Agent/examples/process_pipeline/script/search_eval/server.py

# -*- coding: utf-8 -*-
"""搜索评估案例查看 server。
沿用 图文排版搜索评估.html 的版式(卡片 + dialog 详情 + rubric 评分条)，
数据实时扫描 runs_full/*/form_*.json —— runs_full 下每新增一个 q 文件夹，刷新即出现。
分页：query → 三种形式(A/B/C) → 三个渠道 三行从上到下。

用法：python server.py [port]   默认 8770，浏览器开 http://0.0.0.0:8770
"""
import json, re, glob, sys, pathlib, subprocess, threading
from datetime import datetime
from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
from urllib.parse import urlparse, parse_qs

try:  # Windows 控制台默认 cp1252，中文 print 会崩，统一切 utf-8
    sys.stdout.reconfigure(encoding="utf-8")
except Exception:
    pass

HERE = pathlib.Path(__file__).parent
sys.path.insert(0, str(HERE))
PORT = int(sys.argv[1]) if len(sys.argv) > 1 else 8770

PLAT = {"xhs": "小红书", "gzh": "公众号", "zhihu": "知乎", "x": "X", "bili": "B站", "douyin": "抖音",
        "sph": "视频号", "youtube": "YouTube", "github": "GitHub", "toutiao": "头条", "weibo": "微博"}
KT = {"procedure": "工序", "step": "步骤", "tool": "工具"}

# 从 taxonomy 取动作叶子/类型名，用于把 original_q 解析回原始维度(动作×类型 正交)
# 路径优先级：search_eval/evaluation/（主源，IDE 编辑那份就是 runtime 实际读的）
# → test_script/evaluation/（历史副本兜底）→ script/evaluation/（更老兜底）
# 谁也找不到时整目录扫空，server 仍能起。
EVALDIR = HERE / "evaluation"
if not EVALDIR.exists():
    EVALDIR = HERE.parent.parent / "test_script" / "evaluation"
if not EVALDIR.exists():
    EVALDIR = HERE.parent / "evaluation"
try:
    _jm = json.load(open(EVALDIR / "judged_matrix.json", encoding="utf-8"))
    ACT_L1 = {a["name"]: a["l1"] for a in _jm["actions"]}
    ACTION_SET = set(ACT_L1)
    TYPE_SET = {t["name"] for t in _jm["types"]}
    ACTIONS_TAX = [{"name": a["name"], "l1": a["l1"], "l2": a.get("l2", "")} for a in _jm["actions"]]
    TYPES_TAX = [{"name": t["name"], "l1": t["l1"]} for t in _jm["types"]]
    # taxonomy 顺序沿用 judged_matrix(严格版)；矩阵分值改用 type_action_scores(宽松版) —
    # 两份是同一组 27×50 cell 的独立 gemini judging，前者只 53 格到 tier3，后者 156 格到 score3
    _tas = json.load(open(EVALDIR / "type_action_scores.json", encoding="utf-8"))["scores"]
    _MATRIX = []
    for a in _jm["actions"]:
        row = []
        for t in _jm["types"]:
            rec = _tas.get(t["name"], {}).get(a["name"])
            row.append({"tier": rec["score"], "r": rec.get("reason", "")} if rec else {})
        _MATRIX.append(row)
except Exception:
    ACT_L1, ACTION_SET, TYPE_SET, ACTIONS_TAX, TYPES_TAX, _MATRIX = {}, set(), set(), [], [], []

ACTIVE_TASKS = {}
ACTIVE_REEVALS = {}

from batch_extract_procedures import _short_case, _source_to_dsl_input, _write_meta, _composite_score

def run_extraction_task(q, folder_name, src_path, out_dir, engine, model):
    task_key = f"{q}/{folder_name}"
    log_path = out_dir / "_extract.log"
    try:
        out_dir.mkdir(parents=True, exist_ok=True)
        if engine == "cyber_runner":
            script_path = HERE / "procedure-dsl" / "run_cyber.py"
        else:
            script_path = HERE / "procedure-dsl" / "run_procedure_dsl.py"
            
        cmd = [
            sys.executable, "-u", str(script_path),
            str(src_path),
            "--out-dir", str(out_dir),
            "--model", model,
            "--max-turns", "300"
        ]
        if engine != "cyber_runner":
            cmd.extend(["--max-retries", "3"])
        
        flags = subprocess.CREATE_NEW_PROCESS_GROUP if sys.platform == "win32" else 0
        with open(log_path, "w", encoding="utf-8", buffering=1) as log_fh:
            proc = subprocess.Popen(cmd, stdout=log_fh, stderr=subprocess.STDOUT,
                                    cwd=str(HERE), creationflags=flags)
            ACTIVE_TASKS[task_key]["pid"] = proc.pid
            proc.wait()
            
