + 变化用例数徽章）
  - 用例对比表（通过/失败/改善↑/退步↓）+ 展开规则明细

Co-Authored-By: Claude <noreply@anthropic.com>

.env.example

@@ -27,3 +27,10 @@ AGENTEVAL_OPENCLAW_AUTH_TOKEN=change-me-in-production
 # ── Frontend ───────────────────────────────────────────────────
 # Optional override of the built frontend dist path (default: frontend/web/dist).
 # AGENTEVAL_FRONTEND_DIST_PATH=/app/frontend/web/dist
+
+# ── Webhook ────────────────────────────────────────────────────
+# If set, AgentEvalTool will POST run completion summaries to this URL.
+# Example: AGENTEVAL_WEBHOOK_URL=https://your-service.example.com/webhooks/agenteval
+AGENTEVAL_WEBHOOK_URL=
+# Optional shared secret included as X-Webhook-Secret header for verification.
+AGENTEVAL_WEBHOOK_SECRET=

backend/agenteval/config/settings.py

@@ -65,6 +65,16 @@ class Settings(BaseSettings):
         description="Comma-separated list of allowed file extensions for upload.",
     )
 
+    # ── Webhook ────────────────────────────────────────────────────
+    webhook_url: Optional[str] = Field(
+        default=None,
+        description="If set, POST run completion summaries to this URL.",
+    )
+    webhook_secret: Optional[str] = Field(
+        default=None,
+        description="If set, included as X-Webhook-Secret header for verification.",
+    )
+
     @model_validator(mode="after")
     def _derive_openclaw_ws_origin(self) -> "Settings":
         """Auto-derive openclaw_ws_origin from allowed_origins.

backend/agenteval/evaluation/report.py

@@ -168,6 +168,124 @@ def generate_report(run_id: str, session=None) -> dict[str, Any]:
     }
 
 
+def generate_compare_report(run_id_1: str, run_id_2: str, session=None) -> dict[str, Any]:
+    """Build a side-by-side comparison dict for two runs."""
+    report_a = generate_report(run_id_1, session)
+    report_b = generate_report(run_id_2, session)
+
+    def _summary_delta(key: str) -> float:
+        return report_b["summary"][key] - report_a["summary"][key]
+
+    # Case-level diff: match by case_id
+    cases_a = {c["case_id"]: c for c in report_a.get("cases", [])}
+    cases_b = {c["case_id"]: c for c in report_b.get("cases", [])}
+    all_case_ids = sorted(set(cases_a) | set(cases_b))
+
+    case_diffs = []
+    for cid in all_case_ids:
+        ca = cases_a.get(cid)
+        cb = cases_b.get(cid)
+
+        def _case_passed(c):
+            if not c:
+                return None
+            return all(r["passed"] for r in c.get("results", []))
+
+        case_diffs.append({
+            "case_id": cid,
+            "run_a_passed": _case_passed(ca),
+            "run_b_passed": _case_passed(cb),
+            "changed": _case_passed(ca) != _case_passed(cb),
+            "run_a_results": ca["results"] if ca else [],
+            "run_b_results": cb["results"] if cb else [],
+        })
+
+    return {
+        "run_a": {
+            "run_id": run_id_1,
+            "target_name": report_a.get("target_name"),
+            "scenario_name": report_a.get("scenario_name"),
+            "status": report_a.get("status"),
+            "started_at": report_a.get("started_at"),
+            "summary": report_a["summary"],
+        },
+        "run_b": {
+            "run_id": run_id_2,
+            "target_name": report_b.get("target_name"),
+            "scenario_name": report_b.get("scenario_name"),
+            "status": report_b.get("status"),
+            "started_at": report_b.get("started_at"),
+            "summary": report_b["summary"],
+        },
+        "delta": {
+            "pass_rate": round(_summary_delta("pass_rate"), 4),
+            "passed_cases": int(_summary_delta("passed_cases")),
+            "passed_rules": int(_summary_delta("passed_rules")),
+        },
+        "cases": case_diffs,
+        "changed_cases": sum(1 for c in case_diffs if c["changed"]),
+    }
+
+
+def render_markdown_report(run_id: str, session=None) -> str:
+    """Render a report as Markdown string."""
+    report = generate_report(run_id, session)
+    s = report["summary"]
+    lines: list[str] = [
+        f"# 评测报告 — {report.get('scenario_name', run_id)}",
+        "",
+        f"**评测对象**: {report.get('target_name', '-')}  ",
+        f"**评测场景**: {report.get('scenario_name', '-')}  ",
+        f"**状态**: {report.get('status', '-')}  ",
+        f"**开始时间**: {report.get('started_at', '-')}  ",
+        f"**完成时间**: {report.get('completed_at', '-')}  ",
+        "",
+        "## 汇总",
+        "",
+        "| 指标 | 数值 |",
+        "|------|------|",
+        f"| 总用例数 | {s['total_cases']} |",
+        f"| 通过用例 | {s['passed_cases']} |",
+        f"| 失败用例 | {s['failed_cases']} |",
+        f"| 总规则数 | {s['total_rules']} |",
+        f"| 通过规则 | {s['passed_rules']} |",
+        f"| 通过率 | {s['pass_rate'] * 100:.1f}% |",
+        "",
+        "## 用例明细",
+        "",
+    ]
+
+    for case in report.get("cases", []):
+        case_passed = all(r["passed"] for r in case.get("results", []))
+        badge = "✅" if case_passed else "❌"
+        lines.append(f"### {badge} 用例 `{case['case_id']}`")
+        lines.append("")
+
+        for turn in case.get("turns", []):
+            lines.append(f"**第 {turn['round']} 轮**")
+            lines.append("")
+            lines.append(f"> **用户**: {turn.get('sent_text', '—')}")
+            lines.append("")
+            reply = turn.get("reply_text") or "（无回复）"
+            lines.append(f"> **智能体**: {reply}")
+            if turn.get("latency_ms") is not None:
+                lines.append(f"> *延迟: {turn['latency_ms']}ms*")
+            lines.append("")
+
+        if case.get("results"):
+            lines.append("**规则评估结果**")
+            lines.append("")
+            lines.append("| 规则 | 结果 | 评分 | 说明 |")
+            lines.append("|------|------|------|------|")
+            for r in case["results"]:
+                badge = "✅" if r["passed"] else "❌"
+                score = f"{r['score']:.2f}" if r.get("score") is not None else "-"
+                lines.append(f"| {r['rule_type']} | {badge} | {score} | {r.get('reason', '')} |")
+            lines.append("")
+
+    return "\n".join(lines)
+
+
 def render_json_report(run_id: str, session=None) -> str:
     """Render a report as JSON string."""
     report = generate_report(run_id, session)
@@ -193,6 +311,9 @@ def save_report(run_id: str, fmt: str = "html", output_dir: Optional[Path] = Non
     elif fmt == "json":
         content = render_json_report(run_id)
         path = output_dir / f"report_{run_id}_{timestamp}.json"
+    elif fmt == "markdown":
+        content = render_markdown_report(run_id)
+        path = output_dir / f"report_{run_id}_{timestamp}.md"
     else:
         raise ValueError(f"unsupported report format: {fmt}")
 

