@hupan56/wlkj 2.7.12 → 3.1.2

package/templates/qoder/scripts/validation/eval/qwork_harness.py

@@ -0,0 +1,617 @@
+"""
+qwork_harness.py — QoderWork 命令自动化测试框架 (P2)
+
+走完整个流程: 注入命令 → 轮询 transcript → 遇到 AskUserQuestion 自动代答
+            → 等 turn.finished → 调 transcript_timing 出耗时报告
+
+设计要点:
+  - 真相源是 transcript jsonl (实时追加), 不靠猜 DOM
+  - AskUserQuestion 的 question+options 全在 tool.requested.data.args 里
+  - 代答策略: 关键词匹配 option.label/description, 选最"全"的那个
+    (如 "整体/context/全貌/平台/默认" 等), 也可 --auto-answer=1 指定
+  - 代答回传: 优先点击渲染的选项 DOM; 兜底注入 label 文本 + Enter
+
+用法:
+  python qwork_harness.py "/wl-search 考勤"
+  python qwork_harness.py "/wl-search 考勤" --auto-answer=0   # 选第1个
+  python qwork_harness.py "/wl-search 考勤" --no-answer        # 只跑不等答
+  python qwork_harness.py "/wl-prd-full 用户管理" --auto-answer=1
+
+依赖: websocket-client (pip install websocket-client)
+"""
+import json
+import os
+import re
+import sys
+import time
+import glob
+import urllib.request
+
+try:
+    import websocket
+except ImportError:
+    print("[FATAL] 缺 websocket-client, 请: pip install websocket-client")
+    sys.exit(2)
+
+# 复用 P1 计时引擎
+_HERE = os.path.dirname(os.path.abspath(__file__))
+sys.path.insert(0, _HERE)
+import transcript_timing as tt  # noqa: E402
+
+# ── CDP 连接 ────────────────────────────────────────────────────────
+CDP_PORT = 6529
+
+
+def _get_qwork_ws(port=CDP_PORT):
+    d = json.loads(urllib.request.urlopen(f"http://127.0.0.1:{port}/json", timeout=3).read())
+    for t in d:
+        if t.get("type") == "page" and "QoderWork" in t.get("title", ""):
+            return t["webSocketDebuggerUrl"]
+    raise RuntimeError("没找到 QoderWork 页面 (CDP 6529). QoderWork 开着吗?")
+
+
+class CDP:
+    def __init__(self, ws_url):
+        self.ws = websocket.create_connection(ws_url, timeout=15)
+        self.mid = 0
+
+    def call(self, method, params=None, timeout=15):
+        self.mid += 1
+        self.ws.send(json.dumps({"id": self.mid, "method": method, "params": params or {}}))
+        end = time.time() + timeout
+        while time.time() < end:
+            r = json.loads(self.ws.recv())
+            if r.get("id") == self.mid:
+                return r
+        return {}
+
+    def eval(self, expr, timeout=15):
+        r = self.call("Runtime.evaluate", {"expression": expr, "returnByValue": True}, timeout)
+        body = r.get("result", {})
+        if "exceptionDetails" in body:
+            return {"ERR": str(body["exceptionDetails"])[:200]}
+        res = body.get("result", {})
+        return res.get("value", res)
+
+    def close(self):
+        try:
+            self.ws.close()
+        except Exception:
+            pass
+
+
+# ── 输入框操作 ──────────────────────────────────────────────────────
+def focus_input(cdp):
+    return cdp.eval(r"""(() => {
+      const e = document.querySelector('[contenteditable="true"]');
+      if(!e) return 'NO_INPUT';
+      e.focus();
+      const sel=window.getSelection(); const range=document.createRange();
+      range.selectNodeContents(e); range.collapse(false);
+      sel.removeAllRanges(); sel.addRange(range);
+      return 'ok';
+    })()""")
+
+
+def inject_text(cdp, text):
+    """CDP Input.insertText — 模拟真实键盘输入, 框架必识别。"""
+    cdp.call("Input.insertText", {"text": text}, timeout=8)
+    time.sleep(0.5)
+
+
+def submit_enter(cdp):
+    """CDP 派发 Enter 键提交。"""