            if proc.returncode == 0:
                try:
                    import build_workflows
                    build_workflows.write_one(q, folder_name, runs_dir=HERE / "runs_full")
                    ACTIVE_TASKS[task_key]["status"] = "success"
                except Exception as ex:
                    ACTIVE_TASKS[task_key]["status"] = "failed"
                    ACTIVE_TASKS[task_key]["error"] = f"Workflow compilation failed: {ex}"
                    with open(log_path, "a", encoding="utf-8") as f_err:
                        f_err.write(f"\n[server error] Workflow compilation failed: {ex}\n")
            else:
                ACTIVE_TASKS[task_key]["status"] = "failed"
                ACTIVE_TASKS[task_key]["error"] = f"Runner failed with exit code {proc.returncode}"
    except Exception as e:
        ACTIVE_TASKS[task_key]["status"] = "failed"
        ACTIVE_TASKS[task_key]["error"] = str(e)
        try:
            with open(log_path, "a", encoding="utf-8") as f_err:
                f_err.write(f"\n[server error] Extraction failed: {e}\n")
        except Exception:
            pass

MODSET = {"文", "图", "视频", "音频"}
TOOLQUAL = {"AI": "AI 模型", "软件": "桌面 APP", "电脑端": "桌面 APP", "在线": "云端 Web",
            "网页版": "云端 Web", "代码": "API·CLI", "命令行": "API·CLI", "插件": "插件扩展"}

def parse_dims(oq):
    """把组合 query(如 '文 元素生成 提示词 教程')解析回 {动作, 类型, 动作L1, 约束}。"""
    toks = (oq or "").split()
    action = next((t for t in toks if t in ACTION_SET), None)
    type_ = next((t for t in toks if t in TYPE_SET), None)
    cons = None
    if toks:
        t0 = toks[0]
        if t0 in MODSET:
            cons = {"kind": "模态", "value": t0}
        elif t0 in TOOLQUAL:
            cons = {"kind": "工具类型", "value": TOOLQUAL[t0]}
    return {"action": action, "type": type_, "action_l1": ACT_L1.get(action, ""), "constraint": cons}

def flat_scores(sc):
    f = {}
    for k, v in (sc or {}).items():
        if isinstance(v, dict):
            for kk, vv in v.items():
                try: f[kk] = int(vv)
                except Exception: pass
        else:
            try: f[k] = int(v)
            except Exception: pass
    return f

def _recency_hard(date_str):
    """按 publish_timestamp 头 10 字符（YYYY-MM-DD）算硬时效：半年内=3 / 两年内=2 / 更早=1。

    取代原 LLM 评的 recency 维度——脚本算更稳，发布时间在帖子抓取时就有，无需 LLM token。
    """
    try:
        d = datetime.strptime((date_str or "")[:10], "%Y-%m-%d")
    except (ValueError, TypeError):
        return None
    days = (datetime.now() - d).days
    if days <= 180: return 3
    if days <= 730: return 2
    return 1


def adapt(r, run, form_name=None):
    p = r.get("post", {}); e = r.get("llm_evaluation", {})
    
    # 1. 解析 知识类型 (knowledge_type)
    kt = []
    kt_raw = e.get("知识类型") or e.get("knowledge_type") or []
    for k in kt_raw:
        if k in ("工序", "procedure"): kt.append("procedure")
        elif k in ("能力", "步骤", "step"): kt.append("step")
        elif k in ("工具", "tool"): kt.append("tool")

    fs = {}
    score_reasons = {}

    # 检测是否为 eval_prompt_sample-mod 里的新版 0-10 分数 schema
    is_mod_schema = "相关性" in e and isinstance(e["相关性"], dict) and ("和内容制作知识相关" in e["相关性"] or "和 query 相关" in e["相关性"])

    if is_mod_schema:
        # 新版 0-10 分数格式解析
        # 1. 相关性
        rel = e.get("相关性") or {}
        for subkey, item in rel.items():
            if isinstance(item, dict):
                score_val = item.get("得分")
                reason_val = item.get("理由")
                code_key = None
                if "内容制作" in subkey or "知识" in subkey:
                    code_key = "relevance_production"
                elif "query" in subkey or "检索" in subkey:
                    code_key = "relevance_query"
                if code_key and score_val is not None:
                    try:
                        fs[code_key] = float(score_val)
                        if reason_val:
                            score_reasons[code_key] = reason_val
                    except Exception:
                        pass