backend/agenteval/utils/webhook.py

@@ -0,0 +1,40 @@
+"""Webhook notification for run completion events."""
+
+import logging
+from typing import Any
+
+import httpx
+
+from agenteval.config import get_settings
+
+logger = logging.getLogger(__name__)
+
+
+async def send_run_webhook(run_id: str, status: str, summary: dict[str, Any]) -> None:
+    """POST run completion payload to the configured webhook URL.
+
+    Silently logs and returns on any error — webhook failure must never
+    affect the run itself.
+    """
+    settings = get_settings()
+    if not settings.webhook_url:
+        return
+
+    payload = {
+        "event": "run_completed",
+        "run_id": run_id,
+        "status": status,
+        "summary": summary,
+        "report_url": f"/api/reports/{run_id}",
+    }
+
+    headers = {"Content-Type": "application/json"}
+    if settings.webhook_secret:
+        headers["X-Webhook-Secret"] = settings.webhook_secret
+
+    try:
+        async with httpx.AsyncClient(timeout=10) as client:
+            resp = await client.post(settings.webhook_url, json=payload, headers=headers)
+        logger.info("webhook sent: run=%s status=%s http=%d", run_id, status, resp.status_code)
+    except Exception as exc:
+        logger.warning("webhook failed (non-fatal): run=%s error=%s", run_id, exc)

backend/agenteval/web/routers/reports.py

@@ -1,15 +1,35 @@
 """API routes for evaluation reports."""
 
-from fastapi import APIRouter, Depends, HTTPException, Response
+from fastapi import APIRouter, Depends, HTTPException, Query, Response
 from sqlmodel import Session
 
-from agenteval.evaluation.report import generate_report, render_html_report, render_json_report
+from agenteval.evaluation.report import (
+    generate_compare_report,
+    generate_report,
+    render_html_report,
+    render_json_report,
+    render_markdown_report,
+)
 from agenteval.storage.repository import RunRepository
 from agenteval.web.deps import get_db
 
 router = APIRouter()
 
