batch 20 posts

examples/process_pipeline/script/search_eval/batch_extract_procedures.py

@@ -1,53 +1,60 @@
-"""批量化跑 procedure-dsl: 对每个 q 的三 form 选评分最高的"工序"帖子,
-喂给 procedure-dsl/run_procedure_dsl.py 提取工序, HTML 直接落在
-runs_full/qNNNN/procedures/{form}_{short_case}/ 下。
-
-筛选:  llm_evaluation.知识类型 含 "工序" (兼容英文 "procedure") 且非 _error
-       —— 工序帖才有端到端流程可抽; 能力/工具帖太原子, 抽不出 procedure。
-评分:  两级均值, 避免子项多的组天然占便宜:
-       group_rel  = mean(相关性 下所有得分)
-       group_qual = mean(质量 下所有得分, 含固定+动态)
-       final = mean(group_rel, group_qual)
-选条:  每 form top-K (默认 1)。两层去重, 撞了顺延下一个候选:
-       ① 本 q 跨 form (同帖只跑一次)
-       ② 全局: 扫所有 q*/procedures/*/ 目录名 → 抽 short_case → 已存在即视为占用
-         (无论跑完没有, 跑挂了也算占了 case; 想强复跑用 --force)
+"""批量化跑 procedure-dsl: 把这次列出的所有 q 的全部帖子 (每 q 三 form ≈90 帖) 倒进
+一个大池, 跨 q 按评分降序取全局 top-N (默认 20) 个未被丢弃的"工序"帖, 喂给
+procedure-dsl/run_procedure_dsl.py 提取工序。每条仍落回它自己所属那个 q 的目录:
+runs_full/qNNNN/procedures/{form}_{short_case}/。
+
+  ⚠️ top-N 是"这批 q 加起来总共 N 条", 不是"每 q N 条"。列 5 个 q 也只跑全局最高的 N 条,
+     可能全集中在某几个 q 上。想多跑就调大 --top-n 或换一批 q。
+
+筛选:  ① 非异常 (anomaly) 且 decision == report (即未被丢弃)
+       ② 知识类型 含 "工序" (procedure) —— 工序帖才有端到端流程可抽; 能力/工具帖抽不出
+       丢弃判定与评分都直接复用 server.evaluate_result (跟前端 /api/data 同一真源),
+       schema 再变也不会两边漂移; 阈值见 server.py (制作相关性/时效/综合均分)。
+评分:  server.evaluate_result(r)["overall"] —— 相关性均分与质量均分的两级平均。
+选条:  top-N 是「累计目标」—— workflows/ 里这些 q 最终要有 N 条; 本次预算 = N − 已完成数。
+       全部 q 合池后按 overall 降序取 budget 条。去重:
+       ① 池内同帖 (case_id 撞, 跨 form / 跨 q) 只留最高分那条, 落在它得分最高的 q 下
+       ② 选片只排除「已完成」(workflows/ 里有的); 在跑/跑挂的不排, 交给跑时原子 mkdir 分流
 产物:  HTML/MD/workflow.json/_source.json/_meta.json 全落
        runs_full/qNNNN/procedures/{form}_{short_case}/
-跳过:  目录已存 → skip (--force 可强覆盖)
-并发:  多终端开几个就跑几个; 每条 procedure 一启动就 mkdir, 其他终端立刻见到目录视作已占
+并发:  多窗口直接粘贴同一条指令并行 —— 每条 procedure 靠 mkdir(exist_ok=False) 原子认领,
+       谁先建谁跑、其余撞 FileExistsError 即跳, 自动把 budget 条瓜分到各窗口, 不重不超额。
+       (预算按 workflows/ 已完成数算, 所以各窗口算出的计划对齐同一 top-N 前沿, 不会越跑越多)
+       --force 不走认领锁 (允许覆盖), 故 --force 不适合多窗口并行。
 
 用法:
-  # 单 q 试
-  python batch_extract_procedures.py --only-q q0036
+  # 跨多个 q 合池, 全局取 top-20
+  python batch_extract_procedures.py --only-q q0018,q0036,q2300
 
-  # 区间
+  # 区间内的 q 合池
   python batch_extract_procedures.py --start 0 --count 5
 
-  # top-3 + 强制重跑
-  python batch_extract_procedures.py --only-q q0036 --top-k 3 --force
+  # 全局只取 top-5 + 强制重跑 (--top-k 是 --top-n 的兼容别名)
+  python batch_extract_procedures.py --only-q q0018,q0036 --top-n 5 --force
 
   # 仅列出会跑哪些 (不真跑)
-  python batch_extract_procedures.py --only-q q0036 --dry-run
+  python batch_extract_procedures.py --only-q q0018,q0036 --dry-run
 """
 import argparse
 import json
-import os
 import re
 import subprocess
 import sys
 from datetime import datetime
 from pathlib import Path
-from statistics import mean
-from typing import Any, Dict, List, Optional, Set, Tuple
+from typing import Any, Dict, List, Set, Tuple
 
 import build_workflows  # 复用其 write_run: 跑完一个 q 就把合并 json 写进 workflows/
+from server import evaluate_result  # 评分 + 丢弃判定的唯一真源 (与前端 /api/data 一致)
 
 HERE = Path(__file__).resolve().parent
 RUNS_FULL_DEFAULT = HERE / "runs_full"
 RUN_DSL = HERE / "procedure-dsl" / "run_procedure_dsl.py"
 
-_KT_PROCEDURE = {"工序", "procedure"}   # 工序没改名, 仅兼容英文枚举
+# 图过多 (>此值) 直接不选这帖, 名额顺延给下一名 —— "最次则跳过" 那一档。
+# 上限以内的多图帖由 run_procedure_dsl 侧降分辨率/拼九宫格扛 (见那边 _IMG_MONTAGE_OVER)。
+# 72 = 8 个九宫格; 再多拼图也压不住, 放弃更省。
+_IMG_SKIP_CAP = 72
 
 
 def _norm_q(s: str) -> int:
@@ -57,117 +64,85 @@ def _norm_q(s: str) -> int:
     return int(m.group())
 
 
-def _has_procedure(ev: Dict[str, Any]) -> bool:
-    """llm_evaluation.知识类型 是否声明了 工序 类型 (含英文 procedure)。"""
-    kt = ev.get("知识类型") or ev.get("knowledge_type") or []
-    if isinstance(kt, str):
-        kt = [kt]
-    return any(k in _KT_PROCEDURE for k in kt) if isinstance(kt, list) else False
-
-
-def _safe_float(x: Any) -> Optional[float]:
-    try:
-        return float(x)
-    except (TypeError, ValueError):
-        return None
-
-
-def _all_scores(obj: Any) -> List[float]:
-    """递归挖一棵嵌套 dict 里所有 '得分' 叶子的数值。"""
-    out: List[float] = []
-    if isinstance(obj, dict):
-        for k, v in obj.items():
-            if k == "得分":
-                f = _safe_float(v)
-                if f is not None:
-                    out.append(f)
-            elif isinstance(v, (dict, list)):
-                out.extend(_all_scores(v))
-    elif isinstance(obj, list):
-        for it in obj:
-            out.extend(_all_scores(it))
-    return out
-
-
-def _composite_score(ev: Dict[str, Any]) -> float:
-    """两级均值:
-       group_rel  = mean(相关性 下所有 得分)
-       group_qual = mean(质量 下所有 得分, 含固定+动态全部子项)
-       final = mean(group_rel, group_qual)
-    某一组缺失 (该组下没任何得分字段) → 降级用另一组的均值; 都缺 → 0。"""
-    rel_scores = _all_scores(ev.get("相关性"))
-    qual_scores = _all_scores(ev.get("质量"))
-
-    rel_mean = mean(rel_scores) if rel_scores else None
-    qual_mean = mean(qual_scores) if qual_scores else None
-
-    parts = [m for m in (rel_mean, qual_mean) if m is not None]
-    return mean(parts) if parts else 0.0
-
-
-def _pick_candidates(form_path: Path, exclude_ids: Set[str], top_k: int
-                     ) -> List[Tuple[float, Dict[str, Any]]]:
-    """读 form_*.json, 取评分最高的 top_k 个工序帖, 跳过 exclude_ids 里的 case_id
-    (已被其他 q 跑过 / 正在跑 / 本进程已规划)。exclude_ids 用 _short_case 形式。"""
-    try:
-        d = json.loads(form_path.read_text(encoding="utf-8"))
-    except Exception as e:
-        print(f"   ⚠️ 读 {form_path.name} 失败: {e}")
-        return []
-    cands: List[Tuple[float, Dict[str, Any]]] = []
-    for r in d.get("results", []):
-        ev = r.get("llm_evaluation") or {}
-        if ev.get("_error") or not _has_procedure(ev):
-            continue
-        cands.append((_composite_score(ev), r))
-    cands.sort(key=lambda x: x[0], reverse=True)
-    out: List[Tuple[float, Dict[str, Any]]] = []
-    for s, r in cands:
-        # 用 _short_case 化的 case_id 对齐 global 状态 / 目录命名
-        if _short_case(r.get("case_id") or "") in exclude_ids:
-            continue
-        out.append((s, r))
-        if len(out) >= top_k:
+def _pick_pool_candidates(q_dirs: List[Path], exclude_ids: Set[str], top_n: int
+                          ) -> List[Tuple[str, str, Dict[str, Any], float, float]]:
+    """把列出的所有 q 的全部 form (每 q ≈90 帖) 倒进一个大池, 跨 q 取全局 top_n 个工序帖。
+    返回 [(q_name, form_key, result, norm, raw), ...] (已按归一化分降序)。
+
+    评分: 排名用 *归一化分* norm = overall / 满分 —— 因为数据里有两套评分量纲 (mod 0-10 /
+          old 0-5), 直接比 overall 会让 0-5 的 q 永远被 0-10 压死。norm 统一到 0-1, 跨 q
+          可比; 同 schema 批次里 norm 与 overall 同序, 不影响。raw=overall 仅供展示/落档。
+    筛选: 非异常 + decision==report (未丢弃) + 知识类型含 procedure —— 评分与丢弃判定
+          全部走 server.evaluate_result, 与前端 /api/data 同一真源。
+    去重: ① 池内同 case_id 撞了只留最高分那条 (同帖可能被多 form / 多 q 同时召回,
+            最终落在它得分最高的那个 q 目录下)
+          ② 跳过 exclude_ids (磁盘上 procedure 目录已存在的), 键统一 _short_case 形式。"""
+    # short_case -> (norm, raw, q, form, r)
+    pooled: Dict[str, Tuple[float, float, str, str, Dict[str, Any]]] = {}
+    n_img_skip = 0   # 因图过多 (>_IMG_SKIP_CAP) 被放弃的工序帖数, 末尾汇报
+    for qd in q_dirs:
+        for fp in sorted(qd.glob("form_*.json")):
+            try:
+                d = json.loads(fp.read_text(encoding="utf-8"))
+            except Exception as e:
+                print(f"   ⚠️ 读 {qd.name}/{fp.name} 失败: {e}")
+                continue
+            fkey = fp.stem.replace("form_", "")    # A/B/C
+            for r in d.get("results", []):
+                ev = evaluate_result(r)
+                if ev["anomaly"] or ev["decision"] != "report":
+                    continue                       # 异常 / 被丢弃 → 不参选
+                if "procedure" not in ev["knowledge_type"]:
+                    continue                       # 只跑工序帖
+                n_img = len((r.get("post") or {}).get("images") or [])
+                if n_img > _IMG_SKIP_CAP:
+                    n_img_skip += 1                # 图太多, 拼图也压不住 → 放弃, 名额让给下一名
+                    continue
+                raw = ev["overall"] or 0.0
+                norm = raw / (ev["scale"] or 5)    # 归一化到 0-1, 跨 schema 可比
+                ck = _short_case(r.get("case_id") or "")
+                if ck not in pooled or norm > pooled[ck][0]:
+                    pooled[ck] = (norm, raw, qd.name, fkey, r)  # 同帖跨 form/q 撞分, 留最高
+    if n_img_skip:
+        print(f"   ⓘ 因图过多 (>{_IMG_SKIP_CAP} 张) 放弃 {n_img_skip} 个工序帖 (名额已顺延)")
+    ranked = sorted(pooled.items(), key=lambda kv: kv[1][0], reverse=True)
+    out: List[Tuple[str, str, Dict[str, Any], float, float]] = []
+    for ck, (norm, raw, qname, fkey, r) in ranked:
+        if ck in exclude_ids:
+            continue                           # 磁盘上已占
+        out.append((qname, fkey, r, norm, raw))
+        if len(out) >= top_n:
             break
     return out
 
 
-# ── 全局去重: 扫所有 q*/procedures/*/ 目录名 → short_case 占用集 ──────────────────
-
-def _gather_claimed(runs_full: Path) -> Dict[str, str]:
-    """扫所有 q*/procedures/*/ 返回 {short_case: rel_dir} —— 目录存在即视为占用,
-    无论它含 HTML 还是空、是否还在跑。多终端场景下, 一启动就 mkdir, 其他终端立刻见到。
-
-    key 解析顺序: ① _meta.json 里的 full case_id → _short_case 化
-                 ② 目录名 `{form}_{short_case}` 直接抽 short_case
-    """
-    claimed: Dict[str, str] = {}
-    if not runs_full.is_dir():
-        return claimed
-    for pd in runs_full.glob("q*/procedures/*"):
-        if not pd.is_dir():
-            continue
-        key = None
-        meta_path = pd / "_meta.json"
-        if meta_path.exists():
-            try:
-                full = json.loads(meta_path.read_text(encoding="utf-8")).get("case_id")
-                if full:
-                    key = _short_case(full)
-            except Exception:
-                pass
-        if not key:
-            parts = pd.name.split("_", 1)
-            key = parts[1] if len(parts) == 2 else pd.name
-        claimed[key] = pd.relative_to(runs_full).as_posix()
-    return claimed
-
-
-def _write_meta(out_dir: Path, case_id: str, from_q: str, form: str, score: float) -> None:
-    """_meta.json 记 full case_id + 出处 + 时间, 供其他终端/回查识别。"""
+# ── 累计目标: 扫 workflows/ 里属于这些 q 的「已完成并合并」帖 → short_case 集 ───────
+# top-N 当「累计目标」用: 本次预算 = N − 已完成数 (见 main)。配合跑时原子 mkdir 分流,
+# 多窗口跑同一条指令也只会瓜分到 N 条、不超额 (不排在跑/跑挂的, 那些交给 mkdir 撞锁)。
+
+def _gather_done(workflows_dir: Path, q_names: Set[str]) -> Set[str]:
+    """workflows/ 里属于 q_names 的已完成帖, 返回其 _short_case 集。
+    文件名形如 {qname}_{form}_{short}.json (build_workflows.write_one 落盘约定)。"""
+    done: Set[str] = set()
+    if not workflows_dir.is_dir():
+        return done
+    for qn in q_names:
+        for f in workflows_dir.glob(f"{qn}_*.json"):
+            folder = f.stem[len(qn) + 1:]          # 去 '{qname}_' → '{form}_{short}'
+            parts = folder.split("_", 1)
+            if len(parts) == 2:
+                done.add(parts[1])                 # short_case
+    return done
+
+
+def _write_meta(out_dir: Path, case_id: str, from_q: str, form: str,
+                score: float, score_norm: float) -> None:
+    """_meta.json 记 full case_id + 出处 + 时间, 供其他终端/回查识别。
+    score=原始 overall (量纲随 schema), score_norm=归一化 0-1 (跨 q 排名实际用的分)。"""
     out_dir.mkdir(parents=True, exist_ok=True)
     (out_dir / "_meta.json").write_text(json.dumps({
-        "case_id": case_id, "from_q": from_q, "form": form, "score": round(score, 4),
+        "case_id": case_id, "from_q": from_q, "form": form,
+        "score": round(score, 4), "score_norm": round(score_norm, 4),
         "started_at": datetime.now().isoformat(timespec="seconds"),
     }, ensure_ascii=False, indent=2), encoding="utf-8")
 