+    for et in ("rawKeyDown", "keyDown", "keyUp"):
+        cdp.call("Input.dispatchKeyEvent", {
+            "type": et, "key": "Enter", "code": "Enter", "windowsVirtualKeyCode": 13,
+        }, timeout=5)
+        time.sleep(0.05)
+
+
+def get_input_text(cdp):
+    return cdp.eval(r"""document.querySelector('[contenteditable="true"]')?.textContent || ''""")
+
+
+# ── transcript 实时读取 ─────────────────────────────────────────────
+def _find_session_jsonl():
+    """找当前活跃 session 最新的 jsonl (取 mtime 最新且较大的)。"""
+    base = os.path.expanduser(r"~/.qoderwork/logs/sessions")
+    if not os.path.isdir(base):
+        return None
+    allf = []
+    for p in glob.glob(os.path.join(base, "**", "*.jsonl"), recursive=True):
+        try:
+            allf.append((os.path.getmtime(p), os.path.getsize(p), p))
+        except OSError:
+            continue
+    if not allf:
+        return None
+    allf.sort(reverse=True)
+    # 取最新且 >5KB 的(活跃会话)
+    for mt, sz, p in allf:
+        if sz > 5000:
+            return p
+    return allf[0][2]
+
+
+# ── 系统/后台 turn 识别 ─────────────────────────────────────────────
+# QoderWork 每条用户命令跑完后, 会自动追加一条 [SYSTEM: automated background
+# review] 后台审核 turn。它的 prompt 文本含这些标记。harness 必须跳过它们,
+# 否则 baseline+1 计数会被后台 turn 干扰, 提前误判"命令完成"。
+_SYSTEM_MARKERS = (
+    "automated background review",
+    "must be fulfilled using",
+    "[system:",
+    "this request must be fulfilled",
+)
+
+
+def _is_system_prompt(text):
+    """判断 prompt 文本是否为系统/后台任务(非用户命令)。"""
+    if not text:
+        return False
+    low = text.lower()
+    return any(m in low for m in _SYSTEM_MARKERS)
+
+
+def _ui_is_busy(cdp):
+    """检测 QoderWork 是否正忙(有任务在跑或有队列)。
+    双重判定: ① UI 状态文字 ② transcript 的 turn 配对(started>finished)。
+    返回 (busy:bool, detail:str)。
+    """
+    # ① UI 状态文字
+    res = cdp.eval(r"""(() => {
+      const txts = [...document.querySelectorAll('div,span')]
+        .filter(e => e.offsetParent !== null && e.children.length === 0)
+        .map(e => (e.textContent || '').trim())
+        .filter(t => t && t.length < 30);
+      const joined = txts.join('|');
+      const busy = /处理中|执行中|running/i.test(joined);
+      const queue = (txts.find(t => /\d+\s*个队列任务/.test(t))) || '';
+      return {busy, queue};
+    })()""")
+    ui_busy = False
+    detail = ""
+    if isinstance(res, dict):
+        ui_busy = bool(res.get("busy") or res.get("queue"))
+        detail = res.get("queue", "") or ("处理中" if res.get("busy") else "")
+    # ② transcript turn 配对(更可靠的兜底)
+    jsonl = _find_session_jsonl()
+    if jsonl:
+        ev = read_events(jsonl)
+        ns = sum(1 for e in ev if e.get("type") == "turn.started")
+        nf = sum(1 for e in ev if e.get("type") == "turn.finished")
+        if ns > nf:
+            ui_busy = True
+            detail = detail or f"turn未配对(started={ns}>finished={nf})"
+    return ui_busy, detail
+
+
+def read_events(path):
+    """读 jsonl 全部事件。"""
+    if not path or not os.path.exists(path):
+        return []
+    events = []
+    with open(path, "r", encoding="utf-8") as f:
+        for line in f:
+            line = line.strip()
+            if not line:
+                continue
+            try:
+                events.append(json.loads(line))
+            except json.JSONDecodeError:
+                continue
+    return events
+
+
+def _ts_of(event):