        # 2. 质量
        q_block = e.get("质量") or {}
        fixed = q_block.get("固定维度") or {}
        
        # 固定维度
        fixed_keys = {
            "时效性": "recency",
            "热度性": "popularity",
            "评论反馈": "feedback"
        }
        for cn, code in fixed_keys.items():
            item = fixed.get(cn)
            if isinstance(item, dict):
                score_val = item.get("得分")
                reason_val = item.get("理由")
                if score_val is not None:
                    try:
                        fs[code] = float(score_val)
                        if reason_val:
                            score_reasons[code] = reason_val
                    except Exception:
                        pass
                        
        # 用例 (真实感, 表现力)
        usecase = fixed.get("用例") or {}
        usecase_keys = {
            "真实感": "realism",
            "表现力": "expressiveness"
        }
        for cn, code in usecase_keys.items():
            item = usecase.get(cn)
            if isinstance(item, dict):
                score_val = item.get("得分")
                reason_val = item.get("理由")
                if score_val is not None:
                    try:
                        fs[code] = float(score_val)
                        if reason_val:
                            score_reasons[code] = reason_val
                    except Exception:
                        pass

        # 动态维度
        dynamic = q_block.get("动态维度") or {}
        
        # 工序
        proc = dynamic.get("工序") or {}
        if proc:
            item = proc.get("流程完整性")
            if isinstance(item, dict):
                score_val = item.get("得分")
                reason_val = item.get("理由")
                if score_val is not None:
                    try:
                        fs["procedure_completeness"] = float(score_val)
                        if reason_val:
                            score_reasons["procedure_completeness"] = reason_val
                    except Exception:
                        pass
            field = proc.get("字段完整性") or {}
            field_keys = {
                "输入完整性": "procedure_input",
                "实现完整性": "procedure_implementation",
                "输出完整性": "procedure_output"
            }
            for cn, code in field_keys.items():
                item = field.get(cn)
                if isinstance(item, dict):
                    score_val = item.get("得分")
                    reason_val = item.get("理由")
                    if score_val is not None:
                        try:
                            fs[code] = float(score_val)
                            if reason_val:
                                score_reasons[code] = reason_val
                        except Exception:
                            pass
            item = proc.get("泛化性")
            if isinstance(item, dict):
                score_val = item.get("得分")
                reason_val = item.get("理由")
                if score_val is not None:
                    try:
                        fs["procedure_generality"] = float(score_val)
                        if reason_val:
                            score_reasons["procedure_generality"] = reason_val
                    except Exception:
                        pass

        # 能力
        cap = dynamic.get("能力") or dynamic.get("步骤") or {}
        if cap:
            field = cap.get("字段完整性") or {}
            field_keys = {
                "输入完整性": "step_input",
                "实现完整性": "step_implementation",
                "输出完整性": "step_output"
            }
            for cn, code in field_keys.items():
                item = field.get(cn)
                if isinstance(item, dict):
                    score_val = item.get("得分")
                    reason_val = item.get("理由")
                    if score_val is not None:
                        try:
                            fs[code] = float(score_val)
                            if reason_val:
                                score_reasons[code] = reason_val
                        except Exception:
                            pass
            item = cap.get("泛化性")
            if isinstance(item, dict):
                score_val = item.get("得分")
                reason_val = item.get("理由")
                if score_val is not None:
                    try:
                        fs["step_generality"] = float(score_val)
                        if reason_val:
                            score_reasons["step_generality"] = reason_val
                    except Exception:
                        pass

        # 工具
        tool = dynamic.get("工具") or {}
        if tool:
            tool_keys = {
                "能力边界覆盖": "tool_boundary",
                "有效比较": "tool_comparison",
                "参数/接口具体性": "tool_specificity",
                "实操示例": "tool_example",
                "版本&限制": "tool_limits"
            }
            for cn, code in tool_keys.items():
                item = tool.get(cn)
                if isinstance(item, dict):
                    score_val = item.get("得分")
                    reason_val = item.get("理由")
                    if score_val is not None:
                        try:
                            fs[code] = float(score_val)
                            if reason_val:
                                score_reasons[code] = reason_val
                        except Exception:
                            pass

    else:
        # 兼容老版 1-5 分数 schema (带 "评分" 或 old-style flatness)
        is_new_schema = "评分" in e or "知识类型" in e or "制作相关性" in e
        CN_TO_EN = {
            "相关性": "relevance",
            "成品质量": "result_quality",
            "可信度": "credibility",
            "具体用例": "concrete_use_case",
            "完整性": "completeness",
            "步骤结构": "step_structure",
            "步骤可复现": "step_reproducibility",
            "步骤可复现性": "step_reproducibility",
            "能力定义": "capability_definition",
            "实现深度": "implementation_depth",
            "边界失败": "boundary_failure_eval",
            "通用性": "generality",
            "能力覆盖": "capability_coverage",
            "有效对比": "effective_comparison",
            "参数具体": "param_specificity",
            "实操示例": "worked_example",
            "实操用例": "worked_example",
            "示例完整": "worked_example",
            "版本限制": "version_limits",
            "版本说明": "version_limits",
            "限制说明": "version_limits",
        }
        