@@ -224,7 +199,8 @@ def main() -> None:
                    help="离散指定 q (如 '36' / 'q0036' / '36,55,720')")
     p.add_argument("--start", type=int, default=0, help="区间起始 idx (默认 0)")
     p.add_argument("--count", type=int, default=10, help="区间取几条 q (默认 10)")
-    p.add_argument("--top-k", type=int, default=1, help="每 form 取几条 (默认 1)")
+    p.add_argument("--top-n", "--top-k", dest="top_n", type=int, default=20,
+                   help="每个 q 合池后取评分前几条 (默认 20; --top-k 为兼容别名)")
     p.add_argument("--model", default="claude-sonnet-4-6", help="透传给 run_procedure_dsl")
     p.add_argument("--max-turns", type=int, default=300, help="透传给 run_procedure_dsl")
     p.add_argument("--max-retries", type=int, default=3, help="透传给 run_procedure_dsl")
@@ -253,100 +229,100 @@ def main() -> None:
         q_dirs = all_q[args.start : args.start + args.count]
         if not q_dirs:
             sys.exit(f"❌ 区间 [{args.start}:{args.start+args.count}] 取不到 q")
-    # ── 启动前: 全局扫描所有 q*/procedures/*/ 目录名占用集 ────────────────────
-    global_claimed = {} if args.force else _gather_claimed(runs_full)
-    print(f"📋 {len(q_dirs)} 个 q, top-{args.top_k}/form, "
-          f"目标 ≤{len(q_dirs) * 3 * args.top_k} 条工序 (去重后会少)")
-    print(f"📦 全局占用 case 数 = {len(global_claimed)}"
-          f"{' (--force 已忽略)' if args.force else ''}")
-
-    n_ran = n_skipped = n_failed = n_no_cap = n_global_skip = 0
-    # 本进程已规划的 case_id (即使还没开跑也算"占用", 防同进程后续 q 重复选)
-    local_claimed: Set[str] = set()
-
-    for qd in q_dirs:
-        forms = sorted(qd.glob("form_*.json"))
-        if not forms:
-            print(f"\n⏭️  {qd.name}: 无 form_*.json")
-            continue
-        print(f"\n▶ {qd.name}")
-
-        # exclude = 全局已占用 (磁盘上 procedure 目录存在的) + 本进程已规划 (跨 q 也屏蔽)
-        # 所有键统一 _short_case 形式 (跟目录名约定一致)
-        excluded: Set[str] = set(global_claimed) | local_claimed
-        plan: List[Tuple[str, Dict[str, Any], float]] = []
-        for fp in forms:
-            fkey = fp.stem.replace("form_", "")    # A/B/C
-            picks = _pick_candidates(fp, excluded, args.top_k)
-            for score, r in picks:
-                cid_full = r.get("case_id") or ""
-                cid_key = _short_case(cid_full)
-                plan.append((fkey, r, score))
-                excluded.add(cid_key)        # 本 q 内其他 form 不会再选
-                local_claimed.add(cid_key)   # 后续 q 也不再选
-                title = (r.get("post") or {}).get("title", "")[:30]
-                print(f"   [{fkey}] score={score:.2f} {cid_full[:24]} {title}")
-            if len(picks) < args.top_k:
-                # 想取 top-k 但被 excluded 顶替到不够 → 报一行说明
-                n_global_skip += args.top_k - len(picks)
-                print(f"   [{fkey}] 候选不足 top-{args.top_k} (其他 q 已占了同帖)")
-
-        if not plan:
-            print(f"   ⚠️ {qd.name} 无 工序 类候选 (无新 case 或都被 全局去重 顶了)")
-            n_no_cap += 1
+    # ── 累计目标: top-N 是「workflows/ 里这些 q 最终要有 N 条」, 本次只补差额 ──────
+    q_names = {d.name for d in q_dirs}
+    workflows_dir = Path(build_workflows.OUT_DIR)
+    done_set: Set[str] = set() if args.force else _gather_done(workflows_dir, q_names)
+    budget = args.top_n if args.force else max(0, args.top_n - len(done_set))
+    print(f"📋 {len(q_dirs)} 个 q 合池, 累计目标 top-{args.top_n}; "
+          f"workflows/ 已有 {len(done_set)} 条 → 本次还需 {budget} 条"
+          f"{' (--force 忽略已完成)' if args.force else ''}")
+
+    n_ran = n_skipped = n_failed = 0
+
+    if budget == 0:
+        print("\n✅ 累计目标已达成 (workflows/ 里这些 q 已够 top-N), 无需再跑.")
+        print(f"\n{'='*60}\n📊 完成: 跑 0 / 已存跳过 0 / 失败 0")
+        return
+
+    # ── 跨 q 合池挑 top-budget ──────────────────────────────────────────────────
+    # 只排除「已完成」(done_set); 在跑/跑挂的留在池里, 跑时靠原子 mkdir 占位分流到各窗口。
+    plan = _pick_pool_candidates(q_dirs, done_set, budget)
+
+    if not plan:
+        print("\n⚠️ 这批 q 合池后无可跑候选 (全被丢弃/异常过滤, 或都已完成)")
+        print(f"\n{'='*60}\n📊 完成: 跑 0 / 已存跳过 0 / 失败 0")
+        return
+
+    print(f"\n▶ 本次挑出 {len(plan)} 条 (补差额, score=归一化 0-1, 括号内原始/满分):")
+    for qname, fkey, r, norm, raw in plan:
+        cid_full = r.get("case_id") or ""
+        title = (r.get("post") or {}).get("title", "")[:30]
+        print(f"   [{qname} {fkey}] score={norm:.3f} ({raw:g}) {cid_full[:24]} {title}")
+    if len(plan) < budget:
+        # 池里够格的工序帖不足本次预算 (被丢弃过滤掉 / 已完成)
+        print(f"   ⓘ 可跑候选不足本次预算 {budget} (仅 {len(plan)} 条; 想多跑就加 q 或调大 --top-n)")
+
+    # ── 落产物 + 跑; 每跑完一条就立刻把它的合并 json 写进 workflows/ ──────────────
+    def _emit_workflow(qname: str, folder: str) -> None:
+        """单条 procedure 跑完 (或已存) → 立刻写出 workflows/{qname}_{folder}.json。
+        (dry-run 不产 workflow.json 故跳过; build 失败不应中断批量, 故 try 包住)"""
+        if args.dry_run:
+            return
+        try:
+            if build_workflows.write_one(qname, folder, runs_dir=runs_full):
+                print(f"   🧩 workflows/: {qname}_{folder}.json 已写")
+        except Exception as e:
+            print(f"   ⚠️ build_workflows 失败 ({qname}/{folder}): {e}")
+
+    for qname, fkey, r, norm, raw in plan:
+        cid = r.get("case_id") or "unknown"
+        name = f"{fkey}_{_short_case(cid)}"
+        out_dir = runs_full / qname / "procedures" / name
+
+        # dry-run 绝不碰磁盘 (建目录会污染状态), 只报会不会跑
+        if args.dry_run:
+            mark = ("已完成" if _already_done(out_dir)
+                    else "已占(在跑/跑挂)" if out_dir.exists() else "would run")
+            print(f"   [dry-run] {mark}: {qname}/{name}/")
             continue
 
-        # 落产物
-        proc_root = qd / "procedures"
-        proc_root.mkdir(exist_ok=True)
-        for fkey, r, score in plan:
-            cid = r.get("case_id") or "unknown"
-            name = f"{fkey}_{_short_case(cid)}"
-            out_dir = proc_root / name
-
-            if _already_done(out_dir) and not args.force:
-                print(f"   ⏭️  {name}/ 已存在 case-*.html (用 --force 覆盖)")
-                n_skipped += 1
-                continue
-
-            # 写 _source.json (DSL 输入) + _meta.json (case_id 索引, 供其他终端识别)
-            out_dir.mkdir(parents=True, exist_ok=True)
-            src_path = out_dir / "_source.json"
-            src_path.write_text(
-                json.dumps(_source_to_dsl_input(r), ensure_ascii=False, indent=2),
-                encoding="utf-8")
-
-            if args.dry_run:
-                print(f"   [dry-run] would run on {src_path.name} → {out_dir.name}/")
-                continue
-
-            # _meta.json 即占位标识: 其他终端扫到 out_dir/_meta.json 就视作 已占
-            _write_meta(out_dir, case_id=cid, from_q=qd.name, form=fkey, score=score)
-            try:
-                code = run_one(src_path, out_dir, args)
-            except KeyboardInterrupt:
-                print(f"\n⚠️  Ctrl-C 中断, {name}/ 保留 (含 _meta.json), 下次跑会跳过这条; 想复跑用 --force 或手动删 dir")
-                raise
-            if code == 0:
-                n_ran += 1
-                global_claimed[_short_case(cid)] = (out_dir.relative_to(runs_full)).as_posix()
-                print(f"   ✓ {name} done")
+        # ── 原子认领: mkdir(exist_ok=False) 内核级唯一, 谁先建谁跑 ──────────────
+        # 多窗口跑同一指令时, 同一条 procedure 只有一个窗口能建成目录、其余撞 FileExistsError
+        # → 跳过 (那条已被别窗口领走 / 或是历史跑挂的残目录)。--force 时允许覆盖已存。
+        try:
+            out_dir.mkdir(parents=True, exist_ok=args.force)
+        except FileExistsError:
+            if _already_done(out_dir):
+                _emit_workflow(qname, name)   # 跑完但没合并的, 顺手补进 workflows/
+                print(f"   ⏭️  {qname}/{name}/ 已完成, 跳过")
             else:
-                n_failed += 1
-                print(f"   ❌ {name} exit={code} (目录已占; 想重试 --force 或 手动 rm 该 dir)")
+                print(f"   ⏭️  {qname}/{name}/ 已被别窗口认领 / 或历史跑挂, 跳过")
+            n_skipped += 1
+            continue
 
-        # 本 q 的 procedure 跑完 → 把合并 json 增量写进 workflows/
-        # (dry-run 不产 workflow.json, 跳过; build 失败不应中断批量, 故 try 包住)
-        if not args.dry_run:
-            try:
-                n_wf = build_workflows.write_run(qd.name, runs_dir=runs_full)
-                print(f"   🧩 workflows/: {qd.name} 写出 {n_wf} 个合并 json")
-            except Exception as e:
-                print(f"   ⚠️ build_workflows 失败 ({qd.name}): {e}")
+        # 认领成功 (此刻 out_dir 归本窗口) → 写 _source.json + _meta.json
+        src_path = out_dir / "_source.json"
+        src_path.write_text(
+            json.dumps(_source_to_dsl_input(r), ensure_ascii=False, indent=2),
+            encoding="utf-8")
+        _write_meta(out_dir, case_id=cid, from_q=qname, form=fkey, score=raw, score_norm=norm)
+        try:
+            code = run_one(src_path, out_dir, args)
+        except KeyboardInterrupt:
+            print(f"\n⚠️  Ctrl-C 中断, {qname}/{name}/ 保留 (含 _meta.json), 下次跑会跳过这条; 想复跑用 --force 或手动删 dir")
+            raise
+        if code == 0:
+            n_ran += 1
+            print(f"   ✓ {qname}/{name} done")
+            _emit_workflow(qname, name)   # 跑完一个就出一个, 不等同 q 其他帖子
+        else:
+            n_failed += 1
+            print(f"   ❌ {qname}/{name} exit={code} (目录已占; 想重试 --force 或 手动 rm 该 dir)")
 
     print(f"\n{'='*60}")
-    print(f"📊 完成: 跑 {n_ran} / 本次已存目录跳过 {n_skipped} / 全局占用顶替 {n_global_skip} / "
-          f"失败 {n_failed} / 无候选 q {n_no_cap}")
+    print(f"📊 完成: 跑 {n_ran} / 跳过(已完成或别窗口领走) {n_skipped} / 失败 {n_failed} "
+          f"(累计目标 top-{args.top_n}, 本次预算 {budget}, 选中 {len(plan)})")
 
 
 if __name__ == "__main__":

examples/process_pipeline/script/search_eval/build_workflows.py

@@ -49,14 +49,16 @@ def load_json(path):
         return json.load(f)
 
 
-def build_run(run_id, runs_dir=None):
+def build_run(run_id, runs_dir=None, only_folder=None):
     """为单个 run 目录构建合并结果。
 
     以帖子为单位：返回一个 list，每个元素是 (folder, merged_dict)，
     merged_dict 即单个帖子的合并 JSON（query + post + llm_evaluation + workflow）。
     找不到 procedures 时返回空 list。
     runs_dir 缺省用模块的 RUNS_DIR；外部脚本（如 batch_extract_procedures.py 用了
-    --output-dir）可传入自己的 runs_full，避免两边路径不一致。"""
+    --output-dir）可传入自己的 runs_full，避免两边路径不一致。
+    only_folder 非空时只处理那一个 folder（其余直接跳过、不刷 warn）——给 write_one
+    用，避免同 q 里别的未完成 folder 每次都被扫一遍刷一堆 warn。"""
     run_dir = os.path.join(runs_dir or RUNS_DIR, run_id)
     proc_root = os.path.join(run_dir, "procedures")
     if not os.path.isdir(proc_root):
@@ -80,6 +82,8 @@ def build_run(run_id, runs_dir=None):
 
     out = []
     for folder in sorted(os.listdir(proc_root)):
+        if only_folder and folder != only_folder:
+            continue                       # write_one 只要这一个, 其余静默跳过
         folder_path = os.path.join(proc_root, folder)
         if not os.path.isdir(folder_path):
             continue
@@ -128,19 +132,36 @@ def build_run(run_id, runs_dir=None):
     return out
 
 
+def _dump_merged(run_id, folder, merged):
+    """把单个帖子的 merged 写成 workflows/{run_id}_{folder}.json。"""
+    os.makedirs(OUT_DIR, exist_ok=True)
+    out_path = os.path.join(OUT_DIR, f"{run_id}_{folder}.json")
+    with open(out_path, "w", encoding="utf-8") as f:
+        json.dump(merged, f, ensure_ascii=False, indent=2)
+    print(f"[ok] {run_id}/{folder} -> {os.path.basename(out_path)}")
+
+
 def write_run(run_id, runs_dir=None):
     entries = build_run(run_id, runs_dir=runs_dir)
     if not entries:
         return 0
-    os.makedirs(OUT_DIR, exist_ok=True)
     for folder, merged in entries:
-        out_path = os.path.join(OUT_DIR, f"{run_id}_{folder}.json")
-        with open(out_path, "w", encoding="utf-8") as f:
-            json.dump(merged, f, ensure_ascii=False, indent=2)
-        print(f"[ok] {run_id}/{folder} -> {os.path.basename(out_path)}")
+        _dump_merged(run_id, folder, merged)
     return len(entries)
 
 
+def write_one(run_id, folder, runs_dir=None):
+    """只把指定 procedure folder 的合并 json 写出 —— 跑完一个工序就立刻出一个,
+    无需等同 q 其他帖子。命中并写出返回 1, 没匹配到 (缺 workflow.json 等) 返回 0。
+    复用 build_run 的解析/校验逻辑, 只处理 folder 那一个 (only_folder 让同 q 其余
+    未完成目录被静默跳过, 不刷 warn)。"""
+    for f, merged in build_run(run_id, runs_dir=runs_dir, only_folder=folder):
+        if f == folder:
+            _dump_merged(run_id, folder, merged)
+            return 1
+    return 0
+
+
 def build_runs(run_ids):
     """对一组 run 执行 write_run，返回写出的帖子 json 总数。"""
     total = 0

examples/process_pipeline/script/search_eval/index.html

@@ -424,6 +424,31 @@
       box-shadow: 0 28px 90px rgba(0, 0, 0, .25);
     }
 
+    /* 全屏模式: 占满视口 + 改用 flex 列布局, 让内容面板/工序 iframe 撑满 */
+    dialog.fullscreen {
+      width: 100vw;
+      height: 100vh;
+      max-width: none;
+      max-height: none;
+      border: none;
+      border-radius: 0;
+      display: flex;
+      flex-direction: column;
+    }
+    dialog.fullscreen .modal-head,
+    dialog.fullscreen .modal-tabs {
+      flex: 0 0 auto;            /* 头部/标签固定高, 不被压缩 */
+    }
+    dialog.fullscreen .modal-content,
+    dialog.fullscreen #modalContentProcedure {
+      flex: 1 1 auto;           /* 只有内容面板撑满剩余高度 */
+      min-height: 0;
+      overflow: auto;
+    }
+    dialog.fullscreen #modalContentProcedure {
+      height: auto !important;   /* 覆盖内联的固定 600px, 改由 flex 撑满 */
+    }
+
     dialog::backdrop {
       background: rgba(38, 33, 27, .42);
     }
@@ -1451,7 +1476,10 @@
         <div id="modalMeta" class="meta"></div>
         <h3 id="modalTitle"></h3>
       </div>
-      <button onclick="detailDialog.close()">关闭</button>
+      <div style="display: flex; gap: 8px; flex-shrink: 0;">
+        <button id="fullscreenBtn" onclick="toggleFullscreen()">全屏</button>
+        <button onclick="detailDialog.close()">关闭</button>
+      </div>
     </div>
     
     <div class="modal-tabs" id="modalTabs" style="display: none;">
@@ -2346,9 +2374,23 @@
       }
       
       switchModalTab('detail');
+      applyFullscreenPref();          // 沿用上次的全屏/窗口偏好
       detailDialog.showModal();
     }
 
+    // 全屏开关: 切 .fullscreen class + 记住偏好 (localStorage), 下次开 case 详情沿用
+    function applyFullscreenPref() {
+      const on = localStorage.getItem("caseDetailFullscreen") === "1";
+      detailDialog.classList.toggle("fullscreen", on);
+      const btn = document.getElementById("fullscreenBtn");
+      if (btn) btn.textContent = on ? "退出全屏" : "全屏";
+    }
+    function toggleFullscreen() {
+      const on = !detailDialog.classList.contains("fullscreen");
+      localStorage.setItem("caseDetailFullscreen", on ? "1" : "0");
+      applyFullscreenPref();
+    }
+
     let currentPinnedScoreEl = null;
     function pinScoreReason(el, label, k) {
       const activeIdx = detailDialog.dataset.activeIdx;

examples/process_pipeline/script/search_eval/procedure-dsl/run_procedure_dsl.py

@@ -116,6 +116,14 @@ _MEDIA_TYPE = {
 # 同一 URL 多次跑 / resume 不重复下载. 不影响 SDK / OAuth, 纯本地 IO.
 _IMAGE_CACHE_DIR = DSL_ROOT / ".image_cache"
 