+    return tt._parse_ts(event.get("ts"))
+
+
+# ── AskUserQuestion 代答逻辑 ────────────────────────────────────────
+# AskUserQuestion 在 transcript 里的真实生命周期 (经实测 seq 验证):
+#   tool.requested      (tool_name=="AskUserQuestion", args.questions=完整问题)
+#   permission.requested(同上, UI 此时渲染选项卡)  ← 在此代答(点 DOM)
+#   permission.resolved (updated_args.answers 已填) ← 表示用户/AI 已答
+#   tool.execution.finished
+# 一个 AskUserQuestion 可含多个 question, 每个都要答; multiSelect 可多选。
+ANSWER_TOOL = "AskUserQuestion"
+
+# 关键词优先级: 出现这些词的 option 优先选 (信息最全/最安全)
+PREFER_KEYWORDS = [
+    "整体", "context", "全貌", "全部", "概览",  # 搜索类:选最全的
+    "web", "管理端", "pc", 1,  # 平台: 默认 web
+    "默认", "推荐", "recommended",
+    "是", "确认", "继续", "ok", "yes",
+]
+AVOID_KEYWORDS = ["取消", "否", "不", "cancel", "no", "停止", "abort"]
+
+
+def pick_answer_option(question_obj, force_index=None):
+    """从 AskUserQuestion 的单个 question 里选一个 option。
+    question_obj = {question, options:[{label,description}]}
+    返回 (index, label)。
+    """
+    opts = question_obj.get("options", [])
+    if not opts:
+        return None, None
+    if force_index is not None and 0 <= force_index < len(opts):
+        return force_index, opts[force_index].get("label", "")
+    # 关键词打分
+    best_i, best_score = 0, -1
+    for i, o in enumerate(opts):
+        text = (o.get("label", "") + " " + o.get("description", "")).lower()
+        score = 0
+        for kw in PREFER_KEYWORDS:
+            if str(kw).lower() in text:
+                score += 2
+        for kw in AVOID_KEYWORDS:
+            if str(kw).lower() in text:
+                score -= 3
+        # 默认偏向第一个(通常是推荐项)
+        if i == 0:
+            score += 1
+        if score > best_score:
+            best_score = score
+            best_i = i
+    return best_i, opts[best_i].get("label", "")
+
+
+def find_unanswered_questions(events, baseline_turns, already_seqs):
+    """从 transcript 事件里找出"已渲染但未答"的 AskUserQuestion。
+
+    判定规则(实测 seq 验证):
+      - 定位 permission.requested (tool_name==AskUserQuestion): UI 此时已渲染选项
+      - 同 turn 窗口内若已有 permission.resolved(updated_args.answers 非空) → 已答, 跳过
+      - 用 seq 去重; 已在 already_seqs 里的不再返回
+
+    Returns: [{seq, questions:[...]}] 按 seq 升序
+    """
+    # 1. 收集 AskUserQuestion 的 permission.requested (seq, questions)
+    asks = []
+    for e in events:
+        d = e.get("data", {}) or {}
+        if e.get("type") != "permission.requested":
+            continue
+        if d.get("tool_name") != ANSWER_TOOL:
+            continue
+        seq = e.get("seq")
+        if seq is None or seq in already_seqs:
+            continue
+        questions = (d.get("args", {}) or {}).get("questions", [])
+        if questions:
+            asks.append((seq, e.get("ts"), questions))
+
+    if not asks:
+        return []
+
+    # 2. 收集已 resolved 的 (用 updated_args.answers 非空作为"已答"信号)
+    resolved_seqs = set()
+    for e in events:
+        d = e.get("data", {}) or {}
+        if e.get("type") != "permission.resolved":
+            continue
+        if d.get("tool_name") != ANSWER_TOOL:
+            continue
+        upd = d.get("updated_args", {}) or {}
+        if upd.get("answers"):  # 有答案 = 真正答了
+            resolved_seqs.add(e.get("seq"))
+
+    # permission.resolved 的 seq 比 permission.requested 大 1~2;
+    # 用 questions 文本指纹匹配更稳: 已 resolved 的 question 文本集合
+    answered_qtexts = set()