        if is_new_schema:
            pf = e.get("评分") or {}
            for cat, metrics in pf.items():
                if isinstance(metrics, dict):
                    for metric, val in metrics.items():
                        en_key = CN_TO_EN.get(metric, metric)
                        if isinstance(val, dict) and "得分" in val:
                            try: fs[en_key] = int(val["得分"])
                            except Exception: pass
                        elif isinstance(val, (int, float)):
                            fs[en_key] = int(val)
                        
                        if isinstance(val, dict) and "理由" in val:
                            score_reasons[en_key] = val["理由"]
        else:
            fs = flat_scores(e.get("scores", {}))
        
    # 计算均分 (overall)
    if is_mod_schema:
        rel_keys = {"relevance_production", "relevance_query"}
        rel_vals = [v for k, v in fs.items() if k in rel_keys]
        qual_vals = [v for k, v in fs.items() if k not in rel_keys]
        
        rel_avg = sum(rel_vals) / len(rel_vals) if rel_vals else None
        qual_avg = sum(qual_vals) / len(qual_vals) if qual_vals else None
        
        if rel_avg is not None and qual_avg is not None:
            overall = round((rel_avg + qual_avg) / 2, 1)
        elif rel_avg is not None:
            overall = round(rel_avg, 1)
        elif qual_avg is not None:
            overall = round(qual_avg, 1)
        else:
            overall = 0.0
    else:
        overall = round(sum(fs.values()) / len(fs), 1) if fs else 0
    anomaly = bool(e.get("error")) or not fs
    grade = p.get("_quality_grade", "")
    fb = r.get("found_by_queries", [])
    
    # 4. 解析 制作相关性 (production_relevance)
    if is_mod_schema:
        # 新版使用 "相关性" 中的 "和内容制作知识相关" 代表制作相关性
        production_relevance = fs.get("relevance_production")
    else:
        if is_new_schema:
            pr_block = e.get("制作相关性") or {}
            pr_raw = pr_block.get("得分") if isinstance(pr_block, dict) else pr_block
            if isinstance(pr_block, dict) and "理由" in pr_block:
                score_reasons["production_relevance"] = pr_block["理由"]
        else:
            pr_raw = e.get("production_relevance")
            
        try: production_relevance = int(float(pr_raw)) if pr_raw is not None else None
        except (TypeError, ValueError): production_relevance = None
    
    recency_hard = _recency_hard(p.get("publish_timestamp", ""))
    
    # 5. 解析 判定决策 (decision) 和 理由 (reason)
    reason = e.get("判定理由") or e.get("reason") or ""
    
    # 根据过滤指标决定是否保留 (过滤指标判定逻辑优先，不依赖文字匹配)
    is_discard = False
    
    # 制作相关性低于阈值则丢弃 (新版 0-10 满分，因此低于 4 丢弃；老版低于 2 丢弃)
    if production_relevance is not None:
        threshold = 4 if is_mod_schema else 2
        if production_relevance < threshold:
            is_discard = True
            
    # 时效性低于 2 被丢弃（发布时间超两年的老帖）
    if recency_hard is not None and recency_hard < 2:
        is_discard = True
        
    # 综合均分低于阈值被丢弃 (新版低于 6 丢弃；老版低于 3 丢弃)
    if overall is not None:
        threshold_ov = 6 if is_mod_schema else 3
        if overall < threshold_ov:
            is_discard = True
        
    decision = "discard" if is_discard else "report"

    # Find matching procedure html
    procedure_html = None
    case_id = r.get("case_id", "")
    title = p.get("title", "")
    run_dir = HERE / "runs_full" / run

    if run_dir.is_dir():
        # 1. 优先扫描该帖子对应的文件夹下的任何 HTML 文件 (不限名称)
        # 文件夹名格式: {form}_{platform}_{channel_content_id[:8]}
        content_id = r.get("channel_content_id") or ""
        if not content_id and case_id and "_" in case_id:
            content_id = case_id.split("_", 1)[1]
        plat_key = r.get("platform") or ""
        
        if form_name and plat_key and content_id:
            folder_name = f"{form_name}_{plat_key}_{content_id[:8]}"
            case_dir = run_dir / "procedures" / folder_name
            if case_dir.is_dir():
                html_files = list(case_dir.glob("*.html"))
                if html_files:
                    procedure_html = f"runs_full/{run}/procedures/{folder_name}/{html_files[0].name}"