+# ── 多图压缩策略 (防止几十张图把单轮 output 撑破 32K/64K 上限) ──────────────────
+# 阶梯: 图少 → 单图只降分辨率; 图多 (>OVER) → 拼 3×3 九宫格把"图块数"压 9:1, 模型逐图
+# 铺陈的冲动小了, 骨架/output 也跟着小. (跳过极多图帖那一档在 batch 选片侧做, 见
+# batch_extract_procedures.py 的 _IMG_SKIP_CAP — 那边知道图数又控制 top-N 名额回填.)
+_IMG_MAX_EDGE = 1024      # 单图降分辨率: 长边上限 px
+_IMG_MONTAGE_OVER = 18    # 图数 > 此值 → 拼九宫格 (与 batch 选片阈值需各自维护, 语义不同)
+_IMG_GRID_CELL = 512      # 九宫格每格 px → 3×3 拼图长边 1536, 压在 Anthropic 1568 上限内
+
 # 浏览器 UA, 避免 mmbiz / 其他图床对裸 python-requests 默认 UA 直接 403.
 _DOWNLOAD_UA = (
     "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 "
@@ -151,53 +159,170 @@ def _url_to_cached_path(url: str, timeout: float = 15.0) -> Path:
     return local
 
 
-def _append_image_blocks(blocks: List[Dict[str, Any]], images: List[str]) -> None:
-    """循环把 images 转成 content block 加到 blocks. 单张失败不阻塞整批.
+def _resolve_local(ref: str) -> Path:
+    """图 ref → 本地路径: URL 先下载进缓存 (绕 robots.txt), 本地路径校验存在。"""
+    if ref.startswith(("http://", "https://")):
+        return _url_to_cached_path(ref)
+    p = Path(ref).expanduser().resolve()
+    if not p.exists():
+        raise FileNotFoundError(f"image not found: {ref}")
+    return p
 
-    base / variant driver 共用 — 保证 image 容错策略一致.
-    """
-    n_ok, n_fail = 0, 0
+
+def _pil_to_block(im: Any) -> Dict[str, Any]:
+    """PIL.Image → Anthropic base64 content block (统一 JPEG q90, 体积可控)。"""
+    import io
+    buf = io.BytesIO()
+    im.save(buf, format="JPEG", quality=90)
+    return {
+        "type": "image",
+        "source": {"type": "base64", "media_type": "image/jpeg",
+                   "data": base64.standard_b64encode(buf.getvalue()).decode()},
+    }
+
+
+_OCR_ENGINE: Any = None
+_OCR_UNAVAILABLE = False
+
+
+def _ocr_image(im: Any) -> str:
+    """对一张 PIL.Image 做全分辨率 OCR, 返回拼接后的文字 (失败/无字 → 空串)。
+    九宫格会把图压到 512px 小字糊掉, 故在缩放前先 OCR 把文字走文本通道保住。
+    RapidOCR 未装 → 整体禁用 OCR (只此一次, 之后不再尝试), 退化成纯拼图。"""
+    global _OCR_ENGINE, _OCR_UNAVAILABLE
+    if _OCR_UNAVAILABLE:
+        return ""
+    try:
+        if _OCR_ENGINE is None:
+            from rapidocr_onnxruntime import RapidOCR
+            _OCR_ENGINE = RapidOCR()
+        import numpy as np
+        result, _ = _OCR_ENGINE(np.array(im))
+        if not result:
+            return ""
+        # result: [[box, text, score], ...] 按检测顺序 (大致从上到下) → 直接拼
+        return " ".join(seg[1] for seg in result if seg and len(seg) >= 2 and seg[1]).strip()
+    except ImportError:
+        _OCR_UNAVAILABLE = True
+        print("[ocr] ⚠️ rapidocr 未装, 本次禁用 OCR (九宫格仅拼图无文字旁挂)", flush=True)
+        return ""
+    except Exception as e:
+        print(f"[ocr] skip 1 张 ({type(e).__name__}: {e})", flush=True)
+        return ""
+
+
+def _format_ocr_sidecar(ocr_texts: List[str]) -> str:
+    """把每图 OCR 文字按"拼图#·格#"标注成一个文本块 (跟九宫格阅读顺序对齐)。
+    全空 → 返回空串 (不加块)。i 从 0 起: grid=i//9, cell=i%9 (逐行 左→右 上→下)。"""
+    lines = []
+    for i, t in enumerate(ocr_texts):
+        if not t:
+            continue
+        lines.append(f"■ 拼图{i // 9 + 1}·格{i % 9 + 1}: {t}")
+    if not lines:
+        return ""
+    return ("【配图 OCR 文字】图已压成九宫格, 小字可能糊; 以下是各格**全分辨率** OCR 原文, "
+            "与拼图位置一一对应 (格号即拼图内 从左到右、从上到下 的序), 用它补全图里看不清的"
+            "菜单/参数/prompt 文字:\n" + "\n".join(lines))
+
+
+def _montage(pil_imgs: List[Any], cell: int) -> List[Any]:
+    """把图按每 9 张拼成一张 3×3 九宫格 (每格 cell px, 白底居中)。返回拼图列表。
+    最后一组不足 9 张, 余下格子留白。阅读顺序: 逐行 从左到右、从上到下 (封面=第一张拼图左上)。"""
+    from PIL import Image
+    grids: List[Any] = []
+    for i in range(0, len(pil_imgs), 9):
+        canvas = Image.new("RGB", (cell * 3, cell * 3), (255, 255, 255))
+        for j, im in enumerate(pil_imgs[i:i + 9]):
+            t = im.copy()
+            t.thumbnail((cell, cell))                      # 等比缩进格子
+            row, col = divmod(j, 3)
+            canvas.paste(t, (col * cell + (cell - t.width) // 2,
+                             row * cell + (cell - t.height) // 2))
+        grids.append(canvas)
+    return grids
+
+
+def _append_image_blocks_raw(blocks: List[Dict[str, Any]], images: List[str]) -> None:
+    """回落路径 (PIL 不可用时): 逐图原样 base64, 不降分辨率/不拼图。"""
+    n_ok = n_fail = 0
     for ref in images:
         try:
-            blocks.append(_image_block(ref))
+            local = _resolve_local(ref)
+            blocks.append({
+                "type": "image",
+                "source": {"type": "base64",
+                           "media_type": _MEDIA_TYPE.get(local.suffix.lower(), "image/png"),
+                           "data": base64.standard_b64encode(local.read_bytes()).decode()},
+            })
             n_ok += 1
         except Exception as e:
             n_fail += 1
             print(f"[image] skip {ref[:80]}... ({type(e).__name__}: {e})", flush=True)
     if images:
-        print(f"[image] {n_ok}/{len(images)} 成功下载并 base64 化, {n_fail} 失败已跳过", flush=True)
+        print(f"[image] (无 PIL) {n_ok}/{len(images)} 原样 base64, {n_fail} 失败跳过", flush=True)
 
 
-def _image_block(ref: str) -> Dict[str, Any]:
-    """图 ref → Anthropic content block.
+def _append_image_blocks(blocks: List[Dict[str, Any]], images: List[str]) -> None:
+    """把 images 转成 content block 加到 blocks，按阶梯压缩 (见 _IMG_* 常量):
+       ≤OVER 张 → 逐图降分辨率单发; >OVER 张 → 拼 3×3 九宫格减块数。
+    单张下载/解码失败不阻塞整批。PIL 不可用 → 回落原样 base64。
+    base / variant driver 共用 — 保证 image 容错策略一致。"""
+    if not images:
+        return
+    try:
+        from PIL import Image
+    except Exception:
+        _append_image_blocks_raw(blocks, images)
+        return
 
-    URL → 客户端下载后 base64 (不再走 server-side fetch, 绕开 robots.txt 限制).
-    本地路径 → 直接 base64.
-    下载失败 (网络问题 / 404 等) → raise, 让上层决定跳过还是阻塞.
-    """
-    if ref.startswith(("http://", "https://")):
-        local = _url_to_cached_path(ref)
-        data = base64.standard_b64encode(local.read_bytes()).decode()
-        return {
-            "type": "image",
-            "source": {
-                "type": "base64",
-                "media_type": _MEDIA_TYPE.get(local.suffix.lower(), "image/png"),
-                "data": data,
-            },
-        }
-    p = Path(ref).expanduser().resolve()
-    if not p.exists():
-        raise FileNotFoundError(f"image not found: {ref}")
-    data = base64.standard_b64encode(p.read_bytes()).decode()
-    return {
-        "type": "image",
-        "source": {
-            "type": "base64",
-            "media_type": _MEDIA_TYPE.get(p.suffix.lower(), "image/png"),
-            "data": data,
-        },
-    }
+    pil_imgs: List[Any] = []
+    n_fail = 0
+    for ref in images:
+        try:
+            pil_imgs.append(Image.open(_resolve_local(ref)).convert("RGB"))
+        except Exception as e:
+            n_fail += 1
+            print(f"[image] skip {ref[:80]}... ({type(e).__name__}: {e})", flush=True)
+    n = len(pil_imgs)
+    if n == 0:
+        print(f"[image] 0/{len(images)} 可用 ({n_fail} 失败)", flush=True)
+        return
+
+    if n > _IMG_MONTAGE_OVER:
+        ocr_texts = [_ocr_image(im) for im in pil_imgs]   # 缩放前全分辨率 OCR 保文字
+        grids = _montage(pil_imgs, _IMG_GRID_CELL)
+        for g in grids:
+            blocks.append(_pil_to_block(g))
+        sidecar = _format_ocr_sidecar(ocr_texts)
+        if sidecar:
+            blocks.append({"type": "text", "text": sidecar})
+        n_ocr = sum(1 for t in ocr_texts if t)
+        print(f"[image] {n} 张 > {_IMG_MONTAGE_OVER} → 拼九宫格 {len(grids)} 张 "
+              f"(每张≤9图·格{_IMG_GRID_CELL}px) + {n_ocr} 格 OCR 文字旁挂, "
+              f"{n_fail} 失败跳过", flush=True)
+    else:
+        for im in pil_imgs:
+            im.thumbnail((_IMG_MAX_EDGE, _IMG_MAX_EDGE))   # 等比降分辨率 (小图不放大)
+            blocks.append(_pil_to_block(im))
+        print(f"[image] {n} 张单图 (降分辨率≤{_IMG_MAX_EDGE}px), {n_fail} 失败跳过", flush=True)
+
+
+def _image_prompt_note(n_images: int) -> str:
+    """配图行文案: 跟 _append_image_blocks 的阶梯保持一致, 拼图时务必告诉模型
+    它收到的是九宫格 (否则会以为有 N 张独立图, 引用错位 / 逐图铺陈撑爆 output)。"""
+    if n_images == 0:
+        return "本消息未附图."
+    if n_images > _IMG_MONTAGE_OVER:
+        import math
+        n_grid = math.ceil(n_images / 9)
+        return (f"本消息附了 {n_grid} 张【九宫格拼图】—— 由原始 {n_images} 张配图按 3×3 网格拼接 "
+                f"(每张最多含 9 个子图, 已统一降分辨率)。原图过多, 拼图是为压 token; "
+                f"分析时把**每个格子当一张独立配图**看, 阅读顺序: 每张拼图内 从左到右、从上到下逐行; "
+                f"封面图在第一张拼图的左上格 (老格式 image_type=2 / 新格式原序第一张)。"
+                f"图后另附一个【配图 OCR 文字】文本块 (各格全分辨率 OCR), 拼图里小字糊掉时**以 OCR 文字为准**.")
+    return (f"本消息附了 {n_images} 张图作多模态内容 (已降分辨率; URL 抽自老格式的 image_url_list "
+            f"或新格式的 images; 老格式 image_type=2 封面排最前, 新格式按原序).")
 
 
 # ──── source JSON → 自动拉 image_url_list ────────────────────────────────────
@@ -297,7 +422,7 @@ def _build_initial_blocks(
     · 老格式 (case-*-raw.json): {{title, link, body_text, image_url_list:[{{image_type,image_url}}], publish_timestamp, channel_account_name}}
     · 新格式 (eval_case-*.json): {{title, link, body_text, images:[url_str], videos, channel, content_type, like_count, publish_timestamp, channel_content_id, ...}}
   正文 (body_text) 和元数据从此文件 Read; 两种 schema 的 body_text 含义一致.
-- 配图: 本消息附了 {len(images)} 张图作多模态内容 (URL 抽自老格式的 image_url_list 或新格式的 images; 老格式 image_type=2 封面排最前, 新格式按原序).
+- 配图: {_image_prompt_note(len(images))}
 
 ## 输出目录
 
@@ -623,6 +748,9 @@ Do not output raw dictionary structure or any other nesting. Do not touch or mod
             "ANTHROPIC_API_KEY": "",
             "ANTHROPIC_BASE_URL": "",
             "ANTHROPIC_AUTH_TOKEN": "",
+            # 多图帖 (几十张图) 的提取响应容易顶破默认 32000 output token 上限 →
+            # 调到 sonnet 4.x 的出参天花板 64000。外部已设同名环境变量则以这里为准。
+            "CLAUDE_CODE_MAX_OUTPUT_TOKENS": "64000",
         },
         stderr=_capture_stderr,
     )

examples/process_pipeline/script/search_eval/procedure-dsl/spec/taxonomy/type_suggestions.md