 
+@router.get("/compare")
+def get_compare_report(
+    run1: str = Query(..., description="First run ID"),
+    run2: str = Query(..., description="Second run ID"),
+    session: Session = Depends(get_db),
+) -> dict:
+    repo = RunRepository(session)
+    if not repo.get(run1):
+        raise HTTPException(status_code=404, detail=f"run not found: {run1}")
+    if not repo.get(run2):
+        raise HTTPException(status_code=404, detail=f"run not found: {run2}")
+    return generate_compare_report(run1, run2, session)
+
+
 @router.get("/{run_id}")
 def get_report(run_id: str, session: Session = Depends(get_db)) -> dict:
     run = RunRepository(session).get(run_id)
@@ -34,3 +54,16 @@ def get_json_report(run_id: str, session: Session = Depends(get_db)) -> Response
         raise HTTPException(status_code=404, detail="run not found")
     json_text = render_json_report(run_id, session)
     return Response(content=json_text, media_type="application/json")
+
+
+@router.get("/{run_id}/markdown")
+def get_markdown_report(run_id: str, session: Session = Depends(get_db)) -> Response:
+    run = RunRepository(session).get(run_id)
+    if not run:
+        raise HTTPException(status_code=404, detail="run not found")
+    md = render_markdown_report(run_id, session)
+    return Response(
+        content=md,
+        media_type="text/markdown; charset=utf-8",
+        headers={"Content-Disposition": f'attachment; filename="report-{run_id}.md"'},
+    )

backend/agenteval/web/routers/runs.py

@@ -12,6 +12,7 @@ from agenteval.models import EvalRun, RunStatus
 from agenteval.storage.db import get_session
 from agenteval.storage.repository import RunRepository, ScenarioRepository, TargetRepository
 from agenteval.utils.llm import extract_reply_text
+from agenteval.utils.webhook import send_run_webhook
 from agenteval.web.deps import get_db
 from agenteval.web.websocket import ws_manager
 
@@ -52,6 +53,14 @@ async def _run_evaluation(run_id: str, target_id: str, scenario_id: str) -> None
             progress_callback=lambda event, data: ws_manager.emit(run_id, event, data),
             existing_run=existing_run,
         )
+        # Fire webhook after run completes (non-blocking, best-effort)
+        completed_run = RunRepository(session).get(run_id)
+        if completed_run:
+            await send_run_webhook(
+                run_id=run_id,
+                status=completed_run.status.value,
+                summary=completed_run.summary or {},
+            )
     finally:
         session.close()
         _cancel_tokens.pop(run_id, None)

backend/plugins/openclaw/agenteval_skill.py

@@ -1,96 +1,116 @@
-"""
+"""OpenClaw Skill for AgentEvalTool — HTTP API mode.
-OpenClaw Skill example for AgentEvalTool.
 
-This skill demonstrates how OpenClaw can call the AgentEvalTool CLI
+Triggers an evaluation run via the AgentEvalTool REST API, polls until
-to execute an evaluation run and fetch the report.
+completion, and returns a structured Chinese-language summary suitable for
+display in an OpenClaw conversation.
 
-In OpenClaw, register this as a skill and configure the following parameters:
+Configuration (passed via OpenClaw skill params):
-    - target_id: ID of the registered evaluation target
+    api_base_url    Base URL of AgentEvalTool (e.g. http://192.168.8.145:8001)
-    - scenario_id: ID of the evaluation scenario
+    api_key         X-API-Key header value (if auth is enabled, else omit)
-    - report_format: "json" or "html" (default "json")
+    target_id       ID of the registered evaluation target
-
+    scenario_id     ID of the evaluation scenario
-The skill assumes that the `agenteval` CLI is available on the system PATH.
+    poll_interval   Seconds between status polls (default: 3)
+    timeout         Max seconds to wait for completion (default: 300)
 """
 
-import json
+import time
-import subprocess
 from typing import Any
 
+import httpx
+
 
 class AgentEvalSkill:
-    """OpenClaw skill wrapper for AgentEvalTool."""
+    """OpenClaw skill: trigger evaluation and return a structured summary."""
 
     def __init__(self, config: dict[str, Any]):
-        self.config = config
+        self.api_base = config["api_base_url"].rstrip("/")
+        self.api_key = config.get("api_key")
+        self.target_id = config["target_id"]
+        self.scenario_id = config["scenario_id"]
+        self.poll_interval = float(config.get("poll_interval", 3))
+        self.timeout = float(config.get("timeout", 300))
+
+    def _headers(self) -> dict[str, str]:
+        h = {"Content-Type": "application/json"}
+        if self.api_key:
+            h["X-API-Key"] = self.api_key
+        return h
 
     def run(self) -> dict[str, Any]:
-        target_id = self.config["target_id"]
+        with httpx.Client(base_url=self.api_base, headers=self._headers(), timeout=30) as client:
-        scenario_id = self.config["scenario_id"]
+            # 1. Start the run
-        report_format = self.config.get("report_format", "json")
+            resp = client.post("/api/runs", json={
+                "target_id": self.target_id,
+                "scenario_id": self.scenario_id,
+            })
+            if resp.status_code != 200:
+                return {"ok": False, "error": f"启动评测失败: HTTP {resp.status_code} {resp.text}"}
 