        # 2. 其次匹配标准文件名: case-{case_id}.html 或 {case_id}.html
        candidate_dirs = [run_dir, run_dir / "procedures"]
        if not procedure_html and case_id:
            named_files = [f"case-{case_id}.html", f"{case_id}.html"]
            for d_dir in candidate_dirs:
                if d_dir.is_dir():
                    for name in named_files:
                        if (d_dir / name).is_file():
                            procedure_html = f"runs_full/{run}/procedures/{name}" if d_dir.name == "procedures" else f"runs_full/{run}/{name}"
                            break
                if procedure_html:
                    break

        # 3. 再次匹配 HTML 内部的标准声明 (meta 标签或 HTML 注释)
        if not procedure_html and case_id:
            for d_dir in candidate_dirs:
                if d_dir.is_dir():
                    for html_path in d_dir.glob("*.html"):
                        try:
                            content = html_path.read_text(encoding="utf-8")
                            if f'name="case-id" content="{case_id}"' in content or \
                               f'name="case_id" content="{case_id}"' in content or \
                               f'<!-- case_id: {case_id} -->' in content or \
                               f'<!-- case-id: {case_id} -->' in content:
                                procedure_html = f"runs_full/{run}/procedures/{html_path.name}" if d_dir.name == "procedures" else f"runs_full/{run}/{html_path.name}"
                                break
                        except Exception:
                            continue
                if procedure_html:
                    break

        # 4. 最后使用标题作为兜底模糊匹配
        if not procedure_html and title:
            for d_dir in candidate_dirs:
                if d_dir.is_dir():
                    for html_path in d_dir.glob("*.html"):
                        try:
                            content = html_path.read_text(encoding="utf-8")
                            if title in content:
                                procedure_html = f"runs_full/{run}/procedures/{html_path.name}" if d_dir.name == "procedures" else f"runs_full/{run}/{html_path.name}"
                                break
                        except Exception:
                            continue
                if procedure_html:
                    break

    return {
        "case_id": r.get("case_id", ""),
        "platform": PLAT.get(r.get("platform"), r.get("platform")), "platformKey": r.get("platform"),
        "title": p.get("title", "") or "(无标题)", "date": (p.get("publish_timestamp", "") or "")[:10],
        "url": r.get("source_url", ""), "engagement": f'{p.get("like_count", 0)} 赞',
        "knowledge_type": kt, "decision": decision,
        "tools": [KT.get(k, k) for k in kt] + ([f"质量 {grade}"] if grade else []), "found_by": fb,
        "images": (p.get("images") or [])[:6], "text": p.get("body_text", "") or "",
        "scores": fs, "overall": overall, "reason": reason, "score_reasons": score_reasons,
        "grade": grade, "qscore": p.get("_quality_score", 0), "anomaly": anomaly,
        "production_relevance": production_relevance, "recency_hard": recency_hard,
        "run": run, "procedure_html": procedure_html,
    }

def scan_runs():
    runs = {}
    for f in sorted(glob.glob(str(HERE / "runs_full" / "*" / "form_*.json"))):
        try:
            d = json.load(open(f, encoding="utf-8"))
        except Exception:
            continue
        run = pathlib.Path(f).parent.name
        form_name = d.get("form") or ""
        results = [adapt(r, run, form_name) for r in d.get("results", [])]
        report_val = sum(1 for r in results if r.get("decision") == "report" and not r.get("anomaly"))
        discard_val = sum(1 for r in results if r.get("decision") == "discard" and not r.get("anomaly"))
        
        runs.setdefault(run, []).append({
            "form": d.get("form"), "query": d.get("query"), "original_q": d.get("original_q", ""),
            "requirement": d.get("requirement", ""),
            "platforms": d.get("platforms", []), "total": d.get("total"),
            "report": report_val, "discard": discard_val,
            "results": results,
        })
    for v in runs.values():
        v.sort(key=lambda x: x.get("form") or "")

    def _qnum(name):  # "q156" → 156，按数字排，避免 "q156" < "q99" 的字符串误排
        m = re.search(r"\d+", name)
        return (int(m.group()) if m else 0, name)
    out = []
    for k, v in sorted(runs.items(), key=lambda kv: _qnum(kv[0])):
        oq = v[0].get("original_q") or v[0].get("query") or ""
        seen, hits = set(), 0   # 知识命中数 = 各形式采纳(report)且非异常、按 url 去重后的帖子数
        for f in v:
            for r in f.get("results", []):
                if r.get("decision") == "report" and not r.get("anomaly") and r.get("url") not in seen:
                    seen.add(r.get("url")); hits += 1
        out.append({"key": k, "forms": v, "dims": parse_dims(oq), "original_q": oq,
                    "hits": hits, "tot": sum((f.get("total") or 0) for f in v)})
    active_reevals = {k: v["status"] for k, v in ACTIVE_REEVALS.items()}
    return {"queries": out, "actions": ACTIONS_TAX, "types": TYPES_TAX, "matrix": _MATRIX, "active_reevals": active_reevals}