@@ -195,3 +195,115 @@
 - `解码配置`: 控制解码行为的参数集合，包含temperature缩放系数、logits_bias词级偏置、repetition_penalty重复惩罚系数和bad_words_ids屏蔽列表  (来自 case-C_xhs_69ee4674, extends `生成参数`)
 - `候选Token列表`: 经Top-k/Top-p截断后保留的候选token子集，形状仍[vocab_size,]但非候选位置logits被置为负无穷，等待最终采样决策  (来自 case-C_xhs_69ee4674, extends `样图`)
 - `输出Token`: 经采样或argmax决策后产出的第一个生成token，对应词表中某个整数ID及其解码后的子词或字符，是本次推理的最终交付物  (来自 case-C_xhs_69ee4674, extends `正文`)
+
+- `关键词列表`: 按核心词、场景词、问题词三类分类整理的关键词条目集合，用于后续嵌入文章结构的指定位置  (来自 case-creative_1, extends `知识库`)
+- `含关键词骨架`: 在文章章节骨架基础上标注了关键词嵌入位置规则（标题含核心词、首段前三行融入核心词、正文每300字一个场景词、小标题含问题词）的结构化内容规划  (来自 case-creative_1, extends `大纲`)
+- `文章草稿`: 按价值密度、表达节奏、视觉辅助三维准则展开填充的完整正文中间态，含加粗重点、案例数据、设问、配图说明，尚未添加互动钩子，需后续步骤继续处置  (来自 case-creative_1, extends `脚本`)
+
+- `写作触发`: 触发写作冲动的灵感事件或感受，如读书感受、被某篇文章触动、某事件引发的感慨与写作冲动  (来自 case-creative_2, extends `描述`)
+- `账号定位`: 公众号账号的写作定位与目标受众范围，是跨批次复用的账号策略约束（如：公众号运营、写作成长、退休生活等领域）  (来自 case-creative_2, extends `描述`)
+- `选题方向`: 经爆款潜力与定位符合双重筛选后确定的粗粒度写作话题方向，尚未细化为具体角度  (来自 case-creative_2, extends `描述`)
+- `竞品数据`: 在公众号、知乎、百度等平台搜索同话题历史文章所得的阅读量数据，用于验证选题是否具有读者基础  (来自 case-creative_2, extends `知识库`)
+- `选题`: 经阅读量验证与核心卖点确认后的具体写作选题角度，比选题方向更精细、可直接指导创作  (来自 case-creative_2, extends `描述`)
+- `素材`: 从公众号、知乎、百度及书籍中一次性收集的、支撑各分论点的观点、论据、案例与数据集合  (来自 case-creative_2, extends `知识库`)
+- `草稿`: 包含开头、各分论点展开、段落过渡与结尾的完整文章初稿，已具备正文形态但尚未定稿，需继续修改打磨后才能发布  (来自 case-creative_2, extends `正文`)
+
+- `关键词`: 用于搜索或检索的单个关键词，作为信息获取步骤的查询输入，指向某一主题领域（如「减肥」）  (来自 case-creative_3, extends `提示词`)
+- `竞品内容`: 同行账号在抖音/小红书等平台发布的高播放量内容，作为被研究、借鉴打法的整体范本（含评论区）  (来自 case-creative_3, extends `对标内容`)
+- `痛点列表`: 从单一渠道（评论区或社群）收集到的用户痛点陈述条目列表，各条目独立成行，是进一步汇聚的原始素材  (来自 case-creative_3, extends `大纲`)
+- `社群`: 创作者所在的行业微信群/QQ群等社群，作为用户真实吐槽与痛点的信息采集渠道  (来自 case-creative_3, extends `知识库`)
+- `关键词列表`: 搜索引擎/平台输入关键词后出现的下拉联想词条目列表，反映用户真实搜索意图与痛点  (来自 case-creative_3, extends `大纲`)
+- `痛点清单`: 汇聚自评论区、社群、搜索下拉词三个渠道的用户痛点问题汇总清单，是选题筛选的直接依据  (来自 case-creative_3, extends `大纲`)
+- `选题问题`: 从痛点清单中提炼出的核心问题表述（通常为一个问句），是后续标题创作与内容策划的锚点  (来自 case-creative_3, extends `描述`)
+- `人设规格`: 创作者在短视频平台的角色定位描述，包含职业身份、专业角度与表达风格（如「营养师（专家型）——从代谢角度分析」）  (来自 case-creative_3, extends `描述`)
+- `选题方案`: 结合人设视角对钩子标题进行匹配调整后输出的完整选题设计方案，包含人设切入角度与具体标题  (来自 case-creative_3, extends `描述`)
+- `内容素材`: 支撑选题内容生产的多类型素材集合，包含真实案例故事、权威数据来源与证据截图等，供脚本/正文撰写取用  (来自 case-creative_3, extends `知识库`)
+
+- `需求信号列表`: 由多渠道并行采集汇聚的需求信号集合，包含搜索热点词、同行爆款样本、评论区痛点、头部创作者选题规律等，供后续选题转化步骤消费  (来自 case-creative_4, extends `知识库`)
+- `热搜词`: 从搜索平台检索得到的与内容方向相关的热搜关键词列表，含热度趋势标注，反映当前平台用户搜索需求  (来自 case-creative_4, extends `知识库`)
+- `用户反馈`: 从高互动内容评论区提取的用户高频痛点与诉求聚类，反映目标受众的真实需求与困惑  (来自 case-creative_4, extends `知识库`)
+- `规律总结`: 通过研究头部创作者内容提炼出的选题规律与打法模式，如常用结构组合、高频内容类型、标题惯用套路等  (来自 case-creative_4, extends `知识库`)
+- `候选选题`: 基于选题框架批量生成的选题候选列表，每条注明所用框架，作为待进一步优化打磨的选题骨架  (来自 case-creative_4, extends `大纲`)
+- `选题`: 经优化打磨后具备高传播潜力的内容选题，含差异化亮点标注，作为内容创作的最终选定方向  (来自 case-creative_4, extends `标题`)
+
+- `题目图片`: 包含数学题目内容（公式+图形+文字说明）的截图，用作 GPT Image 2.0 的视觉理解输入  (来自 case-C_gzh_8a75847e, extends `截图`)
+- `课文套组`: 由同一提示词一次生成的两张配套成品图：课文学习海报（知识板块+情境插画）和同步练习卷（分栏题型+分值标注），用于课文教学整套配套  (来自 case-C_gzh_8a75847e, extends `成品图`)
+- `词卡系列`: 批量生成的多张单词卡片集合，每张风格迥异、互不雷同；单次超级提示词产出，各卡自包含单词、音标/读法、中文释义和视觉场景  (来自 case-C_gzh_8a75847e, extends `成品图`)
+
+- `图片策划序列`: LLM输出的N个结构化图片策划块的有序序列，每块含图片定位/核心信息/文字内容/视觉提示词  (来自 case-A_xhs_69663d9d, extends `提示词`)
+- `图片策划`: 单张信息图的结构化策划块，含图片定位/核心信息/文字内容（主标题+副标题+补充说明）/视觉提示词  (来自 case-A_xhs_69663d9d, extends `提示词`)
+- `信息图`: NanoBanana生成的小红书简笔画素描风格单张信息图（竖版3:4，手写体文字），终态成品直接交付  (来自 case-A_xhs_69663d9d, extends `成品图`)
+- `信息图序列`: N张小红书简笔画风格信息图的有序列表，构成完整信息图系列，终态成品集合直接交付  (来自 case-A_xhs_69663d9d, extends `成品图`)
+- `文本`: 用户提供的待可视化原始文字内容（如某主题的教程、知识点或经验分享），在工序中作为LLM内容分析的输入素材  (来自 case-A_xhs_69663d9d, extends `正文`)
+
+- `参数列表`: 从需求描述中提取出的结构化变量参数清单，列出可替换参数与固定元素，供后续提示词模板组合使用  (来自 case-A_zhihu_20382372, extends `描述`)
+
+- `参数列表`: 从需求描述中提取出的结构化变量参数清单，列出可替换参数与固定元素，供后续提示词模板组合使用  (来自 case-?, extends `描述`)
+
+- `文本`: 文本形式的主题关键词、场景描述或内容要求，作为图像生成的描述性输入依据  (来自 case-B_zhihu_20418774, extends `描述`)
+- `生成图`: 由 GPT-Image-2 文生图步骤产出的终态图像，作为最终交付成品  (来自 case-B_zhihu_20418774, extends `成品图`)
+
+- `平台工具`: 操作入口所在的平台及工具信息，说明使用哪个 Web 平台访问哪个 AI 工具  (来自 case-A_xhs_6a16c1d0, extends `生成参数`)
+- `工具配置`: 工具与模型的综合配置项，含平台模型选型、参考模式、画幅规格、时长及配音等  (来自 case-A_xhs_6a16c1d0, extends `生成参数`)
+
+- `平台账号`: 访问特定平台所需的账号凭证，包含登录信息及可用配额（如即梦平台每日赠送的88-100免费积分）。  (来自 case-C_gzh_e55e66b8, extends `生成参数`)
+- `关键词`: 提示词中具有特定视觉或语义影响力的单个词项，用于控制生图结果的局部属性（如光照时段、渲染风格、色调），可被单独替换以观察画面变化。  (来自 case-C_gzh_e55e66b8, extends `提示词`)
+- `提示词能力`: 通过大量关键词替换实验积累形成的感知性知识——掌握「关键词→画面变化」映射规律，能快速将脑中意图转化为AI可理解的精准提示词。  (来自 case-C_gzh_e55e66b8, extends `知识库`)
+- `参考样本`: 已成功完成的样本实例（如已排版的网页框架、成功的文案或代码结构），整体作为AI处理新内容时的风格/结构/逻辑参照。  (来自 case-C_gzh_e55e66b8, extends `对标内容`)
+- `原始素材`: 待按参考样本风格处理的原始内容（如未排版的新文章），在工序中作为被加工的主体输入，经AI处理后产出风格对齐的结果。  (来自 case-C_gzh_e55e66b8, extends `正文`)
+- `AI处理结果`: AI按填入变量后的提示词模板执行任务后产出的内容结果（如产品经理视角的用户留存策略分析），质量稳定可预期，作为工序的最终交付物。  (来自 case-C_gzh_e55e66b8, extends `正文`)
+- `变量值`: 用于填入提示词模板中各占位符的具体参数值组合（如职业/角色、任务描述、受众特征、输出格式等），驱动模板实例化并执行AI任务。  (来自 case-C_gzh_e55e66b8, extends `生成参数`)
+- `提示词武器库`: 持续积累的个人提示词资产库，以飞书或Chat Memo为载体，存储经过验证的可复用提示词与模板，构成长期护城河。  (来自 case-C_gzh_e55e66b8, extends `知识库`)
+- `提示词模板`: 带[变量名]占位符的可复用提示词骨架，固定结构部分不变，可变参数用占位符标记，使用时填入具体变量值即可直接执行。  (来自 case-C_gzh_e55e66b8, extends `模板`)
+
+- `产品图`: 本次工序中待迁移风格的目标产品图（雅诗兰黛 Re-Nutriv 护肤水官方图），被 AI 分析以提取目标产品的外观特征，并在生图时作为图像参考上传到 Jimeng；其角色是'被模仿/被迁移风格的目标实体'，而非单纯的风格范本，因此以 case-specific 类型独立标注。  (来自 case-A_xhs_6968aa4d, extends `参考图`)
+- `工具配置`: Jimeng 文生图工具的综合生图配置集合，包含：已粘贴的反推提示词、已上传的目标产品参考图、画面比例（2:3）、参考强度（100%）、使用模型（图片 4.0）、输出规格（高清 2K）；属于本次生图任务的完整参数束，超出单一生成参数范畴故作 case-specific 注册。  (来自 case-A_xhs_6968aa4d, extends `生成参数`)
+
+- `生成图`: 由文生图模型依据提示词输出的复刻成品图，是本工序的最终交付物；用于区分来自 AI 生成器（而非摄影或设计工具）的成品图像  (来自 case-B_xhs_693adc7f, extends `成品图`)
+
+- `需求文档`: 对目标广告效果等创作需求的文字说明，作为后续生成步骤的输入依据  (来自 case-A_gzh_9091d1e7, extends `描述`)
+- `提示词块`: 用于后续拼合为完整生成提示词的分段描述文本，如卖点描述块、场景构图块、光线风格块  (来自 case-A_gzh_9091d1e7, extends `提示词`)
+- `广告图提示词`: 由多个描述块合并而成的完整文生图提示词，用于驱动广告图像的AI生成  (来自 case-A_gzh_9091d1e7, extends `提示词`)
+- `广告图`: 以实物照为基础、经AI图生图生成的最终交付广告图像  (来自 case-A_gzh_9091d1e7, extends `成品图`)
+- `实体产品`: 待拍摄的实体商品，作为摄影采集的主体对象，其视觉呈现将通过拍摄步骤转化为数字素材  (来自 case-A_gzh_9091d1e7, extends `参考图`)
+- `产品照片`: 手机拍摄得到的产品实物照片，作为图生图换背景的基础输入图像  (来自 case-A_gzh_9091d1e7, extends `底图`)
+- `图生图提示词`: 用于图生图（换背景）操作的提示词，描述目标场景、光线和风格，保留产品主体  (来自 case-A_gzh_9091d1e7, extends `提示词`)
+
+- `图片描述`: AI对参考图结构化识别后生成的描述文本，含主体、背景、视觉风格、色调、构图、光影等维度  (来自 case-C_zhihu_20169233, extends `描述`)
+- `节气海报`: 以中国传统节气为主题的AI批量生成海报图，含节气文字排版、传统文化元素和装饰艺术风格  (来自 case-C_zhihu_20169233, extends `成品图`)
+
+- `创作需求`: 图片主题、用途与内容要点的文字描述，是图像创作的原始需求输入  (来自 case-?, extends `描述`)
+- `图片定位`: 图像的使用场景、目标平台与版式方向的定位性描述  (来自 case-?, extends `描述`)
+- `风格标签`: 图像视觉风格方向的关键词组合标签（如水墨风/微缩景观/诗意高端）  (来自 case-?, extends `描述`)
+- `新创作目标`: 新主题的替换要素说明，含城市名、地标、诗词文案等待替换的核心内容变量  (来自 case-?, extends `描述`)
+
+- `主题构想`: 用户对本次生成目标的主题意图描述，包含风格期望（水墨禅意）与构图诉求（留白供排版），作为构建提示词的原始输入依据  (来自 case-B_zhihu_20365059, extends `描述`)
+- `封面图`: 科技新闻杂志风格的封面图像，包含标题文字、视觉主体与配色排版，可作为底图被后续指令迭代修改  (来自 case-B_zhihu_20365059, extends `成品图`)
+- `修改需求`: 针对已生成背景图提出的变体修改意见，指定画面主体与色调调整方向，驱动下一轮图像迭代  (来自 case-B_zhihu_20365059, extends `评语`)
+- `背景图`: 水墨风格的PPT横版背景图，以远山云雾或竹林为主体，含大面积留白供文字排版，可被后续指令迭代生成变体  (来自 case-B_zhihu_20365059, extends `成品图`)
+
+- `操作界面`: 工具平台的配置操作界面，带有多个可填写区域（如基础设定、提示词、高级设定等），是智能体配置流程的交互载体  (来自 case-A_gzh_7dfd3085, extends `模板`)
+- `配置项`: 智能体配置中已填写完成的具体设定条目，记录名称、简介、头像、提示词等配置内容  (来自 case-A_gzh_7dfd3085, extends `描述`)
+- `AI智能体`: 已配置并发布到平台的AI对话式智能体实例，包含基础设定与系统提示词，可被用户调用执行多步对话任务  (来自 case-A_gzh_7dfd3085, extends `工作流`)
+- `分析报告`: 对图像或内容按结构化公式所做的系统性分析，拆解风格、主题、构图、色彩、情绪等维度，以定性文字输出  (来自 case-A_gzh_7dfd3085, extends `评语`)
+
+- `创意方向`: 视频创作方向与主题意图的文字刻画，含人物设定、场景风格、叙事核心等创意要素，作为手写提示词步骤的原始输入依据  (来自 case-C_gzh_5bdae160, extends `描述`)
+- `生成图`: AI文生图模型（豆包）直接输出的像素风格图像，含水印或去水印后均以此类型流转，后续经去水印处理并作为图生视频步骤的参考帧输入，是图像→视频转换链路的中间态  (来自 case-C_gzh_5bdae160, extends `底图`)
+- `生成视频`: AI图生视频模型（即梦）直接输出的像素风格动态视频片段，含水印或去水印后均以此类型流转，最终作为剪映剪辑合成步骤的主体素材被组装进成片  (来自 case-C_gzh_5bdae160, extends `视频片段`)
+- `音频素材`: 从爆款短视频平台或个人音乐库中挑选的背景音乐，与像素风赛博朋克视频风格匹配，被混入剪映剪辑合成步骤的音轨  (来自 case-C_gzh_5bdae160, extends `BGM`)
+- `合成视频`: 剪映合成后的完整短视频，含配乐BGM、字幕/对话气泡及0.7倍变速处理，时长约7.2秒，是发布前的完整成品形态  (来自 case-C_gzh_5bdae160, extends `视频成品`)
+- `短视频`: 发布到视频号/抖音/小红书等公域平台的像素风格赛博朋克AI短视频，是工序链路的最终交付物  (来自 case-C_gzh_5bdae160, extends `视频成品`)
+
+- `主题词`: 单个主题词条，用于定向替换或标识特定语义对象的极短文字片段（如节气名称、风格标签等）  (来自 case-B_zhihu_20400839, extends `描述`)
+
+- `视觉风格JSON`: 从参考图中提取的多维度视觉风格结构化描述（JSON格式），涵盖颜色、排版、构图、特效、色调、表现手法、氛围等维度，作为后续图像生成步骤的风格控制依据被消费。  (来自 case-B_gzh_e9677a53, extends `描述`)
+
+- `风格描述`: 豆包从参考图提取的JSON格式视觉风格描述，含颜色/排版/构图/特效四个维度，作为生图提示词的风格模板来源  (来自 case-B_gzh_25318352, extends `描述`)
+
+- `迭代记录`: 记录多轮提示词迭代的完整过程日志——每条记录包含该轮的提示词版本、执行结果（是否达预期）、发现的问题以及下一步的修改方向；是对已产出物逐轮评判+修改意见的历史汇总，兼具过程审查与版本溯源作用。  (来自 case-C_zhihu_20313114, extends `评语`)
+
+- `创意意图`: 用户对想生成图像的初始模糊描述，尚未结构化，常表现为堆叠关键词  (来自 case-C_zhihu_19056944, extends `提示词`)
+- `8要素骨架`: 按「主体+场景+动作/关系+构图+光线+风格/媒介+质感+镜头参数」公式填写的结构化要素清单  (来自 case-C_zhihu_19056944, extends `提示词`)
+- `完整提示词方案`: 包含正向提示词（含参考系、构图控制词）与负向提示词的完整 AI 生图提示词集合，可直接粘贴至生图工具  (来自 case-C_zhihu_19056944, extends `提示词`)
+- `出片规划`: 二段式出片的操作计划：草图阶段（先对构图）→ 细节强化阶段（皮肤/眼睛/材质/背景）  (来自 case-C_zhihu_19056944, extends `脚本`)
+- `一致性配置`: 用于系列图输出的一致性策略配置：参考图上传设置 + 固定 Seed + 单变量迭代规则  (来自 case-C_zhihu_19056944, extends `生成参数`)

examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/wf-patch.py

@@ -50,10 +50,15 @@ value 特殊取值:
     - instruction (列表套列表, 手动 Edit; 透传 directive 用 --resolve-passthrough)
     改字段/删字段/改 source 现在都走本工具, 不要再 Read→Edit 改 workflow.json (会反复重读、烧 token).
 
+自动修引号 (load 时兜底):
+    workflow.json 由模型直 Write, 偶尔把中文引号写成未转义的 ASCII " → JSON 崩.
+    本工具 load 失败时会自动把这类误引号修成「」再 parse; 修成功则继续 patch, 并把
+    修复随本次写回落盘 (--dry-run 不写). 修不回才按 exit 2 报错. 不用再手写 _scratch 修复脚本.
+
 退出码:
     0  全部校验通过并写入 (--dry-run 时为校验通过)
     1  有校验失败 (整批未写) / 路径解析失败
-    2  CLI 参数错误 / 文件不存在 / JSON 损坏
+    2  CLI 参数错误 / 文件不存在 / JSON 损坏 (且自动修引号也救不回)
 """
 from __future__ import annotations
 
@@ -86,6 +91,52 @@ KIND_VOCAB = {'step', 'block', 'nested', 'atom'}
 META_REF = re.compile(r'[（(]?\s*同\s*s[\d]|见\s*s[\d]|←\s*s[\d]|同上')
 
 
+# ===========================================================================
+# 自动修引号: 模型直 Write workflow.json 时常把中文引号写成 ASCII " (未转义) → JSON 崩.
+# 仅在 json.loads 失败时兜底调用 (合法文件零开销). 判别: 串内一个 ASCII " 之后第一个
+# 非空白字符 ∈ {,:}]} 或 EOF → 真·字符串定界符 (保留); 否则是误写的内容引号 → 换直角
+# 引号「」(串内交替 开「/闭」). 逻辑独立内置于本文件 (不 import 任何外部模块);
+# scratch/repair_workflow_quotes.py 是同款独立实现, 二者无依赖关系. 改完必须能 parse 才用.
+# ===========================================================================
+
+_STRUCT_AFTER = set(',:}]')
+
+
+def repair_ascii_quotes(raw: str):
+    """→ (修后文本, 改动的内容引号数). 纯走字符, 不依赖能否 parse."""
+    out, i, n = [], 0, len(raw)
+    in_str = esc = False
+    open_q = True
+    changes = 0
+    while i < n:
+        c = raw[i]
+        if not in_str:
+            out.append(c)
+            if c == '"':
+                in_str, esc, open_q = True, False, True
+            i += 1
+            continue
+        if esc:
+            out.append(c); esc = False; i += 1; continue
+        if c == '\\':
+            out.append(c); esc = True; i += 1; continue
+        if c == '"':
+            j = i + 1
+            while j < n and raw[j] in ' \t\r\n':
+                j += 1
+            nxt = raw[j] if j < n else ''
+            if nxt == '' or nxt in _STRUCT_AFTER:
+                out.append(c); in_str = False          # 真·结束符
+            else:
+                out.append('「' if open_q else '」')     # 误写的内容引号
+                open_q = not open_q
+                changes += 1
+            i += 1
+            continue
+        out.append(c); i += 1
+    return ''.join(out), changes
+
+
 class PathError(Exception):
     """路径无法解析到 workflow.json 里的目标位置."""
 