-        # 1. Trigger evaluation run
+            run = resp.json()
-        run_cmd = [
+            run_id: str = run["id"]
-            "agenteval",
-            "run",
-            "start",
-            "--target-id",
-            target_id,
-            "--scenario-id",
-            scenario_id,
-        ]
-        run_result = subprocess.run(run_cmd, capture_output=True, text=True, check=False)
-        if run_result.returncode != 0:
-            return {
-                "ok": False,
-                "error": f"evaluation run failed: {run_result.stderr}",
-                "stdout": run_result.stdout,
-            }
 
-        # Extract run_id from CLI output (last line contains "run_id=xxx")
+            # 2. Poll until completion or timeout
-        run_id = None
+            deadline = time.time() + self.timeout
-        for line in reversed(run_result.stdout.strip().splitlines()):
+            while time.time() < deadline:
-            if "run_id=" in line:
+                time.sleep(self.poll_interval)
-                run_id = line.split("run_id=")[-1].strip().split()[0]
+                status_resp = client.get(f"/api/runs/{run_id}")
+                if status_resp.status_code != 200:
+                    continue
+                run = status_resp.json()
+                if run.get("status") in ("completed", "failed"):
                     break
-
+            else:
-        if not run_id:
                 return {
                     "ok": False,
-                "error": "could not extract run_id from CLI output",
-                "stdout": run_result.stdout,
-            }
-
-        # 2. Fetch report
-        report_cmd = [
-            "agenteval",
-            "report",
-            "show",
-            run_id,
-            "--format",
-            report_format,
-        ]
-        report_result = subprocess.run(report_cmd, capture_output=True, text=True, check=False)
-        if report_result.returncode != 0:
-            return {
-                "ok": False,
-                "error": f"report fetch failed: {report_result.stderr}",
                     "run_id": run_id,
+                    "error": f"评测超时（>{self.timeout}s 未完成）",
                 }
 
-        report_data = report_result.stdout
+            # 3. Fetch structured report
-        if report_format == "json":
+            report_resp = client.get(f"/api/reports/{run_id}")
-            try:
+            report = report_resp.json() if report_resp.status_code == 200 else {}
-                report_data = json.loads(report_data)
+
-            except json.JSONDecodeError:
+            return self._format_result(run, report)
-                pass
+
+    def _format_result(self, run: dict[str, Any], report: dict[str, Any]) -> dict[str, Any]:
+        status = run.get("status", "unknown")
+        summary = run.get("summary") or {}
+        pass_rate = summary.get("pass_rate", 0.0)
+        passed_cases = summary.get("passed_cases", 0)
+        total_cases = summary.get("total_cases", 0)
+        passed_rules = summary.get("passed_rules", 0)
+        total_rules = summary.get("total_rules", 0)
+
+        # Build Chinese natural-language summary for display in chat
+        if status == "completed":
+            verdict = "✅ 评测完成" if pass_rate >= 0.8 else "⚠️ 评测完成（有失败项）"
+            summary_text = (
+                f"{verdict}\n"
+                f"• 评测对象：{report.get('target_name', self.target_id)}\n"
+                f"• 评测场景：{report.get('scenario_name', self.scenario_id)}\n"
+                f"• 通过率：{pass_rate * 100:.1f}%\n"
+                f"• 用例：{passed_cases}/{total_cases} 通过\n"
+                f"• 规则：{passed_rules}/{total_rules} 通过\n"
+                f"• 报告链接：{self.api_base}/api/reports/{run['id']}"
+            )
+        else:
+            error_info = summary.get("error", {})
+            if isinstance(error_info, dict):
+                error_msg = error_info.get("message", str(error_info))
+            else:
+                error_msg = str(error_info)
+            summary_text = f"❌ 评测失败：{error_msg}"
 
         return {
-            "ok": True,
+            "ok": status == "completed",
-            "run_id": run_id,
+            "run_id": run.get("id"),
-            "report": report_data,
+            "status": status,
+            "pass_rate": pass_rate,
+            "summary_text": summary_text,
+            "report": report,
         }
 