class H(BaseHTTPRequestHandler):
    def _send(self, code, body, ctype):
        b = body.encode("utf-8") if isinstance(body, str) else body
        self.send_response(code)
        if ctype.startswith("text/") or ctype == "application/json" or ctype == "application/javascript":
            self.send_header("Content-Type", ctype + "; charset=utf-8")
        else:
            self.send_header("Content-Type", ctype)
        self.send_header("Content-Length", str(len(b))); self.end_headers(); self.wfile.write(b)
    def do_GET(self):
        parsed = urlparse(self.path)
        path = parsed.path
        params = parse_qs(parsed.query)

        if path in ("/", "/index.html"):
            try:
                page = (HERE / "index.html").read_text(encoding="utf-8")
                self._send(200, page, "text/html")
            except Exception as e:
                self._send(500, f"Error reading index.html: {e}", "text/plain")
        elif path == "/api/data":
            self._send(200, json.dumps(scan_runs(), ensure_ascii=False), "application/json")
        elif path == "/api/procedure_status":
            q = (params.get("q") or [""])[0].strip()
            form = (params.get("form") or [""])[0].strip()
            case_id = (params.get("case_id") or [""])[0].strip()
            
            if not q or not form or not case_id:
                self._send(400, "missing q, form, or case_id", "text/plain")
                return
                
            folder_name = f"{form}_{_short_case(case_id)}"
            task_key = f"{q}/{folder_name}"
            
            if task_key in ACTIVE_TASKS:
                task = ACTIVE_TASKS[task_key]
                res = {
                    "status": task["status"],
                    "error": task["error"]
                }
                if task["status"] == "success":
                    out_dir = HERE / "runs_full" / q / "procedures" / folder_name
                    html_files = list(out_dir.glob("*.html")) if out_dir.is_dir() else []
                    if html_files:
                        res["procedure_html"] = f"runs_full/{q}/procedures/{folder_name}/{html_files[0].name}"
                self._send(200, json.dumps(res, ensure_ascii=False), "application/json")
                return
                
            out_dir = HERE / "runs_full" / q / "procedures" / folder_name
            html_files = list(out_dir.glob("*.html")) if out_dir.is_dir() else []
            if html_files:
                self._send(200, json.dumps({
                    "status": "success",
                    "procedure_html": f"runs_full/{q}/procedures/{folder_name}/{html_files[0].name}"
                }, ensure_ascii=False), "application/json")
                return
                
            log_path = out_dir / "_extract.log"
            if log_path.is_file():
                self._send(200, json.dumps({"status": "failed", "error": "Not running, but no HTML output found (possibly crashed)."}, ensure_ascii=False), "application/json")
                return
                
            self._send(200, json.dumps({"status": "not_started"}, ensure_ascii=False), "application/json")
        elif path == "/api/procedure_log":
            q = (params.get("q") or [""])[0].strip()
            form = (params.get("form") or [""])[0].strip()
            case_id = (params.get("case_id") or [""])[0].strip()
            
            if not q or not form or not case_id:
                self._send(400, "missing q, form, or case_id", "text/plain")
                return
                
            folder_name = f"{form}_{_short_case(case_id)}"
            log_path = HERE / "runs_full" / q / "procedures" / folder_name / "_extract.log"
            
            if not log_path.is_file():
                self._send(200, json.dumps({"log": ""}, ensure_ascii=False), "application/json")
                return
                