@@ -574,11 +625,21 @@ def main() -> None:
     if not wf.exists():
         print(f'wf-patch: 文件不存在 {wf}', file=sys.stderr)
         sys.exit(2)
+    raw = wf.read_text(encoding='utf-8')
+    repaired = 0
     try:
-        data = json.loads(wf.read_text(encoding='utf-8'))
+        data = json.loads(raw)
     except json.JSONDecodeError as e:
-        print(f'wf-patch: {wf} 不是合法 JSON: {e}', file=sys.stderr)
-        sys.exit(2)
+        # 兜底: 试着把误写成 ASCII 的中文引号修成「」再 parse (模型直 Write 常见崩因)
+        fixed, repaired = repair_ascii_quotes(raw)
+        try:
+            data = json.loads(fixed)
+        except json.JSONDecodeError:
+            print(f'wf-patch: {wf} 不是合法 JSON (自动修引号也救不回): {e}', file=sys.stderr)
+            sys.exit(2)
+        print(f'[wf-patch] ⚠️ 原文件 JSON 非法 ({e.msg} @ line {e.lineno}); 已自动把 '
+              f'{repaired} 处误写的 ASCII 引号修成「」→ 解析成功, 修复将随本次写回落盘',
+              file=sys.stderr)
 
     patches = load_patches(args)
     unsets = args.unset or []
@@ -654,16 +715,19 @@ def main() -> None:
 
     n_changes = len(plan) + len(del_plan) + len(filled)
     if args.dry_run:
-        print(f'\n--dry-run: 预演 {n_changes} 处改动, 未写入.')
+        extra = f' (+ 自动修复 {repaired} 处引号, dry-run 同样不写)' if repaired else ''
+        print(f'\n--dry-run: 预演 {n_changes} 处改动{extra}, 未写入.')
         sys.exit(0)
 
-    if n_changes == 0:
+    # repaired>0 时即便无字段改动也要落盘 (否则修好的引号没存下来, 文件还是坏的)
+    if n_changes == 0 and not repaired:
         print('\n没有改动 (透传 value 都已填好 / 无可赋值), 未写文件.')
         sys.exit(0)
 
     # 落盘 (安全序列化, 你从不手写 JSON)
     wf.write_text(json.dumps(data, ensure_ascii=False, indent=2) + '\n', encoding='utf-8')
-    print(f'\n已写入 {n_changes} 处到 {wf.name}.')
+    tail = f' (含自动修复 {repaired} 处引号→「」)' if repaired else ''
+    print(f'\n已写入 {n_changes} 处到 {wf.name}{tail}.')
     sys.exit(0)
 
 

examples/process_pipeline/script/search_eval/runs_full/q0000/procedures/B_xhs_6963be77/_source.json

@@ -1,26 +0,0 @@
-{
-  "title": "码住🔥AI如何抄参考图| 附详细教程🔥",
-  "link": "https://www.xiaohongshu.com/explore/6963be77000000000903b23b",
-  "body_text": "[doge][doge]超实用的城市建筑图来啦\n用AI简直太方便[斜眼R][斜眼R]\n本期ji-MengAI辅助完成✅仅供娱乐学习[斜眼R]\n[自拍R]欢迎大家一起解锁更多玩法\n#AI工具[话题]# #ai关键词[话题]# #AIGC[话题]# #AI生成[话题]# #自媒体工具[话题]# #作图神器[话题]# #AI绘画[话题]# #电商设计邪修神器[话题]# #图片王者[话题]# #数字造梦季[话题]#",
-  "images": [
-    "https://ci.xiaohongshu.com/1040g00831r737ihvn2005pjeua00u96o76aanfg?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g00831r73520l720g5pjeua00u96odskpi9g?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g00831r73520l72005pjeua00u96of4luje8?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g00831r73520l72105pjeua00u96oups6a88?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g00831r737ihvn20g5pjeua00u96omdjtmp0?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g00831r73520l721g5pjeua00u96oposh8d8?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g00831r73520l72505pjeua00u96op4191io?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g00831r73520l72405pjeua00u96o02293e0?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g00831r73520l724g5pjeua00u96o6le2mq8?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g00831r73520l723g5pjeua00u96ou4bseq8?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g00831r73520l72605pjeua00u96okj9vcho?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g00831r73520l725g5pjeua00u96onuv9278?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g00831r73520l726g5pjeua00u96oc3kudf0?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g00831r737ihv72005pjeua00u96o1q4rk1g?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g00831r73520l722g5pjeua00u96odp9rgqg?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g00831r73520l72205pjeua00u96o85le8l8?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g00831r73520l72305pjeua00u96o79gfl98?imageView2/2/w/1080/format/webp"
-  ],
-  "publish_timestamp": "2026-01-11 23:15:03",
-  "channel_account_name": "xhs"
-}

examples/process_pipeline/script/search_eval/runs_full/q0002/procedures/A_xhs_6885b23c/_source.json

@@ -1,15 +0,0 @@
-{
-  "title": "如何让AI去抄你的参考图？（附教程）",
-  "link": "https://www.xiaohongshu.com/explore/6885b23c0000000023007ff7",
-  "body_text": "1️⃣参考图给AI 反推提示词\n2️⃣然后再修改提示词，根据自己的需求\n3️⃣用AI2.0 指令词和参考图去生图，积分每天有送的\n（可以反复抽卡，得到自己想要的为止）多生成几次！效果会越来越好\n#设计教程[话题]#  #AI绘画[话题]#   #创客贴AI[话题]# #海报设计[话题]#   #AI生图[话题]#   #AI抄参考图[话题]#  #电商海报[话题]# #水果海报[话题]# #水果海报设计[话题]# #创客贴[话题]#",
-  "images": [
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831ke86ss0io3g49hepiq0g8opv425fj8?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831ke86ss0io4049hepiq0g8opblaahjg?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831ke86ss0io4g49hepiq0g8opt7mljuo?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831ke86ss0io5049hepiq0g8opjc30ndg?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831ke86ss0io5g49hepiq0g8opci4qdr0?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831ke86ss0io6049hepiq0g8opqc5qvpg?imageView2/2/w/1080/format/webp"
-  ],
-  "publish_timestamp": "2025-07-27 12:59:40",
-  "channel_account_name": "xhs"
-}

examples/process_pipeline/script/search_eval/runs_full/q0002/procedures/A_xhs_69f31a3d/_source.json

@@ -1,15 +0,0 @@
-{
-  "title": "如何让AI复刻电商海报？",
-  "link": "https://www.xiaohongshu.com/explore/69f31a3d00000000350223cf",
-  "body_text": "如何让AI复刻电商海报？（附详细教程）\n不会写提示词的宝子们看过来，教你几分钟做出满意的海报！\n1、寻找参考图。优先去找和自身产品差不多透视的参考图。\n2、反推提示词。向任意大语言模型发出指令：请分析这张图片，并生成一个能够指导AI作图工具重新创作类似作品的文生图提示词。提示词中需从主体内容、场景设定、风格参考、色调色彩、构图视角、细节补充等这些角度用文字描述图片。\n3、修改提示词。按照自身需求修改大语言模型生成的提示词。\n4、生成图片。把修改后的提示词喂给图片生成软件，挑选出满意的图片。",
-  "images": [
-    "https://ci.xiaohongshu.com/1040g2sg31vj34dr6jmkg5qd2v8v76qarjlpdb9g?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g2sg31vj34dr6jmjg5qd2v8v76qarsmco5c0?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g2sg31vj34dr6jmj05qd2v8v76qar9fqb5k0?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g2sg31vj34dr6jmfg5qd2v8v76qar16uegh8?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g2sg31vj34dr6jmh05qd2v8v76qarh2pv80o?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/1040g2sg31vj34dr6jmk05qd2v8v76qarnq00s58?imageView2/2/w/1080/format/webp"
-  ],
-  "publish_timestamp": "2026-04-30 17:00:45",
-  "channel_account_name": "xhs"
-}

examples/process_pipeline/script/search_eval/runs_full/q0002/procedures/B_xhs_69832685/_source.json

@@ -1,13 +0,0 @@
-{
-  "title": "1:1快速用豆包反推提示词，太香了！",
-  "link": "https://www.xiaohongshu.com/explore/698326850000000009038add",
-  "body_text": "提示词放评论区了 需要自取吧！\n#AI工具[话题]# #ai关键词[话题]# #AI人工智能[话题]# #人工智能[话题]# #ai[话题]# #作图神器[话题]# #AI画图[话题]# #搜索关键词[话题]#",
-  "images": [
-    "https://ci.xiaohongshu.com/spectrum/1040g0k031s5oko2fla705q1dfpujif7he3hn8d8?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/spectrum/1040g34o31s5oktlk5c105q1dfpujif7h043vg58?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/spectrum/1040g34o31s5oktlk5c1g5q1dfpujif7hka5offg?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/spectrum/1040g34o31s5ol5ta58105q1dfpujif7h8r1ek10?imageView2/2/w/1080/format/webp"
-  ],
-  "publish_timestamp": "2026-02-04 18:59:17",
-  "channel_account_name": "xhs"
-}

examples/process_pipeline/script/search_eval/runs_full/q0002/procedures/C_xhs_68400e68/_source.json

@@ -1,17 +0,0 @@
-{
-  "title": "【豆包一键推】图片反推提示词技巧✨",
-  "link": "https://www.xiaohongshu.com/explore/68400e680000000023010d19",
-  "body_text": "看到绝美图片或作品，只会”哇塞”却不会描述？别求人啦！\n豆包隐藏技能简直开挂——3秒反推【精准提示词】1:1还原神图！\n\t\n👇👇👇保姆级教程：\n\t\n🌟1：豆包深度解析\n直接把图 “喂给” 给豆包对话框，3 秒生成【精准提示词】，风格、主体、配文、材质、光线，直接1:1还原神图！\n\t\n🌟2：即梦精准复刻\n复制关键词直接生成，连色板都原汁原味～成图相似度90%！\n\t\n豆包AI #即梦教程 #提示词黑科技 #AI绘画 #设计素材自由 #即梦AI #AI绘画教程 #人物一致性 #垫图大法 #抽卡省钱攻略 #即梦 #即梦AI #图片风格统一 #aigc #ai绘画 #视频脚本 #咒语 #风格 #即梦教程 #ai工具 #即梦AI #AI绘画 #风格码 #抽卡救星 #即梦AI #AI绘画 #效率神器 #AI绘本创作 #AI工具 #即梦AI教程",
-  "images": [
-    "https://ci.xiaohongshu.com/spectrum/1040g34o31ia7vbi26m405nrg3u20bj9198ahilg?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/spectrum/1040g34o31ia7vbi26m4g5nrg3u20bj91ob0u1l8?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/spectrum/1040g34o31ia7vbi26m505nrg3u20bj91hmcka68?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/spectrum/1040g34o31ia7vbi26m5g5nrg3u20bj910ce3r00?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/spectrum/1040g34o31ia7vbi26m605nrg3u20bj91qa6q2o8?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/spectrum/1040g34o31ia7vbi26m6g5nrg3u20bj91tnll7t8?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/spectrum/1040g34o31ia7vbi26m705nrg3u20bj91kkqs1m0?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/spectrum/1040g34o31ia7vbi26m7g5nrg3u20bj914dmebj0?imageView2/2/w/1080/format/webp"
-  ],
-  "publish_timestamp": "2025-06-04 17:14:16",
-  "channel_account_name": "xhs"
-}

examples/process_pipeline/script/search_eval/runs_full/q0003/procedures/B_xhs_6a146c58/_source.json

@@ -1,23 +0,0 @@
-{
-  "title": "别再问AI图里是啥prompt，自己搭个反推器",
-  "link": "https://www.xiaohongshu.com/explore/6a146c58000000003501f0d8",
-  "body_text": "之前发了10组反推提示词框架，很多兄弟说：框架有了，但怎么让它变成一个上传图就能用的工具？\n我花了3小时搭了一个「AI提示词反推器」，上传图片3秒出结果。\n🔧 6步搞定：\n1⃣ 打开主流G开头工具，点\"探索GPT\"\n2⃣ 点右上角\"创建\"\n3⃣ 填名称+描述+指令（指令是核心，决定了它反推还是描述）\n4⃣ 上传核心规则文件，选最新思考模型\n5⃣ 预览测试，确认输出格式\n6⃣ 设置共享，保存\n💡 关键在第3步的指令： 不要让它\"描述图片\"，要让它\"拆生成逻辑\"。 构图怎么复现？镜头感怎么写？光线从哪来？哪些东西必须排除？\n\t\n搭建过程我截图了，跟着做就行。\n\t\n#它工具[话题]# #提示词技巧[话题]# #设计师效率[话题]# #它绘图[话题]# #反推器[话题]# #设计师日常[话题]# #内容创作[话题]# #效率翻倍[话题]# #ai[话题]# #ai智能体[话题]#",
-  "images": [
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320jjueim5cdg5paffnv206a2bfa56eo?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320jjueim5c705paffnv206a2o861iqo?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320jjueim5c7g5paffnv206a2aj0alug?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320jjueim5c805paffnv206a2cparidg?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320jjueim5c8g5paffnv206a2leme488?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320jjueim5c905paffnv206a2rq14emo?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320jjueim5c9g5paffnv206a2dm39ig0?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320jjueim5ca05paffnv206a29h6kebg?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320jjueim5cag5paffnv206a2nc1h398?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320jjueim5cb05paffnv206a2r7gasro?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320jjueim5cbg5paffnv206a2t1jdbn0?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320jjueim5cc05paffnv206a2pspe6v0?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320jjueim5ccg5paffnv206a2nbqor20?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320jjueim5cd05paffnv206a2ss39gpo?imageView2/2/w/1080/format/webp"
-  ],
-  "publish_timestamp": "2026-05-25 23:35:52",
-  "channel_account_name": "xhs"
-}

examples/process_pipeline/script/search_eval/runs_full/q0004/procedures/C_xhs_691847c8/_source.json

@@ -1,17 +0,0 @@
-{
-  "title": "如何让AI抄产品参考图？（附完整指令词）",
-  "link": "https://www.xiaohongshu.com/explore/691847c800000000040281ff",
-  "body_text": "电商人马不停蹄的继续肝双十二！真的很需要！邪修玩法复刻这种光影融合的产品图！还可以进行产品图精修，实现摄影棚效果！效果直接拉满[自拍R] 指令词都在p8啦很多样式的都可以用这个思路[暗中观察R]\n💡在立【ji】想要造｛meng｝的完整指令词：\n超写实商业产品摄影，高端冷奢视觉氛围，以大量透明立方体冰块为背景填充，LANCÔME肌底液瓶身斜向嵌入冰堆中层（与原场景产品角度一致）；瓶身主体为黑渐变透明玻璃+亮面银灰金属瓶盖，表面附着细密水珠，文字为白色“LANCÔME”“ADVANCED GÉNIFIQUE”；调整色调为冷黑+冰灰+银白的低饱和冷奢色系，冰块保留晶莹通透质感但弱化蓝调，光线用柔和侧逆光突出玻璃的通透光泽与金属的冷冽反光，浅景深特写视角，主体位于画面中偏右区域，背景为纯净浅灰渐变，呈现冰感与奢华质感融合的护肤产品视觉效果。\n💡 仅供学习交流分享，参考图来源🕸️\n#电商设计[话题]# #产品摄影[话题]# #产品拍摄[话题]# #产品精修[话题]#  #电商设计邪修神器[话题]# #设计教程[话题]# #小红书创作灵感设计大赛[话题]# #设计灵感[话题]# #产品设计[话题]# @小红书设计周",
-  "images": [
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831otct283ha7049hepiq0g8op3oqv1kg?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831otct283ha7g49hepiq0g8op62s7s3g?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831otct283ha8049hepiq0g8opn7dkqgo?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831otct283ha8g49hepiq0g8op1pt6j0o?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831otct283ha9049hepiq0g8opj09lq30?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831otct283ha9g49hepiq0g8opbgs6st8?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831otct283haa049hepiq0g8opiqk88gg?imageView2/2/w/1080/format/webp",
-    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831otct283haag49hepiq0g8op8fv71q0?imageView2/2/w/1080/format/webp"
-  ],
-  "publish_timestamp": "2025-11-15 17:28:40",
-  "channel_account_name": "xhs"
-}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/.session_id