 
-# Example entrypoint for OpenClaw runtime.
 def execute(config: dict[str, Any]) -> dict[str, Any]:
+    """OpenClaw skill entrypoint."""
     return AgentEvalSkill(config).run()

frontend/web/src/api.ts

@@ -138,6 +138,9 @@ export const reportsApi = {
   get: (runId: string) => api.get(`/reports/${runId}`),
   html: (runId: string) => api.get(`/reports/${runId}/html`),
   json: (runId: string) => api.get(`/reports/${runId}/json`),
+  markdownUrl: (runId: string) => `/api/reports/${runId}/markdown`,
+  compare: (run1: string, run2: string) =>
+    api.get(`/reports/compare`, { params: { run1, run2 } }),
 }
 
 export const statsApi = {

frontend/web/src/pages/Reports.tsx

@@ -1,12 +1,13 @@
 import { useEffect, useState } from 'react'
 import { useSearchParams } from 'react-router-dom'
 import {
-  Button, Card, Col, Collapse, Descriptions, Empty, Row,
+  Button, Card, Col, Collapse, Descriptions, Empty, Row, Segmented,
-  Select, Space, Spin, Statistic, Table, Tag, Tooltip,
+  Select, Space, Spin, Statistic, Table, Tag, Tooltip, Badge,
 } from 'antd'
 import {
   CheckCircleOutlined, CloseCircleOutlined,
   DownloadOutlined, UserOutlined, RobotOutlined, ReloadOutlined,
+  DiffOutlined, FileMarkdownOutlined,
 } from '@ant-design/icons'
 import { reportsApi, runsApi, type Run } from '../api'
 import { colors } from '../tokens'
@@ -32,6 +33,7 @@ interface CaseReport {
 }
 
 interface Report {
+  run_id: string
   target_name: string
   scenario_name: string
   status: string
@@ -48,17 +50,38 @@ interface Report {
   cases: CaseReport[]
 }
 
+interface CompareResult {
+  run_a: { run_id: string; target_name: string; scenario_name: string; status: string; started_at: string; summary: Report['summary'] }
+  run_b: { run_id: string; target_name: string; scenario_name: string; status: string; started_at: string; summary: Report['summary'] }
+  delta: { pass_rate: number; passed_cases: number; passed_rules: number }
+  cases: Array<{
+    case_id: string
+    run_a_passed: boolean | null
+    run_b_passed: boolean | null
+    changed: boolean
+    run_a_results: RuleResultData[]
+    run_b_results: RuleResultData[]
+  }>
+  changed_cases: number
+}
+
+type ViewMode = 'single' | 'compare'
+
 export default function ReportsPage() {
   const [searchParams, setSearchParams] = useSearchParams()
   const runQuery = searchParams.get('run') ?? ''
 
   const [runs, setRuns] = useState<Run[]>([])
   const [selectedRunId, setSelectedRunId] = useState<string>('')
+  const [compareRunId, setCompareRunId] = useState<string>('')
   const [report, setReport] = useState<Report | null>(null)
+  const [compareResult, setCompareResult] = useState<CompareResult | null>(null)
   const [loading, setLoading] = useState(false)
+  const [viewMode, setViewMode] = useState<ViewMode>('single')
 
   const loadReport = async (runId: string) => {
     setSelectedRunId(runId)
+    setCompareResult(null)
     setLoading(true)
     try {
       const res = await reportsApi.get(runId)
@@ -93,10 +116,21 @@ export default function ReportsPage() {
     loadReport(runId)
   }
 
-  const handleExportHtml = () => {
+  const handleCompare = async () => {
-    if (!selectedRunId) return
+    if (!selectedRunId || !compareRunId) return
-    window.open(`/api/reports/${selectedRunId}/html`, '_blank')
+    setLoading(true)
+    try {
+      const res = await reportsApi.compare(selectedRunId, compareRunId)
+      setCompareResult(res.data as CompareResult)
+    } finally {
+      setLoading(false)
     }
+  }
+
+  const runSelectOptions = runs.map((r) => ({
+    value: r.id,
+    label: `${r.id.slice(0, 8)}… | ${r.started_at ? formatDateTime(r.started_at) : ''}`,
+  }))
 