            try:
                content = log_path.read_text(encoding="utf-8", errors="replace")
                self._send(200, json.dumps({"log": content}, ensure_ascii=False), "application/json")
            except Exception as e:
                self._send(500, json.dumps({"error": str(e)}, ensure_ascii=False), "application/json")
        elif path == "/api/spec_content":
            file_name = (params.get("file") or [""])[0].strip()
            allowed = [
                "README.md",
                "tools.md",
                "extraction/phase1-skeleton.md",
                "extraction/phase2-normalize.md",
                "extraction/phase3-finalize.md",
                "taxonomy/type_suggestions.md"
            ]
            if file_name not in allowed:
                self._send(400, "invalid file parameter", "text/plain")
                return
            target_path = HERE / "procedure-dsl" / "spec" / file_name
            if not target_path.is_file():
                self._send(404, "spec file not found", "text/plain")
                return
            try:
                content = target_path.read_text(encoding="utf-8", errors="replace")
                self._send(200, json.dumps({"content": content}, ensure_ascii=False), "application/json")
            except Exception as e:
                self._send(500, json.dumps({"error": str(e)}, ensure_ascii=False), "application/json")
        elif path == "/api/reeval_status":
            q = (params.get("q") or [""])[0].strip()
            if not q:
                self._send(400, "missing q", "text/plain")
                return
            if q in ACTIVE_REEVALS:
                self._send(200, json.dumps({
                    "status": ACTIVE_REEVALS[q]["status"],
                    "error": ACTIVE_REEVALS[q].get("error")
                }, ensure_ascii=False), "application/json")
            else:
                self._send(200, json.dumps({"status": "not_started"}, ensure_ascii=False), "application/json")
        elif self.path.startswith("/runs_full/"):
            try:
                clean_path = self.path.split("?")[0]
                parts = clean_path.strip("/").split("/")
                target_file = HERE
                for part in parts:
                    target_file = target_file / part
                runs_dir = HERE / "runs_full"
                if runs_dir.resolve() in target_file.resolve().parents and target_file.is_file():
                    content = target_file.read_bytes()
                    ext = target_file.suffix.lower()
                    ctype = "text/html"
                    if ext in (".png", ".webp"):
                        ctype = f"image/{ext[1:]}"
                    elif ext in (".jpg", ".jpeg"):
                        ctype = "image/jpeg"
                    elif ext == ".json":
                        ctype = "application/json"
                    elif ext == ".js":
                        ctype = "application/javascript"
                    elif ext == ".css":
                        ctype = "text/css"
                    self._send(200, content, ctype)
                else:
                    self._send(404, "not found", "text/plain")
            except Exception as e:
                self._send(500, f"Error: {e}", "text/plain")
        else:
            self._send(404, "not found", "text/plain")
    def do_POST(self):
        if self.path == "/api/generate_procedure":
            length = int(self.headers.get("Content-Length") or 0)
            raw = self.rfile.read(length).decode("utf-8") if length > 0 else "{}"
            try:
                payload = json.loads(raw)
            except Exception as e:
                self._send(400, json.dumps({"error": f"bad json: {e}"}), "application/json"); return
                
            q = (payload.get("q") or "").strip()
            form = (payload.get("form") or "").strip()
            case_id = (payload.get("case_id") or "").strip()
            engine = (payload.get("engine") or "cyber_runner").strip()
            model = (payload.get("model") or "google/gemini-3.1-flash-lite").strip()
            
            if not re.match(r"^q\d+$", q):
                self._send(400, json.dumps({"error": f"bad q (expect 'qNN'): {q!r}"}, ensure_ascii=False), "application/json"); return
            if form not in ("A", "B", "C"):
                self._send(400, json.dumps({"error": f"bad form: {form!r}"}, ensure_ascii=False), "application/json"); return
            if not case_id:
                self._send(400, json.dumps({"error": "missing case_id"}, ensure_ascii=False), "application/json"); return
                
            q_dir = HERE / "runs_full" / q
            form_file = q_dir / f"form_{form}.json"
            if not form_file.is_file():
                self._send(404, json.dumps({"error": f"form file not found: {form_file.name}"}, ensure_ascii=False), "application/json"); return
                
            try:
                with open(form_file, encoding="utf-8") as f:
                    form_data = json.load(f)
            except Exception as e:
                self._send(500, json.dumps({"error": f"failed to read form: {e}"}, ensure_ascii=False), "application/json"); return
                
            matching_result = None
            for r in form_data.get("results", []):
                if r.get("case_id") == case_id:
                    matching_result = r
                    break
                    
            if not matching_result:
                self._send(404, json.dumps({"error": f"case_id {case_id} not found in form {form}"}, ensure_ascii=False), "application/json"); return
                
            folder_name = f"{form}_{_short_case(case_id)}"
            out_dir = q_dir / "procedures" / folder_name
            out_dir.mkdir(parents=True, exist_ok=True)
            
            src_path = out_dir / "_source.json"
            try:
                with open(src_path, "w", encoding="utf-8") as f:
                    json.dump(_source_to_dsl_input(matching_result), f, ensure_ascii=False, indent=2)
                    
                score = _composite_score(matching_result.get("llm_evaluation") or {})
                _write_meta(out_dir, case_id=case_id, from_q=q, form=form, score=score)
            except Exception as e:
                self._send(500, json.dumps({"error": f"failed to write inputs: {e}"}, ensure_ascii=False), "application/json"); return
                
            task_key = f"{q}/{folder_name}"
            ACTIVE_TASKS[task_key] = {
                "status": "running",
                "start_time": datetime.now().isoformat(),
                "pid": None,
                "error": None
            }
            