@@ -0,0 +1 @@
+0f073b67-c727-4574-a6dc-232d83a8cdb4

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/_meta.json

@@ -0,0 +1,8 @@
+{
+  "case_id": "xhs_69663d9d000000000a02adc4",
+  "from_q": "q0018",
+  "form": "A",
+  "score": 9.4,
+  "score_norm": 0.94,
+  "started_at": "2026-06-03T13:49:49"
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/_scratch/anchors.json

@@ -0,0 +1,8 @@
+[
+  {"path": "p1.s1.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p1.s1.outputs[0].anchor", "value": "→ s2.图片策划序列"},
+  {"path": "p1.s2.inputs[0].anchor",  "value": "← s1.图片策划序列"},
+  {"path": "p1.s2.outputs[0].anchor", "value": "→ 返回 信息图序列"},
+  {"path": "p1.s2.1.inputs[0].anchor",  "value": "← 图片策划序列[i]"},
+  {"path": "p1.s2.1.outputs[0].anchor", "value": "→ 信息图序列.追加"}
+]

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/_scratch/patch_2a.json

@@ -0,0 +1,16 @@
+[
+  {"path": "p1.s1.effect", "value": "主体生成"},
+  {"path": "p1.s1.action", "value": "生成/数组生成"},
+  {"path": "p1.s2.1.effect", "value": "主体生成"},
+  {"path": "p1.s2.1.action", "value": "生成/元素生成"},
+  {"path": "p1.type_registry.文本.extends", "value": "正文"},
+  {"path": "p1.type_registry.文本.desc", "value": "用户提供的待可视化原始文字内容（如某主题的教程、知识点或经验分享），在工序中作为LLM内容分析的输入素材"},
+  {"path": "p1.type_registry.图片策划序列.extends", "value": "提示词"},
+  {"path": "p1.type_registry.图片策划序列.desc", "value": "LLM输出的N个结构化图片策划块的有序序列，每块含图片定位/核心信息/文字内容/视觉提示词"},
+  {"path": "p1.type_registry.图片策划.extends", "value": "提示词"},
+  {"path": "p1.type_registry.图片策划.desc", "value": "单张信息图的结构化策划块，含图片定位/核心信息/文字内容（主标题+副标题+补充说明）/视觉提示词"},
+  {"path": "p1.type_registry.信息图.extends", "value": "成品图"},
+  {"path": "p1.type_registry.信息图.desc", "value": "NanoBanana生成的小红书简笔画素描风格单张信息图（竖版3:4，手写体文字，黑白灰+单色点缀，右下角「御风」水印），作为终态成品直接交付"},
+  {"path": "p1.type_registry.信息图序列.extends", "value": "成品图"},
+  {"path": "p1.type_registry.信息图序列.desc", "value": "N张小红书简笔画风格信息图的有序列表，构成完整信息图系列，作为终态成品集合直接交付"}
+]

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/_scratch/patch_2b.json

@@ -0,0 +1,14 @@
+[
+  {"path": "p1.s1.inputs[0].substance", "value": "/理念/知识/思想/知识类型"},
+  {"path": "p1.s1.inputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/流程递进"},
+  {"path": "p1.s1.outputs[0].substance", "value": "__null__"},
+  {"path": "p1.s1.outputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/逐条罗列/通用列举"},
+  {"path": "p1.s2.inputs[0].substance", "value": "__null__"},
+  {"path": "p1.s2.inputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/逐条罗列/通用列举"},
+  {"path": "p1.s2.outputs[0].substance", "value": "/表象/视觉/实体/物品/影像物品/艺术创作"},
+  {"path": "p1.s2.outputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/逐条罗列/通用列举"},
+  {"path": "p1.s2.1.inputs[0].substance", "value": "__null__"},
+  {"path": "p1.s2.1.inputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+  {"path": "p1.s2.1.outputs[0].substance", "value": "/表象/视觉/实体/物品/影像物品/艺术创作"},
+  {"path": "p1.s2.1.outputs[0].form", "value": "/呈现/视觉/视觉制作/构图编排/版面设计/标注图示 + /呈现/视觉/视觉风格/视觉艺术风格/绘画风格/素描 + /呈现/视觉/视觉制作/构图编排/版面设计/版面结构"}
+]

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/_scratch/task_2a.json

@@ -0,0 +1,136 @@
+{
+  "case_id": "A_xhs_69663d9d",
+  "task": "Phase 2A (effect/action/type normalization)",
+  "image_url_list": [],
+  "steps": [
+    {
+      "path": "p1.s1",
+      "name": "",
+      "action": "生成/关系生成",
+      "effect": "工艺规约",
+      "kind": "step",
+      "type_registry": {
+        "图片策划序列": {
+          "extends": "提示词",
+          "desc": "LLM输出的N个结构化图片策划块的有序序列，每块含图片定位/核心信息/文字内容/视觉提示词"
+        },
+        "图片策划": {
+          "extends": "提示词",
+          "desc": "单张信息图的结构化策划块，含图片定位/核心信息/文字内容（主标题+副标题+补充说明）/视觉提示词"
+        },
+        "信息图": {
+          "extends": "参考图",
+          "desc": "NanoBanana生成的小红书简笔画素描风格单张信息图（竖版3:4，手写体文字）"
+        },
+        "信息图序列": {
+          "extends": "参考图",
+          "desc": "N张小红书简笔画风格信息图的有序列表，构成完整信息图系列"
+        }
+      },
+      "inputs": [
+        {
+          "path": "p1.s1.inputs[0]",
+          "name": "待可视化内容",
+          "value": "<用户提供的待可视化文本内容，如某主题的教程、知识点或经验分享>",
+          "type": "文本",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p1.s1.outputs[0]",
+          "name": "图片策划序列",
+          "value": "<LLM输出的N个结构化图片策划块序列（N由内容复杂度决定），每块含：第X张/共N张 + 图片定位（封面图/内容图/结尾图）+ 核心信息 + 文字内容（主标题/副标题/补充说明）+ 视觉提示词>",
+          "type": "图片策划序列",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p1.s2",
+      "name": "",
+      "action": "",
+      "effect": "",
+      "kind": "block",
+      "type_registry": {
+        "图片策划序列": {
+          "extends": "提示词",
+          "desc": "LLM输出的N个结构化图片策划块的有序序列，每块含图片定位/核心信息/文字内容/视觉提示词"
+        },
+        "图片策划": {
+          "extends": "提示词",
+          "desc": "单张信息图的结构化策划块，含图片定位/核心信息/文字内容（主标题+副标题+补充说明）/视觉提示词"
+        },
+        "信息图": {
+          "extends": "参考图",
+          "desc": "NanoBanana生成的小红书简笔画素描风格单张信息图（竖版3:4，手写体文字）"
+        },
+        "信息图序列": {
+          "extends": "参考图",
+          "desc": "N张小红书简笔画风格信息图的有序列表，构成完整信息图系列"
+        }
+      },
+      "inputs": [
+        {
+          "path": "p1.s2.inputs[0]",
+          "name": "图片策划序列",
+          "value": "<N个结构化图片策划块序列>",
+          "type": "图片策划序列",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p1.s2.outputs[0]",
+          "name": "信息图序列",
+          "value": "<N张生成的小红书风格简笔画信息图有序列表>",
+          "type": "信息图序列",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p1.s2.1",
+      "name": "",
+      "action": "生成/元素生成",
+      "effect": "主体生成",
+      "kind": "nested",
+      "type_registry": {
+        "图片策划序列": {
+          "extends": "提示词",
+          "desc": "LLM输出的N个结构化图片策划块的有序序列，每块含图片定位/核心信息/文字内容/视觉提示词"
+        },
+        "图片策划": {
+          "extends": "提示词",
+          "desc": "单张信息图的结构化策划块，含图片定位/核心信息/文字内容（主标题+副标题+补充说明）/视觉提示词"
+        },
+        "信息图": {
+          "extends": "参考图",
+          "desc": "NanoBanana生成的小红书简笔画素描风格单张信息图（竖版3:4，手写体文字）"
+        },
+        "信息图序列": {
+          "extends": "参考图",
+          "desc": "N张小红书简笔画风格信息图的有序列表，构成完整信息图系列"
+        }
+      },
+      "inputs": [
+        {
+          "path": "p1.s2.1.inputs[0]",
+          "name": "图片策划",
+          "value": "<单张图的结构化策划块，含图片定位/核心信息/文字内容/视觉提示词>",
+          "type": "图片策划",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p1.s2.1.outputs[0]",
+          "name": "信息图",
+          "value": "<NanoBanana生成的小红书简笔画素描风格信息图，竖版3:4比例，手写体文字，黑白灰+单色点缀，简笔画元素，右下角「御风」水印>",
+          "type": "信息图",
+          "related_images": []
+        }
+      ]
+    }
+  ]
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/_scratch/task_2b.json

@@ -0,0 +1,49 @@
+{
+  "case_id": "A_xhs_69663d9d",
+  "task": "Phase 2B (substance/form matching)",
+  "image_url_list": [],
+  "items_to_match": [
+    {
+      "path": "p1.s1.inputs[0]",
+      "name": "待可视化内容",
+      "value": "<用户提供的待可视化文本内容，如某主题的教程、知识点或经验分享>",
+      "type": "文本",
+      "related_images": []
+    },
+    {
+      "path": "p1.s1.outputs[0]",
+      "name": "图片策划序列",
+      "value": "<LLM输出的N个结构化图片策划块序列（N由内容复杂度决定），每块含：第X张/共N张 + 图片定位（封面图/内容图/结尾图）+ 核心信息 + 文字内容（主标题/副标题/补充说明）+ 视觉提示词>",
+      "type": "图片策划序列",
+      "related_images": []
+    },
+    {
+      "path": "p1.s2.inputs[0]",
+      "name": "图片策划序列",
+      "value": "<N个结构化图片策划块序列>",
+      "type": "图片策划序列",
+      "related_images": []
+    },
+    {
+      "path": "p1.s2.outputs[0]",
+      "name": "信息图序列",
+      "value": "<N张生成的小红书风格简笔画信息图有序列表>",
+      "type": "信息图序列",
+      "related_images": []
+    },
+    {
+      "path": "p1.s2.1.inputs[0]",
+      "name": "图片策划",
+      "value": "<单张图的结构化策划块，含图片定位/核心信息/文字内容/视觉提示词>",
+      "type": "图片策划",
+      "related_images": []
+    },
+    {
+      "path": "p1.s2.1.outputs[0]",
+      "name": "信息图",
+      "value": "<NanoBanana生成的小红书简笔画素描风格信息图，竖版3:4比例，手写体文字，黑白灰+单色点缀，简笔画元素，右下角「御风」水印>",
+      "type": "信息图",
+      "related_images": []
+    }
+  ]
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/_trace.md