   return (
     <div style={{ height: '100%', display: 'flex', flexDirection: 'column', overflow: 'hidden' }}>
@@ -115,26 +149,63 @@ export default function ReportsPage() {
         padding: '8px 16px', flexShrink: 0,
         borderBottom: `1px solid ${colors.border}`,
         background: colors.bgSubtle,
-        display: 'flex', alignItems: 'center', gap: 12,
+        display: 'flex', alignItems: 'center', gap: 12, flexWrap: 'wrap',
       }}>
+        <Segmented
+          value={viewMode}
+          onChange={(v) => { setViewMode(v as ViewMode); setCompareResult(null) }}
+          options={[
+            { label: '单次报告', value: 'single' },
+            { label: '对比报告', value: 'compare', icon: <DiffOutlined /> },
+          ]}
+        />
+
         <Select
-          style={{ width: 400 }}
+          style={{ width: 340 }}
-          placeholder="选择已完成的评测记录"
+          placeholder={viewMode === 'compare' ? '选择报告 A' : '选择已完成的评测记录'}
           value={selectedRunId || undefined}
           onChange={handleView}
           showSearch
           filterOption={(input, opt) =>
             (opt?.label as string ?? '').toLowerCase().includes(input.toLowerCase())}
-          options={runs.map((r) => ({
+          options={runSelectOptions}
-            value: r.id,
-            label: `${r.id.slice(0, 8)}... | ${r.started_at ? formatDateTime(r.started_at) : ''}`,
-          }))}
         />
-        {report && (
+
-          <Button icon={<DownloadOutlined />} onClick={handleExportHtml}>
+        {viewMode === 'compare' && (
+          <>
+            <span style={{ color: colors.textMuted, fontSize: 12 }}>vs</span>
+            <Select
+              style={{ width: 340 }}
+              placeholder="选择报告 B"
+              value={compareRunId || undefined}
+              onChange={setCompareRunId}
+              showSearch
+              filterOption={(input, opt) =>
+                (opt?.label as string ?? '').toLowerCase().includes(input.toLowerCase())}
+              options={runSelectOptions.filter((o) => o.value !== selectedRunId)}
+            />
+            <Button
+              type="primary"
+              icon={<DiffOutlined />}
+              disabled={!selectedRunId || !compareRunId}
+              onClick={handleCompare}
+            >
+              对比
+            </Button>
+          </>
+        )}
+
+        {viewMode === 'single' && report && (
+          <Space>
+            <Button icon={<DownloadOutlined />} onClick={() => window.open(`/api/reports/${selectedRunId}/html`, '_blank')}>
               导出 HTML
             </Button>
+            <Button icon={<FileMarkdownOutlined />} onClick={() => window.open(reportsApi.markdownUrl(selectedRunId), '_blank')}>
+              导出 MD
+            </Button>
+          </Space>
         )}
+
         <div style={{ marginLeft: 'auto' }}>
           <Tooltip title="刷新列表">
             <Button size="middle" icon={<ReloadOutlined />} onClick={loadRuns} />
@@ -142,10 +213,31 @@ export default function ReportsPage() {
         </div>
       </div>
 
-      {/* 报告内容 — 可滚动 */}
+      {/* 报告内容区 */}
       <div style={{ flex: 1, minHeight: 0, overflowY: 'auto', padding: '16px' }}>
         <Spin spinning={loading}>
-          {report ? (
+          {viewMode === 'compare'
+            ? <CompareView result={compareResult} />
+            : <SingleReportView report={report} />
+          }
+        </Spin>
+      </div>
+    </div>
+  )
+}
+
+// ── Single Report View ───────────────────────────────────────────────────
+
+function SingleReportView({ report }: { report: Report | null }) {
+  if (!report) {
+    return (
+      <div style={{ display: 'flex', alignItems: 'center', justifyContent: 'center', minHeight: 200 }}>
+        <Empty description="请选择一个评测记录查看报告" />
+      </div>
+    )
+  }
+
+  return (
     <>
       <Row gutter={[16, 16]} style={{ marginBottom: 16 }}>
         <Col xs={24} sm={12} md={6}>
@@ -153,33 +245,20 @@ export default function ReportsPage() {
         </Col>
         <Col xs={24} sm={12} md={6}>
           <Card>
-                    <Statistic
+            <Statistic title="通过用例" value={report.summary.passed_cases}
-                      title="通过用例"
+              valueStyle={{ color: '#3f8600' }} prefix={<CheckCircleOutlined />} />
-                      value={report.summary.passed_cases}
-                      valueStyle={{ color: '#3f8600' }}
-                      prefix={<CheckCircleOutlined />}
-                    />
           </Card>
         </Col>
         <Col xs={24} sm={12} md={6}>
           <Card>
-                    <Statistic
+            <Statistic title="失败用例" value={report.summary.failed_cases}
-                      title="失败用例"
+              valueStyle={{ color: '#cf1322' }} prefix={<CloseCircleOutlined />} />
-                      value={report.summary.failed_cases}
-                      valueStyle={{ color: '#cf1322' }}
-                      prefix={<CloseCircleOutlined />}
-                    />
           </Card>
         </Col>
         <Col xs={24} sm={12} md={6}>
           <Card>
-                    <Statistic
+            <Statistic title="通过率" value={report.summary.pass_rate * 100} precision={1} suffix="%"
-                      title="通过率"
+              valueStyle={{ color: report.summary.pass_rate >= 0.8 ? '#3f8600' : '#cf1322' }} />
-                      value={report.summary.pass_rate * 100}
-                      precision={1}
-                      suffix="%"
-                      valueStyle={{ color: report.summary.pass_rate >= 0.8 ? '#3f8600' : '#cf1322' }}
-                    />
           </Card>
         </Col>
       </Row>
@@ -188,31 +267,32 @@ export default function ReportsPage() {
         <Descriptions size="small" column={2}>
           <Descriptions.Item label="评测对象">{report.target_name}</Descriptions.Item>
           <Descriptions.Item label="评测场景">{report.scenario_name}</Descriptions.Item>
-                  <Descriptions.Item label="开始时间">
+          <Descriptions.Item label="开始时间">{formatDateTime(report.started_at)}</Descriptions.Item>
-                    {report.started_at ? formatDateTime(report.started_at) : '-'}
+          <Descriptions.Item label="完成时间">{report.completed_at ? formatDateTime(report.completed_at) : '-'}</Descriptions.Item>
-                  </Descriptions.Item>
-                  <Descriptions.Item label="完成时间">
-                    {report.completed_at ? formatDateTime(report.completed_at) : '-'}
-                  </Descriptions.Item>
         </Descriptions>
       </Card>
 