            t = threading.Thread(target=run_extraction_task, args=(q, folder_name, src_path, out_dir, engine, model))
            t.daemon = True
            t.start()
            
            self._send(200, json.dumps({
                "status": "started",
                "task_key": task_key,
                "log": f"runs_full/{q}/procedures/{folder_name}/_extract.log"
            }, ensure_ascii=False), "application/json")
        elif self.path == "/api/reeval":
            length = int(self.headers.get("Content-Length") or 0)
            raw = self.rfile.read(length).decode("utf-8") if length > 0 else "{}"
            try:
                payload = json.loads(raw)
            except Exception as e:
                self._send(400, json.dumps({"error": f"bad json: {e}"}), "application/json"); return
            q = (payload.get("q") or "").strip()
            if not re.match(r"^q\d+$", q):
                self._send(400, json.dumps({"error": f"bad q (expect 'qNN'): {q!r}"},
                                           ensure_ascii=False), "application/json"); return
            q_dir = HERE / "runs_full" / q
            if not q_dir.is_dir():
                self._send(404, json.dumps({"error": f"runs_full/{q} not found"}, ensure_ascii=False),
                           "application/json"); return
            log_path = q_dir / "_reeval.log"
            try:
                log_fh = open(log_path, "w", encoding="utf-8", buffering=1)
                cmd = [sys.executable, "-u", str(HERE / "batch_3forms.py"),
                       "--reeval", "--reeval-q", q, "--output-dir", str(HERE / "runs_full")]
                flags = subprocess.CREATE_NEW_PROCESS_GROUP if sys.platform == "win32" else 0
                proc = subprocess.Popen(cmd, stdout=log_fh, stderr=subprocess.STDOUT,
                                        cwd=str(HERE), creationflags=flags)
                
                ACTIVE_REEVALS[q] = {
                    "status": "running",
                    "pid": proc.pid,
                    "error": None
                }
                
                def wait_reeval(q_key, p_obj, fh):
                    try:
                        p_obj.wait()
                        if p_obj.returncode == 0:
                            ACTIVE_REEVALS[q_key]["status"] = "success"
                        else:
                            ACTIVE_REEVALS[q_key]["status"] = "failed"
                            ACTIVE_REEVALS[q_key]["error"] = f"Subprocess exited with code {p_obj.returncode}"
                    except Exception as ex:
                        ACTIVE_REEVALS[q_key]["status"] = "failed"
                        ACTIVE_REEVALS[q_key]["error"] = str(ex)
                    finally:
                        try:
                            fh.close()
                        except Exception:
                            pass
                
                t = threading.Thread(target=wait_reeval, args=(q, proc, log_fh))
                t.daemon = True
                t.start()
                
                self._send(200, json.dumps(
                    {"status": "started", "pid": proc.pid, "q": q,
                     "log": str(log_path.relative_to(HERE))},
                    ensure_ascii=False), "application/json")
            except Exception as e:
                self._send(500, json.dumps({"error": f"failed to start: {e}"},
                                           ensure_ascii=False), "application/json")
        elif self.path == "/api/save_spec":
            length = int(self.headers.get("Content-Length") or 0)
            raw = self.rfile.read(length).decode("utf-8") if length > 0 else "{}"
            try:
                payload = json.loads(raw)
            except Exception as e:
                self._send(400, json.dumps({"error": f"bad json: {e}"}), "application/json"); return
            file_name = (payload.get("file") or "").strip()
            content = payload.get("content") or ""
            allowed = [
                "README.md",
                "tools.md",
                "extraction/phase1-skeleton.md",
                "extraction/phase2-normalize.md",
                "extraction/phase3-finalize.md",
                "taxonomy/type_suggestions.md"
            ]
            if file_name not in allowed:
                self._send(400, json.dumps({"error": "invalid file parameter"}), "application/json"); return
            target_path = HERE / "procedure-dsl" / "spec" / file_name
            try:
                target_path.parent.mkdir(parents=True, exist_ok=True)
                target_path.write_text(content, encoding="utf-8")
                self._send(200, json.dumps({"status": "ok"}, ensure_ascii=False), "application/json")
            except Exception as e:
                self._send(500, json.dumps({"error": str(e)}, ensure_ascii=False), "application/json")
        else:
            self._send(404, json.dumps({"error": "not found"}), "application/json")
    def log_message(self, *a): pass

if __name__ == "__main__":
    n = len(scan_runs()["queries"])
    print(f"搜索评估查看 server：http://0.0.0.0:{PORT}   (runs_full/ 下 {n} 个 query，实时扫描)")
    ThreadingHTTPServer(("0.0.0.0", PORT), H).serve_forever()