@@ -0,0 +1,502 @@
+
+
+---
+
+## ▶ Fresh @ 2026-06-03 13:49:51
+
+- case: `A_xhs_69663d9d`
+- source: `C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\runs_full\q0018\procedures\A_xhs_69663d9d\_source.json`
+- spec: `spec`
+- images: `5`
+- model: `claude-sonnet-4-6`
+- max_turns: `300`
+
+_session=0f073b67 · model='claude-sonnet-4-6' · apiKeySource='none'_
+
+
+### Turn 1 · 13:49:59
+
+
+### Turn 2 · 13:50:00
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/README.md`
+
+
+### Turn 3 · 13:50:01
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools.md`
+
+
+### Turn 4 · 13:50:05
+
+
+### Turn 5 · 13:50:05
+> Now let me read the source JSON and the Phase 1 spec files simultaneously:
+
+
+
+### Turn 6 · 13:50:07
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/_source.json`
+
+
+### Turn 7 · 13:50:08
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/syntax.md`
+
+
+### Turn 8 · 13:50:08
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/fields.md`
+
+
+### Turn 9 · 13:50:11
+
+
+### Turn 10 · 13:50:11
+> Now let me read the remaining Phase 1 spec files simultaneously:
+
+
+
+### Turn 11 · 13:50:13
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/control-flow.md`
+
+
+### Turn 12 · 13:50:14
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/phase1-skeleton.md`
+
+
+### Turn 13 · 13:50:14
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/format/md-structure.md`
+
+
+### Turn 14 · 13:50:18
+
+
+### Turn 15 · 13:50:18
+> Now let me read the template and phase2-normalize spec:
+
+
+
+### Turn 16 · 13:50:20
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/templates/workflow.template.json`
+
+
+### Turn 17 · 13:50:20
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/phase2-normalize.md`
+
+
+### Turn 18 · 13:50:24
+
+
+### Turn 19 · 13:50:24
+> Now let me read the phase3 spec and format files:
+
+
+
+### Turn 20 · 13:50:26
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/phase3-finalize.md`
+
+
+### Turn 21 · 13:50:27
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/format/procedure-table.md`
+
+
+### Turn 22 · 13:50:27
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/format/case-data.schema.json`
+
+
+### Turn 23 · 13:59:06
+
+
+### Turn 24 · 13:59:06
+> Now I have all the spec files loaded. Let me proceed to Phase 1.1 — writing the understanding.md:
+
+
+
+### Turn 25 · 13:59:38
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/understanding.md` (1,835 chars)
+
+
+### Turn 26 · 13:59:45
+
+
+### Turn 27 · 13:59:45
+> Now Phase 1.2 — Write the workflow.json skeleton:
+
+
+
+### Turn 28 · 14:00:31
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/workflow.json` (5,527 chars)
+
+
+### Turn 29 · 14:00:51
+
+
+### Turn 30 · 14:00:51
+> Phase 1.3 — add anchors. First write the patch list, then apply:
+
+
+
+### Turn 31 · 14:00:53
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/_scratch/anchors.json` (378 chars)
+
+
+### Turn 32 · 14:00:57
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/wf-patch.py --workflow "../runs_full/q0018/procedures/A_xhs_69663d9d/work...`
+
+
+### Turn 33 · 14:01:14
+
+
+### Turn 34 · 14:01:14
+> Phase 1.3 complete — all 6 anchors set. Now Phase 2. First, check if prepare-subtask.py exists and run it:
+
+
+
+### Turn 35 · 14:01:14
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/prepare-subtask.py --workflow "../runs_full/q0018/procedures/A_xhs_69663d...`
+
+
+### Turn 36 · 14:01:18
+
+
+### Turn 37 · 14:01:19
+> Good. Now dispatch Phase 2A and 2B sub-agents in parallel:
+
+
+
+### Turn 38 · 14:01:37
+- `Agent` — `Phase 2A: effect/action/type normalization for A_xhs_69663d9d` [subagent=phase-2a-normalizer]
+
+
+### Turn 39 · 14:01:41
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/_scratch/task_2a.json`
+
+
+### Turn 40 · 14:01:41
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/taxonomy/effect.json`
+
+
+### Turn 41 · 14:01:42
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/taxonomy/action.json`
+
+
+### Turn 42 · 14:01:42
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/taxonomy/type.json`
+
+
+### Turn 43 · 14:01:49
+- `Agent` — `Phase 2B: substance/form taxonomy matching for A_xhs_69663d9d` [subagent=phase-2b-matcher]
+
+
+### Turn 44 · 14:01:52
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/_scratch/task_2b.json`
+
+
+### Turn 45 · 14:01:57
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "教程 知识点 经验分享 文本内容"`
+
+
+### Turn 46 · 14:01:58
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "图片策划序列 结构化策划 视觉提示词 图文策划"`
+
+
+### Turn 47 · 14:01:59
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "信息图 简笔画 infographic 可视化图片"`
+
+
+### Turn 48 · 14:02:02
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "策划 创意策划 内容策划 脚本"`
+
+
+### Turn 49 · 14:02:04
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "教程 知识 文本 经验"`
+
+
+### Turn 50 · 14:02:05
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "信息图 简笔画 素描 infographic 竖版"`
+
+
+### Turn 51 · 14:02:08
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "图片策划 结构化文档 策划方案 提示词"`
+
+
+### Turn 52 · 14:02:10
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "图文内容 小红书 可视化内容"`
+
+
+### Turn 53 · 14:02:14
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "知识 教程 经验"`
+
+
+### Turn 54 · 14:02:15
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "结构化文本 列表 序列"`
+
+
+### Turn 55 · 14:02:19
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "创意方案 内容方案 图像生成指令"`
+
+
+### Turn 56 · 14:02:21
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "信息图 图文 可视化 竖版图片"`
+
+
+### Turn 57 · 14:02:24
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "指令 方案 脚本 文案策划"`
+
+
+### Turn 58 · 14:02:25
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/_scratch/patch_2a.json` (1,029 chars)
+
+
+### Turn 59 · 14:02:25
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "视觉设计 图像 生成图片"`
+
+
+### Turn 60 · 14:02:29
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "文案 脚本 内容规划"`
+
+
+### Turn 61 · 14:02:30
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "手绘 简笔画 素描风格 图片"`
+
+
+### Turn 62 · 14:02:33
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "内容创作 视觉内容 图文创作 信息可视化"`
+
+
+### Turn 63 · 14:02:34
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "素描 简笔画 绘画风格 信息图"`
+
+
+### Turn 64 · 14:02:39
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "/理念/知识/思想/知识类型"`
+
+
+### Turn 65 · 14:02:40
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "/架构/逻辑/信息编排逻辑/条目列举/分段结构/流程递进"`
+
+
+### Turn 66 · 14:02:44
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "/理念/知识/思想/知识类型" 2>&1`
+
+
+### Turn 67 · 14:02:47
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/taxonomy-lookup.py` (offset=0, limit=60)
+
+
+### Turn 68 · 14:02:50
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/taxonomy-lookup.py` (offset=60, limit=100)
+
+
+### Turn 69 · 14:02:52
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/taxonomy-lookup.py` (offset=200, limit=80)
+
+
+### Turn 70 · 14:02:56
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/taxonomy-lookup.py` (offset=280, limit=80)
+
+
+### Turn 71 · 14:03:00
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "理念/知识/思想/知识类型" 2>&1`
+
+
+### Turn 72 · 14:03:02
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "架构/逻辑/信息编排逻辑/条目列举/分段结构/流程递进" 2>&1`
+
+
+### Turn 73 · 14:03:07
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "理念/方法/生活技巧/经验心得" 2>&1`
+
+
+### Turn 74 · 14:03:08
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "架构/逻辑/信息编排逻辑/条目列举/逐条罗列/通用列举" 2>&1`
+
+
+### Turn 75 · 14:03:10
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 呈现 --validate "呈现/视觉/视觉制作/构图编排/版面设计/标注图示" 2>&1 ;...`
+
+
+### Turn 76 · 14:03:15
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "呈现/视觉/视觉风格/视觉艺术风格/绘画风格/素描" 2>&1`
+
+
+### Turn 77 · 14:03:16
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "呈现/视觉/视觉制作/构图编排/版面设计/版面结构" 2>&1`
+
+
+### Turn 78 · 14:03:25
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --narrow "图片策划 视觉策划 内容策划 创意指令" 2>&1`
+
+
+### Turn 79 · 14:03:26
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "知识 图文 可视化 信息" 2>&1`
+
+
+### Turn 80 · 14:03:29
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --subtree "理念/知识/思想" 2>&1 | head -60`
+
+
+### Turn 81 · 14:03:36
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "图片 生成 插画 绘图" 2>&1`
+
+
+### Turn 82 · 14:03:38
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "架构/逻辑/信息编排逻辑/条目列举/分段结构/流程递进" 2>&1`
+
+
+### Turn 83 · 14:03:50
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架" 2>&1`
+
+
+### Turn 84 · 14:03:51
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "表象/视觉/实体/物品/影像物品/艺术创作" 2>&1`
+
+
+### Turn 85 · 14:04:03
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --subtree "理念/方法" 2>&1 | head -40`
+
+
+### Turn 86 · 14:04:16
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/_scratch/patch_2b.json` (941 chars)
+
+
+### Turn 87 · 14:04:52
+
+
+### Turn 88 · 14:04:52
+> Both sub-agents completed. Now apply the patches — first 2A, then 2B:
+
+
+
+### Turn 89 · 14:04:53
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/wf-patch.py --workflow "../runs_full/q0018/procedures/A_xhs_69663d9d/work...`
+
+
+### Turn 90 · 14:05:11
+
+
+### Turn 91 · 14:05:11
+> The action `生成/数组生成` is invalid. The tool suggests the leaf is `数组生成`. Let me apply the rest of the 2A patch without that one entry, then fix s1.action separately:
+
+
+
+### Turn 92 · 14:05:17
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/wf-patch.py --workflow "../runs_full/q0018/procedures/A_xhs_69663d9d/work...`
+
+
+### Turn 93 · 14:05:21
+
+
+### Turn 94 · 14:05:22
+> Now fix s1.action using the leaf name hint:
+
+
+
+### Turn 95 · 14:05:23
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/wf-patch.py --workflow "../runs_full/q0018/procedures/A_xhs_69663d9d/work...`
+
+
+### Turn 96 · 14:05:26
+
+
+### Turn 97 · 14:05:27
+> Good — s1.action normalized to `生成/关系生成/数组生成`. Now apply Phase 2B:
+
+
+
+### Turn 98 · 14:05:28
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/wf-patch.py --workflow "../runs_full/q0018/procedures/A_xhs_69663d9d/work...`
+
+
+### Turn 99 · 14:05:55
+
+
+### Turn 100 · 14:05:55
+> Phase 2 complete. Now Phase 3 — start with lint:
+
+
+
+### Turn 101 · 14:05:56
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/lint-case.py --workflow "../runs_full/q0018/procedures/A_xhs_69663d9d/wor...`
+
+
+### Turn 102 · 14:06:51
+
+
+### Turn 103 · 14:06:51
+> Lint clean. Now Phase 3.2 — refine intent fields with proper `{kind:value}` tokens:
+
+
+
+### Turn 104 · 14:06:53
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/wf-patch.py --workflow "../runs_full/q0018/procedures/A_xhs_69663d9d/work...`
+
+
+### Turn 105 · 14:07:01
+
+
+### Turn 106 · 14:07:01
+> Now render the HTML with render-case.py:
+
+
+
+### Turn 107 · 14:07:04
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/render-case.py \
+  --workflow "../runs_full/q0018/procedures/A_xhs_69663d...`
+
+
+### Turn 108 · 14:07:24
+
+
+### Turn 109 · 14:07:24
+> HTML rendered (72,775 chars). Now Phase 3.3 — write the .md file:
+
+
+
+### Turn 110 · 14:08:39
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/case-A_xhs_69663d9d-xhs-simpledraw-infographic.md` (4,980 chars)
+
+
+### Turn 111 · 14:08:55
+
+
+### Turn 112 · 14:09:07
+> All phases complete. Here's the final summary:
+>
+> ---
+>
+> ## 工序梗概
+>
+> **小红书风格信息图系列生成**（1 个工序，3 个节点）：
+> - **s1** (`<llm>` + `生成/关系生成/数组生成`)：将用户文本内容 + 系统提示词模板输入 LLM，输出 N 个结构化图片策划块（含封面/内容/结尾分类、核心信息、文字内容、视觉提示词）
+> - **s2**（`block/遍历`）：遍历图片策划序列
+> - **s2.1**（`nano_banana` + `生成/元素生成`）：每个策划块的视觉提示词送 NanoBanana，生成一张简笔画素描风格信息图，追加进信息图序列
+>
+> ---
+>
+> ## 输出文件路径
+>
+> | 文件 | 路径 |
+> |------|------|
+> | 心智模型 | `.../A_xhs_69663d9d/understanding.md` |
+> | 工序数据 | `.../A_xhs_69663d9d/workflow.json` |
+> | HTML 可视化 | `.../A_xhs_69663d9d/case-A_xhs_69663d9d-xhs-simpledraw-infographic.html` |
+> | DSL 文本版 | `.../A_xhs_69663d9d/case-A_xhs_69663d9d-xhs-simpledraw-infographic.md` |
+>
+> ---
+>
+> ## 对 DSL 的关键发现
+>
+> 1. **`信息图` extends `成品图` 而非 `参考图`**：Phase 2A 子 Agent 发现原始骨架误用 `参考图` 作为 extends。NanoBanana 生成的信息图是直接交付的终态成品，不是供参考的素材，纠正后更准确表达了类型在工序链中的角色。
+>
+> 2. **系统提示词是完整的工艺规约**：`instruction[directive]` 一次性承载了角色定义、图片数量决策树、视觉约束（包含明确正反例："禁止印刷体/禁止写实"）、输出格式模板四层内容，是 prompt 工程中"工艺规约即 directive"的典型范式。
+>
+> 3. **`via = <llm>` 的工具无关性**：原帖不指定 LLM，核心竞争力在 directive 质量而非工具选择。DSL 的通用占位符准确捕捉了这种"工具无关、prompt 有关"的工艺模式。
+>
+> 4. **遍历结构不可拍平**：LLM 一次输出 N 个策划块（N 动态决定），每块独立送图像生成，必须建模为 `block + nested` 循环，不能表达为单步批量操作。
+
+
+
+### ◀ Result · 14:09:07
+
+- subtype: `success` · is_error: `False`
+- num_turns: `29` · duration: `1153088ms` · wall: `1155.1s`
+- tokens: in=22 out=55,525 cache_w=141,739 cache_r=2,166,366
+- cost: sdk=$2.7459, est_if_api=$2.0144

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/case-A_xhs_69663d9d-xhs-simpledraw-infographic.md

@@ -0,0 +1,216 @@
+# Case A_xhs_69663d9d: 小红书信息图系列生成
+
+**Source**: 小红书 · 御风 · 2026-01-13  
+**URL**: https://www.xiaohongshu.com/explore/69663d9d000000000a02adc4  
+**主题**: 分享一套 LLM 系统提示词，将任意文本内容拆解为小红书简笔画素描风格系列信息图
+
+---
+
+## 工序梗概 (人话)
+
+本 case 包含 **1 个工序**：**小红书风格信息图系列生成**。
+
+工序分两阶段：① 将用户的文本内容（教程/知识点/经验分享）连同一套精心设计的系统提示词一并喂给 LLM，由 LLM 以"小红书视觉内容策划师"角色输出 N 个结构化图片策划块（每块含图片定位/核心信息/文字内容/视觉提示词）；② 遍历每个策划块，将其视觉提示词部分输入 NanoBanana，生成对应的简笔画素描风格信息图。最终返回 N 张构成完整系列的小红书信息图。
+
+核心知识资产是这套系统提示词本身，它定义了角色、拆解原则、图片数量判断规则、视觉风格约束（竖版 3:4、手写体、黑白灰+单彩色、无写实元素）和固定输出格式，是整个工序的工艺规约。
+
+---
+
+## 引用的类型 (stdlib + 自定义)
+
+### stdlib 类型 (直接引用)
+
+无直接引用 stdlib 叶子（所有 IO 类型均为 case-specific）
+
+> 注：`文本` 也作为 case-specific 注册（extends `正文`），因为 type.json 中 `正文` 是其最近 stdlib 叶子；`提示词`、`成品图` 是作为 extends 桥接目标的 stdlib 叶子（被 type_registry 引用，本身不直接出现为 IO type）。
+
+### case-specific 类型 (type_registry)
+
+| 类型名 | extends | 描述 |
+|--------|---------|------|
+| 文本 | 正文 | 用户提供的待可视化原始文字内容（如教程/知识点/经验分享），LLM 内容分析的输入素材 |
+| 图片策划序列 | 提示词 | LLM 输出的 N 个结构化图片策划块的有序序列，每块含图片定位/核心信息/文字内容/视觉提示词 |
+| 图片策划 | 提示词 | 单张信息图的结构化策划块，含图片定位/核心信息/文字内容（主标题+副标题+补充说明）/视觉提示词 |
+| 信息图 | 成品图 | NanoBanana 生成的小红书简笔画素描风格单张信息图（竖版 3:4，手写体文字），终态成品直接交付 |
+| 信息图序列 | 成品图 | N 张小红书简笔画风格信息图的有序列表，构成完整信息图系列，终态成品集合直接交付 |
+
+---
+
+## L1 外部函数库
+
+| 工具名 | 类型 | 说明 |
+|--------|------|------|
+| `<llm>` | 通用 LLM 占位符 | 接受系统提示词+用户内容，以"策划师"角色输出结构化图片策划序列；具体模型不指定，任意 LLM 均可 |
+| `nano_banana` | AI 图像生成 | 接受视觉提示词，生成简笔画素描风格信息图；产品即 NanoBanana |
+
+---
+
+## L2 抽象动作 + impl 关系
+
+| 动作路径 | 抽象语义 | impl |
+|----------|---------|------|
+| `生成/关系生成/数组生成` | 将单一输入按结构化规则拆解，生成多个输出单元组成的有序数组 | via `<llm>` |
+| `生成/元素生成` | 按视觉提示词生成单个独立图像元素 | via `nano_banana` |
+
+---
+
+## L3 工序模板
+
+### 模板：小红书风格信息图系列生成 (p1)
+
+**purpose**: 将任意文本内容拆解为小红书简笔画风格系列信息图  
+**category**: 产物创造 | **platform**: xhs | **author**: 御风
+
+**声明输入**:
+- `文本` — `待可视化内容`：需要被拆解为系列信息图的文本内容
+
+**返回**: `信息图序列`
+
+**步骤**:
+
+```
+s1  [step] 用 <llm> 数组生成
+    in:  待可视化内容 (文本)  ← 工序输入
+    out: 图片策划序列         → s2
+
+s2  [block | 遍历]
+    in:  图片策划序列         ← s1.图片策划序列
+    out: 信息图序列           → 返回 信息图序列
+
+  s2.1  [nested] 用 nano_banana 元素生成
+        in:  图片策划         ← 图片策划序列[i]
+        out: 信息图           → 信息图序列.追加
+```
+
+---
+
+## L4 工序实例
+
+### P1 实例 — 小红书风格信息图系列生成
+
+**inputs (本次实际值)**
+
+| 变量名 | 类型 | 值 |
+|--------|------|---|
+| 待可视化内容 | 文本 | `<用户提供的待可视化文本内容，如某主题的教程、知识点或经验分享>` |
+
+*注：原帖配图展示的 5 张信息图为此工序在「小红书简笔画信息图制作指南」主题内容上的运行实例产出*
+
+**bindings (本次工具选择)**
+
+- `s1.via` = `<llm>`（通用 LLM，原帖未指定具体模型）
+- `s2.1.via` = `nano_banana`（NanoBanana AI 图像生成工具）
+
+**extracted_values (本次中间产物)**
+
+**s1 directive — 系统提示词原文**：
+
+```
+角色定义：
+你是一位专业的小红书视觉内容策划师，擅长将复杂内容拆解为简笔画素描风格系列信息图。
+
+任务：
+请分析以下输入内容，将其拆解为 1-10 张小红书风格的系列信息图，并为每张图片输出独立的生成提示词。
+
+拆解原则：
+1. 封面图（第1张）：必须有强烈视觉冲击力，包含核心标题和吸引点
+2. 内容图（中间）：每张聚焦1个核心观点，信息密度适中
+3. 结尾图（最后1张）：总结/行动号召/金句收尾
+
+图片数量判断标准：
+- 简单观点/单一主题：2-3 张
+- 中等复杂度/教程类：4-6 张
+- 深度干货/多维度分析：7-10 张
+
+视觉风格规范（重要）
+
+基础设定：
+- 图片类型：信息图（Infographic）
+- 方向比例：竖版，3:4 或 9:16
+- 整体风格：简笔画素描风格（Simple Line Sketch Style）
+
+背景与配色：
+- 背景：纯白/米白/浅灰/牛皮纸色
+- 配色：黑白灰为主，可点缀单一彩色
+- 线条：黑色/深灰，干净利落
+
+文字风格：
+- 必须使用简笔画手写体文字
+- 大标题突出醒目，可加粗放大
+- 可使用下划线、圈注强调关键词
+- 禁止印刷体或复杂字体
+
+装饰元素：
+- 加入少量简笔画图标、火柴人、简单物品轮廓
+- 可使用：手绘箭头、简单边框、对话气泡
+- 人物用火柴人或简单轮廓，禁止写实
+- 线条特点：简洁明了、一笔成型
+
+排版原则：
+- 信息精简，突出关键词与核心概念
+- 多留白，易于一眼抓住重点
+- 要点分条呈现，层次清晰
+
+输出格式：
+对于每张图片，请按以下结构输出：
+---
+第 X 张 / 共 N 张
+图片定位：[封面图 / 内容图 / 结尾图]
+核心信息：[这张图要传达的1句话核心]
+
+文字内容：
+- 主标题：xxx
+- 副标题/要点：xxx
+- 补充说明（如有）：xxx
+
+视觉提示词：
+小红书风格信息图，竖版（3:4），简笔画素描风格（Simple Line Sketch），手写体文字，[具体背景色]背景。
+[具体内容布局描述]
+加入简笔画图标和简单手绘元素：[具体元素描述]
+整体风格：简笔画、黑白灰为主、简洁干净，信息精简，多留白。所有元素为简笔画风格，线条简单一笔成型，无写实元素。
+右下角水印："御风"
+```
+
+**s1 output 示例**（原帖配图示范的5张策划块）：
+
+第 1 张 / 共 5 张（封面图）：
+- 图片定位：封面图 — 核心信息：小红书简笔画信息图制作指南
+- 视觉提示词（示例）：小红书风格信息图，竖版（3:4），简笔画素描风格，手写体大标题「小红书简笔画信息图制作指南」，平板电脑+手写笔为主图，周围散落：咖啡杯、笔记本、火柴人、思考气泡等简笔画图标，牛皮纸色背景，右下角水印"御风"
+
+第 2 张 / 共 5 张（内容图）：
+- 核心信息：核心步骤—拆解三原则（封面图→内容图→结尾图）
+
+第 3 张 / 共 5 张（内容图）：
+- 核心信息：图片数量判断（简单2-3/中等4-6/深度7-10）+ 视觉规范
+
+第 4 张 / 共 5 张（内容图）：
+- 核心信息：正确示范（手写体✓）vs 错误示范（印刷体✗、写实人物✗）
+
+第 5 张 / 共 5 张（结尾图）：
+- 核心信息：标准输出格式卡（图片定位/核心信息/文字内容/视觉提示词）+ 行动号召
+
+**s2.1 directive 模板**（从视觉提示词段落提取）：
+
+```
+小红书风格信息图，竖版（3:4），简笔画素描风格（Simple Line Sketch），手写体文字，[具体背景色]背景。
+[具体内容布局描述]
+加入简笔画图标和简单手绘元素：[具体元素描述]
+整体风格：简笔画、黑白灰为主、简洁干净，信息精简，多留白。所有元素为简笔画风格，线条简单一笔成型，无写实元素。
+右下角水印："御风"
+```
+
+**trace**: （原帖未提供运行时间/费用数据）
+
+---
+
+## 这个 case 对 DSL 设计的关键启发
+
+1. **系统提示词 = 完整工艺规约的数据模型**：这套提示词融合了角色定义（L2 动作的执行者）、工序参数（图片数量决策树）、风格约束（视觉规范）和输出模板（格式定义）四层内容于一体，是"工艺规约"在 prompt 形式上的完整表达。DSL 中 `instruction[directive]` 字段承载了这个完整的规约文本。
+
+2. **正反例约束的 prompt 工程模式**：配图第4张明确用"正确示范 vs 错误示范"对比展示核心约束（手写体✓ vs 印刷体✗，简笔画✓ vs 写实人物✗）。这说明视觉风格约束更多依赖*负例排除*而非正例描述——prompt 中"禁止写实"、"禁止印刷体"是关键约束词，而非可选风格描述。DSL 的 directive 字段若区分正向 constraint vs 负向 constraint (exclusion) 会更精确。
+
+3. **N 个结构化块的遍历不能拍平**：LLM 单次调用产出 N 个策划块（N 由内容复杂度动态决定），每个块独立送图像生成——这是典型的 `block + nested` 遍历结构。如果拍平为"批量生成步骤"会丢失"每张图独立生成"的工艺语义。
+
+4. **`via = <llm>` 的工具无关性设计**：原帖没有指定具体 LLM，提示词本身已经自洽。这说明某些工序的核心不在 L1 工具选择（哪个模型）而在 directive 设计（系统提示词质量）。DSL 中 `via = <llm>` 占位符正确捕捉了这种"工具无关、prompt 有关"的工艺模式。
+
+5. **信息图 type 的归属：extends `成品图` 而非 `参考图`**：NanoBanana 生成的信息图是直接交付给用户的终态产物，不是用于参考/模仿的素材。2A 归一化发现原始骨架中误用了 `参考图` 作为 extends，纠正为 `成品图` 更准确反映了该类型在工序链中的角色（生产目标，非生产素材）。