+    for e in events:
+        if e.get("type") != "permission.resolved":
+            continue
+        d = e.get("data", {}) or {}
+        if d.get("tool_name") != ANSWER_TOOL:
+            continue
+        upd = d.get("updated_args", {}) or {}
+        for k in (upd.get("answers") or {}).keys():
+            answered_qtexts.add(k.strip())
+
+    out = []
+    for seq, ts, questions in asks:
+        # 本 AskUserQuestion 的所有 question 是否都已被答(resolved)
+        qtexts = {(q.get("question") or "").strip() for q in questions}
+        if qtexts and qtexts <= answered_qtexts:
+            continue
+        out.append({"seq": seq, "ts": ts, "questions": questions})
+    return out
+
+
+def answer_via_dom(cdp, question_obj, chosen_label):
+    """策略A: 点击 DOM 里渲染的选项卡 (按 label 文字匹配)。
+    QoderWork 选项卡结构: <BUTTON class="w-full flex..."> 内含序号+label+description。
+    必须点 BUTTON 本身(点内层 span/div 不触发 onClick), 且优先精确匹配。
+    返回 True 如果成功点到。"""
+    js = r"""((label) => {
+      const target = label.replace(/\s+/g,' ').trim().toLowerCase();
+      // 第1轮: 优先 <button>, 文本含 label 即可(文本带序号前缀如"2新系统...")
+      const buttons = [...document.querySelectorAll('button')].filter(b => b.offsetParent !== null);
+      for(const el of buttons){
+        const txt = (el.textContent||'').replace(/\s+/g,' ').trim();
+        if(txt.length > 200) continue;          // 排除消息体大按钮
+        if(txt.toLowerCase().includes(target)){
+          el.scrollIntoView({block:'center'});
+          el.click();
+          return {clicked:true, via:'button', txt:txt.slice(0,50)};
+        }
+      }
+      // 第2轮: 退而求其次, 任何 pointer 元素 (div/span) 文本精确等于 label
+      const cands = [...document.querySelectorAll('div,span,[role="button"],[role="option"]')]
+        .filter(e => e.offsetParent !== null);
+      for(const el of cands){
+        const cs = getComputedStyle(el);
+        if(cs.cursor !== 'pointer') continue;
+        const txt = (el.textContent||'').replace(/\s+/g,' ').trim();
+        if(!txt || txt.length > 80) continue;
+        if(txt.toLowerCase() === target){
+          el.scrollIntoView({block:'center'});
+          el.click();
+          return {clicked:true, via:'pointer', txt:txt.slice(0,50)};
+        }
+      }
+      return {clicked:false};
+    })(""" + json.dumps(chosen_label) + r""")"""
+    res = cdp.eval(js)
+    return isinstance(res, dict) and res.get("clicked")
+
+
+def answer_via_text(cdp, chosen_label, question_obj):
+    """策略B(兜底): 把 option.label 当回复文本注入 + Enter。"""
+    focus_input(cdp)
+    time.sleep(0.3)
+    # 清空已有内容
+    cdp.eval(r"""(() => {
+      const e=document.querySelector('[contenteditable="true"]');
+      if(e){ e.focus(); e.textContent=''; }
+      return true;
+    })()""")
+    time.sleep(0.2)
+    inject_text(cdp, chosen_label)
+    time.sleep(0.3)
+    submit_enter(cdp)
+    return True
+
+
+# ── 主流程 ──────────────────────────────────────────────────────────
+def run(command, force_answer=None, timeout=600, no_answer=False, verbose=True):
+    """跑一条命令到完成。
+    Args:
+        command: "/wl-search 考勤"
+        force_answer: int|None, 强制选项序号(0-based)
+        timeout: 最大等待秒数
+        no_answer: True 则不自动代答(遇到提问就停, 返回待答信息)
+        verbose: 打印过程
+    Returns dict: {status, report, turns}
+    """
+    def log(*a):
+        if verbose:
+            print(*a)
+
+    log(f"[harness] 连接 CDP...")