       <Card title="用例明细">
-                <Collapse
+        <Collapse items={report.cases.map((c) => ({
-                  items={report.cases.map((c) => ({
           key: c.case_id,
           label: (
             <Space>
               <span style={{ fontWeight: 500 }}>{c.case_id}</span>
-                        {c.results.every((r) => r.passed) ? (
+              {c.results.every((r) => r.passed)
-                          <Tag color="success">全部通过</Tag>
+                ? <Tag color="success">全部通过</Tag>
-                        ) : (
+                : <Tag color="error">存在失败</Tag>}
-                          <Tag color="error">存在失败</Tag>
-                        )}
               <Tag>{c.turns.length} 轮对话</Tag>
             </Space>
           ),
-                    children: (
+          children: <CaseDetail c={c} />,
+        }))} />
+      </Card>
+    </>
+  )
+}
+
+function CaseDetail({ c }: { c: CaseReport }) {
+  return (
     <div>
       <div style={{ marginBottom: 16 }}>
         {c.turns.map((turn, idx) => (
@@ -236,46 +316,139 @@ export default function ReportsPage() {
                 <RobotOutlined style={{ marginRight: 6, color: '#52c41a' }} />
                 {turn.reply_text || '（无回复）'}
                 {turn.latency_ms != null && (
-                                    <span style={{ color: '#999', fontSize: 11, marginLeft: 8 }}>
+                  <span style={{ color: '#999', fontSize: 11, marginLeft: 8 }}>{turn.latency_ms}ms</span>
-                                      {turn.latency_ms}ms
-                                    </span>
                 )}
               </div>
             </div>
           </div>
         ))}
       </div>
-                        <Table
+      <Table size="small" pagination={false} dataSource={c.results} rowKey={(_, idx) => String(idx)}
-                          size="small"
-                          pagination={false}
-                          dataSource={c.results}
-                          rowKey={(_, idx) => String(idx)}
         columns={[
-                            { title: '规则', dataIndex: 'rule_type', width: 150 },
+          { title: '规则', dataIndex: 'rule_type', width: 160 },
           { title: '结果', dataIndex: 'passed', width: 80,
-                              render: (p: boolean) => p
+            render: (p: boolean) => p ? <Tag color="success">通过</Tag> : <Tag color="error">失败</Tag> },
-                                ? <Tag color="success">通过</Tag>
-                                : <Tag color="error">失败</Tag>,
-                            },
           { title: '评分', dataIndex: 'score', width: 80,
-                              render: (s: number | null) => s != null ? s.toFixed(2) : '-',
+            render: (s: number | null) => s != null ? s.toFixed(2) : '-' },
-                            },
           { title: '说明', dataIndex: 'reason' },
         ]}
       />
     </div>
+  )
+}
+
+// ── Compare View ─────────────────────────────────────────────────────────
+
+function CompareView({ result }: { result: CompareResult | null }) {
+  if (!result) {
+    return (
+      <div style={{ display: 'flex', alignItems: 'center', justifyContent: 'center', minHeight: 200 }}>
+        <Empty description="选择两次评测记录后点击「对比」" />
+      </div>
+    )
+  }
+
+  const { run_a, run_b, delta, cases, changed_cases } = result
+  const deltaColor = (v: number) => v > 0 ? '#3f8600' : v < 0 ? '#cf1322' : colors.textMuted
+  const deltaSign = (v: number) => v > 0 ? `+${v}` : String(v)
+
+  return (
+    <>
+      {/* 汇总对比 */}
+      <Row gutter={16} style={{ marginBottom: 16 }}>
+        <Col span={11}>
+          <Card title={<span style={{ color: '#1677ff' }}>报告 A — {run_a.run_id.slice(0, 8)}…</span>} size="small">
+            <Descriptions size="small" column={1}>
+              <Descriptions.Item label="场景">{run_a.scenario_name}</Descriptions.Item>
+              <Descriptions.Item label="时间">{formatDateTime(run_a.started_at)}</Descriptions.Item>
+              <Descriptions.Item label="通过率">{(run_a.summary.pass_rate * 100).toFixed(1)}%</Descriptions.Item>
+              <Descriptions.Item label="用例">{run_a.summary.passed_cases}/{run_a.summary.total_cases}</Descriptions.Item>
+            </Descriptions>
+          </Card>
+        </Col>