+    cdp = CDP(_get_qwork_ws())
+    log("[harness] 已连上 QoderWork")
+
+    # 命令清洗: Git bash(MSYS)会把 /wl-xxx 误转成 Windows 路径
+    # 如 "/wl-search 考勤" → "D:/.../Git/wl-search 考勤" (斜杠被吃, Git/wl-search)
+    # 检测: 命令不以 /wl- 开头, 但内部有 wl- 词 → 截到 wl- 前补斜杠。
+    stripped = command.lstrip()
+    if not stripped.startswith("/wl-") and "wl-" in stripped:
+        m = re.search(r"(^|[/\\])(wl-[a-z][a-z-]*)", stripped)
+        if m:
+            # 保留 wl-xxx 及之后的所有参数(参数在第一个空格后, 未被污染)
+            head = "/" + m.group(2)
+            # 找原始命令第一个空格后的参数部分(参数含中文, 不会被 MSYS 转)
+            sp = stripped.find(" ")
+            tail = stripped[sp:] if sp > 0 else ""
+            clean = head + tail
+            log(f"[harness] 命令清洗(MSYS修复): {command!r} → {clean!r}")
+            command = clean
+
+    # 0. 注入前等待 QoderWork 空闲(防队列堆积)
+    wait_start = time.time()
+    while time.time() - wait_start < 120:
+        busy, detail = _ui_is_busy(cdp)
+        if not busy:
+            break
+        if int(time.time() - wait_start) % 10 < 2:
+            log(f"[harness] 等待 QoderWork 空闲... ({detail})")
+        time.sleep(3)
+    else:
+        log("[harness] ⚠ QoderWork 120s 仍忙碌, 强行注入(可能进队列)")
+    log("[harness] QoderWork 空闲, 开始注入")
+
+    # 1. 注入命令
+    log(f"[harness] 注入命令: {command}")
+    focus_input(cdp)
+    time.sleep(0.3)
+    inject_text(cdp, command)
+    entered = get_input_text(cdp)
+    if command.strip().split()[0] not in str(entered):
+        log(f"[harness][ERR] 注入失败, 输入框={entered!r}")
+        cdp.close()
+        return {"status": "inject_failed", "input": entered}
+    log("[harness] 注入成功, 提交...")
+    submit_enter(cdp)
+
+    start_time = time.time()
+
+    # 注入后重新锁定 session: 注入可能创建新 jsonl, 必须跟那个文件走。
+    # 等 2s 让 QoderWork 落盘, 再取当前最新 session 作为本轮监控目标。
+    time.sleep(2)
+    monitor_jsonl = _find_session_jsonl()
+    baseline_events = read_events(monitor_jsonl) if monitor_jsonl else []
+    # 记录注入前的 prompt 总数, 用于识别"我这条命令的 prompt"
+    baseline_user_prompts = sum(
+        1 for e in baseline_events
+        if e.get("type") == "input.prompt.submitted"
+        and not _is_system_prompt((e.get("data", {}) or {}).get("text_preview", ""))
+    )
+    log(f"[harness] 锁定 session: {os.path.basename(monitor_jsonl or '')} | 基线user-prompt={baseline_user_prompts}")
+
+    answered_questions = []
+    new_turn_finished = False
+    my_prompt_ts = None   # 我这条命令的 prompt.submitted 时间戳
+    answered_req_seqs = set()  # 已代答/已答的 tool.requested seq
+
+    # 2. 轮询 transcript
+    while time.time() - start_time < timeout:
+        time.sleep(2)
+        elapsed = int(time.time() - start_time)
+
+        # ── session 切换保护 ─────────────────────────────
+        # QoderWork 后台记忆任务可能开新 session jsonl, 导致锁定文件不再更新。
+        # 检测到更新的 session 文件就切过去跟踪。
+        cur_latest = _find_session_jsonl()
+        if cur_latest and cur_latest != monitor_jsonl:
+            try:
+                mt_locked = os.path.getmtime(monitor_jsonl) if monitor_jsonl else 0
+                mt_latest = os.path.getmtime(cur_latest)
+                if mt_latest > mt_locked + 5:  # 新文件明显更新
+                    log(f"[harness][{elapsed}s] 检测到 session 切换: "
+                        f"{os.path.basename(monitor_jsonl or '')[:16]} → {os.path.basename(cur_latest)[:16]}")
+                    monitor_jsonl = cur_latest
+                    base_ev = read_events(monitor_jsonl)
+                    baseline_user_prompts = sum(
+                        1 for e in base_ev
+                        if e.get("type") == "input.prompt.submitted"
+                        and not _is_system_prompt((e.get("data", {}) or {}).get("text_preview", "")))
+                    my_prompt_ts = None  # 重新锁定我的命令
+            except OSError:
+                pass
+
+        jsonl = monitor_jsonl
+        if not jsonl or not os.path.exists(jsonl):
+            jsonl = _find_session_jsonl()
+        if not jsonl:
+            continue
+        events = read_events(jsonl)
+
+        # ── 检测 AskUserQuestion (未答的) ─────────────────
+        if not no_answer:
+            pending = find_unanswered_questions(events, baseline_user_prompts, answered_req_seqs)
+            for info in pending:
+                seq = info["seq"]
+                if seq in answered_req_seqs:
+                    continue
+                q0 = info["questions"][0]
+                qtxt = q0.get("question", "")
+                # 同问题文本不重复代答
+                if any(aq.get("question") == qtxt for aq in answered_questions):
+                    answered_req_seqs.add(seq)
+                    continue
+
+                idx, label = pick_answer_option(q0, force_answer)
+                log(f"[harness][{elapsed}s] AI提问(seq={seq}): {qtxt}")
+                log(f"[harness][{elapsed}s]   选项: {[o.get('label','') for o in q0.get('options',[])]}")
+                log(f"[harness][{elapsed}s]   代答: #{idx} {label!r}")
+
+                # 策略A: 点 DOM 选项
+                ok = answer_via_dom(cdp, q0, label)
+                if not ok:
+                    log(f"[harness][{elapsed}s]   DOM点击失败, 兜底注入文本...")