+        <Col span={2} style={{ display: 'flex', alignItems: 'center', justifyContent: 'center', flexDirection: 'column', gap: 8 }}>
+          <div style={{ fontSize: 11, color: colors.textMuted }}>变化</div>
+          <div style={{ fontWeight: 700, color: deltaColor(delta.pass_rate), fontSize: 16 }}>
+            {deltaSign(Math.round(delta.pass_rate * 1000) / 10)}%
+          </div>
+          <Badge count={changed_cases} color={changed_cases > 0 ? 'orange' : 'green'}
+            title={`${changed_cases} 个用例结果变化`} />
+        </Col>
+        <Col span={11}>
+          <Card title={<span style={{ color: '#52c41a' }}>报告 B — {run_b.run_id.slice(0, 8)}…</span>} size="small">
+            <Descriptions size="small" column={1}>
+              <Descriptions.Item label="场景">{run_b.scenario_name}</Descriptions.Item>
+              <Descriptions.Item label="时间">{formatDateTime(run_b.started_at)}</Descriptions.Item>
+              <Descriptions.Item label="通过率">{(run_b.summary.pass_rate * 100).toFixed(1)}%</Descriptions.Item>
+              <Descriptions.Item label="用例">{run_b.summary.passed_cases}/{run_b.summary.total_cases}</Descriptions.Item>
+            </Descriptions>
+          </Card>
+        </Col>
+      </Row>
+
+      {/* 用例对比表 */}
+      <Card title={`用例对比（${changed_cases} 个结果变化）`}>
+        <Table
+          size="small"
+          pagination={false}
+          dataSource={cases}
+          rowKey="case_id"
+          rowClassName={(r) => r.changed ? 'run-row' : ''}
+          columns={[
+            { title: '用例', dataIndex: 'case_id', width: 180,
+              render: (id: string, r) => (
+                <Space>
+                  {r.changed && <Badge dot color="orange" />}
+                  <span style={{ fontWeight: r.changed ? 600 : 400 }}>{id}</span>
+                </Space>
               ),
-                  }))}
+            },
+            { title: '报告 A', dataIndex: 'run_a_passed', width: 100,
+              render: (p: boolean | null) =>
+                p === null ? <Tag>无</Tag>
+                  : p ? <Tag color="success">通过</Tag>
+                  : <Tag color="error">失败</Tag>,
+            },
+            { title: '报告 B', dataIndex: 'run_b_passed', width: 100,
+              render: (p: boolean | null) =>
+                p === null ? <Tag>无</Tag>
+                  : p ? <Tag color="success">通过</Tag>
+                  : <Tag color="error">失败</Tag>,
+            },
+            { title: '变化', width: 80,
+              render: (_: any, r) => {
+                if (!r.changed) return <span style={{ color: colors.textMuted }}>—</span>
+                if (r.run_b_passed && !r.run_a_passed) return <Tag color="success">改善 ↑</Tag>
+                if (!r.run_b_passed && r.run_a_passed) return <Tag color="error">退步 ↓</Tag>
+                return <Tag color="orange">变化</Tag>
+              },
+            },
+          ]}
+          expandable={{
+            expandedRowRender: (r) => (
+              <Row gutter={16}>
+                <Col span={12}>
+                  <div style={{ fontSize: 12, color: colors.textMuted, marginBottom: 4 }}>报告 A 规则</div>
+                  {r.run_a_results.map((res, i) => (
+                    <div key={i} style={{ fontSize: 12, marginBottom: 2 }}>
+                      {res.passed ? '✅' : '❌'} {res.rule_type}: {res.reason}
+                    </div>
+                  ))}
+                </Col>
+                <Col span={12}>
+                  <div style={{ fontSize: 12, color: colors.textMuted, marginBottom: 4 }}>报告 B 规则</div>
+                  {r.run_b_results.map((res, i) => (
+                    <div key={i} style={{ fontSize: 12, marginBottom: 2 }}>
+                      {res.passed ? '✅' : '❌'} {res.rule_type}: {res.reason}
+                    </div>
+                  ))}
+                </Col>
+              </Row>
+            ),
+            rowExpandable: (r) => r.run_a_results.length > 0 || r.run_b_results.length > 0,
+          }}
         />
       </Card>
     </>
-          ) : (
-            <div style={{ display: 'flex', alignItems: 'center', justifyContent: 'center', minHeight: 200 }}>
-              <Empty description="请选择一个评测记录查看报告" />
-            </div>
-          )}
-        </Spin>
-      </div>
-    </div>
   )
 }