+                    answer_via_text(cdp, label, q0)
+                answered_questions.append({"question": qtxt, "answer_index": idx,
+                                           "answer_label": label})
+                answered_req_seqs.add(seq)
+                time.sleep(2)
+
+        # ── 定位"我这条命令"的 prompt.submitted ──────────
+        if my_prompt_ts is None:
+            user_prompts = [(e.get("ts"), (e.get("data", {}) or {}).get("text_preview", ""))
+                            for e in events
+                            if e.get("type") == "input.prompt.submitted"
+                            and not _is_system_prompt((e.get("data", {}) or {}).get("text_preview", ""))]
+            if len(user_prompts) > baseline_user_prompts:
+                # 取第 (baseline+1) 条用户 prompt = 我注入的
+                my_prompt_ts = user_prompts[baseline_user_prompts][0]
+                log(f"[harness][{elapsed}s] 锁定我的命令 prompt: {my_prompt_ts}")
+
+        # ── 检测 turn.finished (只认我的命令对应的那个) ──
+        if my_prompt_ts is not None:
+            # 找 my_prompt_ts 之后的 turn.finished (跳过系统后台 turn)
+            # 系统 turn 的特征: 它前面紧跟的 prompt 是系统类
+            for e in events:
+                if e.get("type") != "turn.finished":
+                    continue
+                if my_prompt_ts and e.get("ts", "") < my_prompt_ts:
+                    continue  # 早于我的命令, 忽略
+                # 判断这个 turn.finished 是否归属"用户命令"(非系统后台):
+                # 取本 turn 窗口内最近的 prompt.submitted, 若是系统类则跳过
+                prompt_before = None
+                for pe in events:
+                    if pe.get("type") != "input.prompt.submitted":
+                        continue
+                    if pe.get("ts", "") > e.get("ts", ""):
+                        break
+                    prompt_before = pe
+                if prompt_before and _is_system_prompt(
+                        (prompt_before.get("data", {}) or {}).get("text_preview", "")):
+                    continue  # 这是后台 review turn, 跳过
+                # 归属我的命令 → 完成
+                new_turn_finished = True
+                dur = (e.get("data", {}) or {}).get("duration_ms")
+                log(f"[harness][{elapsed}s] ✓ 我的命令 turn.finished! (dur={dur}ms)")
+                break
+
+        if new_turn_finished:
+            break  # 命令已完成, 退出轮询
+
+        # 状态提示 (每 ~6s 一次, 不刷屏)
+        cur_turns = sum(1 for e in events if e.get("type") == "turn.finished")
+        if elapsed % 6 < 2:
+            log(f"[harness][{elapsed}s] 运行中... (events={len(events)}, "
+                f"total_turns={cur_turns}, 已答提问={len(answered_questions)})")
+
+    cdp.close()
+
+    if not new_turn_finished:
+        log(f"[harness] ⚠ 超时 {timeout}s, 未等到 turn.finished")
+        return {"status": "timeout", "timeout_s": timeout,
+                "answered": answered_questions, "jsonl": monitor_jsonl}
+
+    # 3. 算耗时报告
+    jsonl = monitor_jsonl or _find_session_jsonl()
+    log(f"[harness] 出报告: {os.path.basename(jsonl)}")
+    report = tt.analyze(jsonl)
+    return {"status": "completed", "report": report, "answered": answered_questions,
+            "jsonl": jsonl, "elapsed_s": round(time.time() - start_time, 1)}
+
+
+def main():
+    import argparse
+    ap = argparse.ArgumentParser()
+    ap.add_argument("command", help='要跑的命令, 如 "/wl-search 考勤"')
+    ap.add_argument("--auto-answer", type=int, default=None,
+                    help="强制选项序号(0-based), 如 --auto-answer=0")
+    ap.add_argument("--no-answer", action="store_true", help="不自动代答")
+    ap.add_argument("--timeout", type=int, default=600)
+    ap.add_argument("--json", action="store_true")
+    args = ap.parse_args()
+
+    result = run(args.command, force_answer=args.auto_answer,
+                 timeout=args.timeout, no_answer=args.no_answer)
+
+    if args.json:
+        print(json.dumps(result, ensure_ascii=False, indent=2, default=str))
+    else:
+        print("\n" + "=" * 60)
+        if result.get("status") == "completed":
+            print("✅ 完成! 耗时报告:")
+            print("=" * 60)
+            print(tt._format_report(result["report"]))
+            print("\n自动代答:")
+            for a in result.get("answered", []):
+                print(f"  Q: {a['question']}")
+                print(f"  A: #{a['answer_index']} {a['answer_label']}")
+            print(f"\n实际墙钟: {result.get('elapsed_s')}s")
+        elif result.get("status") == "timeout":
+            print(f"⏰ 超时 {result.get('timeout_s')}s")
+        else:
+            print(f"❌ {result.get('status')}: {result}")
+
+
+if __name__ == "__main__":
+    main()