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_69663d9d/understanding.md

@@ -0,0 +1,72 @@
+# Phase 1.1 心智模型 — A_xhs_69663d9d
+
+## 工序数量判断
+
+本 case 含 **1 个独立工序**。
+
+**判断依据**：
+- 原文提供一套完整 LLM 系统提示词模板，核心价值是这套模板本身
+- 唯一的输入→终态产出链：文本内容 → LLM（用模板）生成 N 个图片策划 → NanoBanana 生成 N 张信息图
+- 无第二条不同工艺链，无第二类终态产物
+- 配图5张是该工序在某输入上的实际运行示例（信息图系列输出示范）
+
+## 工序概述
+
+**工序名**: 小红书风格信息图系列生成
+
+**终态产物**: 小红书简笔画素描风格系列信息图（N 张，N ∈ [1,10]）
+
+**大致步骤数**: 3 个节点（s1 + s2 控制块 + s2.1 嵌套执行步）
+
+**工艺类型**: 文本→提示词→图像，两阶段转换（LLM 规划 + 图像生成）
+
+## 步骤识别
+
+### s1：内容拆解与图片策划（LLM 执行）
+
+输入用户的待可视化文本内容 + 系统提示词模板（作为 directive 喂给 LLM）。
+
+LLM 扮演"小红书视觉内容策划师"角色，按模板规则拆解内容为 N 个结构化图片策划块。每块包含：
+- 图片定位（封面图 / 内容图 / 结尾图）
+- 核心信息
+- 文字内容（主标题 / 副标题/要点 / 补充说明）
+- 视觉提示词（Small Line Sketch 风格的图像生成 prompt）
+
+输出：图片策划序列（有序的 N 个策划块）
+
+### s2（block，遍历）：逐张生成信息图
+
+遍历图片策划序列，对每个策划块执行一次 NanoBanana 图像生成。
+
+### s2.1（nested）：单张信息图生成（NanoBanana）
+
+将单个图片策划（的视觉提示词部分）作为 directive 输入 NanoBanana，生成一张小红书风格信息图。输出追加到信息图序列。
+
+## 外部工具依赖
+
+- `<llm>`（任意 LLM，原文未指定具体模型）：s1 内容拆解与策划
+- `nano_banana`（NanoBanana，AI 图像生成工具）：s2.1 图像生成
+
+## 隐含工艺规约
+
+1. **LLM 角色定义是核心**：系统提示词定义了"专业的小红书视觉内容策划师"角色，这是工序的知识资产
+2. **图片数量由复杂度决定**：简单 2-3 张 / 中等 4-6 张 / 深度 7-10 张
+3. **视觉风格强约束**：竖版 3:4、简笔画素描风格、手写体文字、黑白灰+单彩色点缀、多留白、无写实元素
+4. **固定输出格式**：每张图策划块含4个分区（定位/核心信息/文字内容/视觉提示词），且视觉提示词包含右下角"御风"水印
+
+## 变量追踪
+
+| 变量名 | 类型 | 流转 |
+|--------|------|------|
+| 待可视化内容 | 文本 | 工序输入 → s1.inputs |
+| 图片策划序列 | 图片策划序列（case-specific） | s1.outputs → s2.inputs |
+| 图片策划（循环变量） | 图片策划（case-specific） | 图片策划序列[i] → s2.1.inputs |
+| 信息图（单次产出） | 信息图（case-specific） | s2.1.outputs → 信息图序列.追加 |
+| 信息图序列 | 信息图序列（case-specific） | s2.outputs → 返回 |
+
+## 关键发现
+
+1. **系统提示词是工序的灵魂**：原文的核心是这套精心设计的系统提示词，包含角色定义+拆解原则+视觉规范+输出格式，构成完整的工艺规约；用户使用此工序时真正"输入"的只是待可视化内容，系统提示词已预设在 directive 中
+2. **两阶段架构**：LLM 负责内容→结构化策划（策划师角色），NanoBanana 负责策划→视觉图像；中间产物是结构化文本（图片策划序列），非图像
+3. **遍历循环是必要结构**：单次 LLM 调用产生 N 个策划块，每个独立送图像生成，自然形成遍历（不能拍平为"批量生成"单步）
+4. **风格限制是负例学习**：配图第4张明确对比"手写体✓ vs 印刷体✗"、"简笔画✓ vs 写实人物✗"，说明风格约束是有意识设计的核心约束，非审美偏好

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/.session_id

@@ -0,0 +1 @@
+240b8a33-638d-4206-bc0c-899512a2092d

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/_meta.json

@@ -0,0 +1,8 @@
+{
+  "case_id": "xhs_6968aa4d000000000a03329c",
+  "from_q": "q0018",
+  "form": "A",
+  "score": 8.9,
+  "score_norm": 0.89,
+  "started_at": "2026-06-03T17:07:16"
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/_scratch/anchors.json

@@ -0,0 +1,10 @@
+[
+  {"path": "p1.s1.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p1.s1.inputs[1].anchor",  "value": "← 工序输入"},
+  {"path": "p1.s1.outputs[0].anchor", "value": "→ s2"},
+  {"path": "p1.s2.inputs[0].anchor",  "value": "← s1.反推提示词"},
+  {"path": "p1.s2.inputs[1].anchor",  "value": "← 工序输入"},
+  {"path": "p1.s2.outputs[0].anchor", "value": "→ s3"},
+  {"path": "p1.s3.inputs[0].anchor",  "value": "← s2.生图配置"},
+  {"path": "p1.s3.outputs[0].anchor", "value": "→ 返回 产品海报"}
+]

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/_scratch/patch_2a.json

@@ -0,0 +1,20 @@
+[
+  {"path": "p1.s1.effect", "value": "预处理"},
+  {"path": "p1.s1.action", "value": "提取/化学提取/反推"},
+  {"path": "p1.s1.inputs[0].type", "value": "参考图"},
+  {"path": "p1.s1.inputs[1].type", "value": "产品图"},
+  {"path": "p1.s1.outputs[0].type", "value": "提示词"},
+  {"path": "p1.s2.effect", "value": "预处理"},
+  {"path": "p1.s2.action", "value": "获取/录入/键入"},
+  {"path": "p1.s2.inputs[0].type", "value": "提示词"},
+  {"path": "p1.s2.inputs[1].type", "value": "产品图"},
+  {"path": "p1.s2.outputs[0].type", "value": "工具配置"},
+  {"path": "p1.s3.effect", "value": "主体生成"},
+  {"path": "p1.s3.action", "value": "生成/元素生成"},
+  {"path": "p1.s3.inputs[0].type", "value": "工具配置"},
+  {"path": "p1.s3.outputs[0].type", "value": "成品图"},
+  {"path": "p1.type_registry.产品图.extends", "value": "参考图"},
+  {"path": "p1.type_registry.产品图.desc", "value": "本次工序中待迁移风格的目标产品图（雅诗兰黛 Re-Nutriv 护肤水官方图），被 AI 分析以提取目标产品的外观特征，并在生图时作为图像参考上传到 Jimeng；其角色是'被模仿/被迁移风格的目标实体'，而非单纯的风格范本，因此以 case-specific 类型独立标注。"},
+  {"path": "p1.type_registry.工具配置.extends", "value": "生成参数"},
+  {"path": "p1.type_registry.工具配置.desc", "value": "Jimeng 文生图工具的综合生图配置集合，包含：已粘贴的反推提示词、已上传的目标产品参考图、画面比例（2:3）、参考强度（100%）、使用模型（图片 4.0）、输出规格（高清 2K）；属于本次生图任务的完整参数束，超出单一生成参数范畴故作 case-specific 注册。"}
+]

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/_scratch/patch_2b.json

@@ -0,0 +1,66 @@
+[
+  {
+    "path": "p1.s1.inputs[0].substance",
+    "value": "/表象/视觉/实体/物品/起居物品/美妆护肤"
+  },
+  {
+    "path": "p1.s1.inputs[0].form",
+    "value": "/呈现/视觉/视觉风格/视觉艺术风格/现代气质/科技未来 + /呈现/视觉/视觉制作/后期处理/图像合成处理/三维建模合成"
+  },
+  {
+    "path": "p1.s1.inputs[1].substance",
+    "value": "/表象/视觉/实体/物品/起居物品/美妆护肤"
+  },
+  {
+    "path": "p1.s1.inputs[1].form",
+    "value": "/呈现/视觉/视觉制作/实景拍摄/拍摄主体/实物呈现"
+  },
+  {
+    "path": "p1.s1.outputs[0].substance",
+    "value": "/理念/知识/商业/前沿技术/AI智能/AI应用"
+  },
+  {
+    "path": "p1.s1.outputs[0].form",
+    "value": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"
+  },
+  {
+    "path": "p1.s2.inputs[0].substance",
+    "value": "/理念/知识/商业/前沿技术/AI智能/AI应用"
+  },
+  {
+    "path": "p1.s2.inputs[0].form",
+    "value": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"
+  },
+  {
+    "path": "p1.s2.inputs[1].substance",
+    "value": "/表象/视觉/实体/物品/起居物品/美妆护肤"
+  },
+  {
+    "path": "p1.s2.inputs[1].form",
+    "value": "/呈现/视觉/视觉制作/实景拍摄/拍摄主体/实物呈现"
+  },
+  {
+    "path": "p1.s2.outputs[0].substance",
+    "value": "__null__"
+  },
+  {
+    "path": "p1.s2.outputs[0].form",
+    "value": "__null__"
+  },
+  {
+    "path": "p1.s3.inputs[0].substance",
+    "value": "__null__"
+  },
+  {
+    "path": "p1.s3.inputs[0].form",
+    "value": "__null__"
+  },
+  {
+    "path": "p1.s3.outputs[0].substance",
+    "value": "/表象/视觉/实体/物品/起居物品/美妆护肤"
+  },
+  {
+    "path": "p1.s3.outputs[0].form",
+    "value": "/呈现/视觉/视觉制作/后期处理/图像合成处理/AI生成合成 + /呈现/视觉/视觉风格/视觉艺术风格/现代气质/科技未来 + /呈现/视觉/视觉制作/构图编排/版面设计/版面结构"
+  }
+]

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/_scratch/task_2a.json

@@ -0,0 +1,99 @@
+{
+  "case_id": "A_xhs_6968aa4d",
+  "task": "Phase 2A (effect/action/type normalization)",
+  "image_url_list": [],
+  "steps": [
+    {
+      "path": "p1.s1",
+      "name": "",
+      "action": "分析参考图并反推文生图提示词",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p1.s1.inputs[0]",
+          "name": "参考图",
+          "value": "<魔法世家蓝色精华水产品CG风格图：倾斜悬浮特写构图，被透明流体光带环绕，背景浅灰渐变，科技感美妆氛围，银色金属瓶盖>",
+          "type": "参考图",
+          "related_images": []
+        },
+        {
+          "path": "p1.s1.inputs[1]",
+          "name": "目标产品图",
+          "value": "<雅诗兰黛Re-Nutriv护肤水官方图：瓶身为透明金色玻璃材质（内含浅金色液体），搭配金色金属质感瓶盖；瓶身印有'Re-Nutriv''ESTÉE LAUDER'字样及产品说明，风格雍容高贵>",
+          "type": "产品图",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p1.s1.outputs[0]",
+          "name": "反推提示词",
+          "value": "未来感极简美妆 CG 风格，倾斜悬浮特写：雅诗兰黛 Re-Nutriv 护肤水（通透金色玻璃瓶身 + 浅金色液体，配金色金属瓶盖，印'Re-Nutriv''ESTÉE LAUDER'字样）被透明流体光带环绕托举；背景浅灰渐变，色调清冷奢华，突出瓶身通透感、瓶盖金属质感与流体光带动态纹理，营造科技感高端护肤氛围",
+          "type": "提示词",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p1.s2",
+      "name": "",
+      "action": "在Jimeng中输入提示词并配置生图参数",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p1.s2.inputs[0]",
+          "name": "反推提示词",
+          "value": "未来感极简美妆 CG 风格，倾斜悬浮特写：雅诗兰黛 Re-Nutriv 护肤水（通透金色玻璃瓶身 + 浅金色液体，配金色金属瓶盖，印'Re-Nutriv''ESTÉE LAUDER'字样）被透明流体光带环绕托举；背景浅灰渐变，色调清冷奢华，突出瓶身通透感、瓶盖金属质感与流体光带动态纹理，营造科技感高端护肤氛围",
+          "type": "提示词",
+          "related_images": []
+        },
+        {
+          "path": "p1.s2.inputs[1]",
+          "name": "目标产品图",
+          "value": "<雅诗兰黛Re-Nutriv护肤水官方图：瓶身为透明金色玻璃材质（内含浅金色液体），搭配金色金属质感瓶盖；瓶身印有'Re-Nutriv''ESTÉE LAUDER'字样及产品说明，风格雍容高贵>",
+          "type": "产品图",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p1.s2.outputs[0]",
+          "name": "生图配置",
+          "value": "<Jimeng生图配置：反推提示词已粘贴，目标产品图已上传，画面比例2:3，参考强度100%，图片4.0模型，高清2K>",
+          "type": "工具配置",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p1.s3",
+      "name": "",
+      "action": "文生图生成并人工选图",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p1.s3.inputs[0]",
+          "name": "生图配置",
+          "value": "<Jimeng生图配置：反推提示词已粘贴，目标产品图已上传，画面比例2:3，参考强度100%，图片4.0模型，高清2K>",
+          "type": "工具配置",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p1.s3.outputs[0]",
+          "name": "产品海报",
+          "value": "<雅诗兰黛Re-Nutriv护肤水高端产品海报：倾斜悬浮CG风格，透明流体光带环绕瓶身，背景浅灰渐变底色，清冷金色调（金色+透明+浅灰），高清2K，竖版（2:3比例）>",
+          "type": "图像",
+          "related_images": []
+        }
+      ]
+    }
+  ]
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/_scratch/task_2b.json

@@ -0,0 +1,63 @@
+{
+  "case_id": "A_xhs_6968aa4d",
+  "task": "Phase 2B (substance/form matching)",
+  "image_url_list": [],
+  "items_to_match": [
+    {
+      "path": "p1.s1.inputs[0]",
+      "name": "参考图",
+      "value": "<魔法世家蓝色精华水产品CG风格图：倾斜悬浮特写构图，被透明流体光带环绕，背景浅灰渐变，科技感美妆氛围，银色金属瓶盖>",
+      "type": "参考图",
+      "related_images": []
+    },
+    {
+      "path": "p1.s1.inputs[1]",
+      "name": "目标产品图",
+      "value": "<雅诗兰黛Re-Nutriv护肤水官方图：瓶身为透明金色玻璃材质（内含浅金色液体），搭配金色金属质感瓶盖；瓶身印有'Re-Nutriv''ESTÉE LAUDER'字样及产品说明，风格雍容高贵>",
+      "type": "产品图",
+      "related_images": []
+    },
+    {
+      "path": "p1.s1.outputs[0]",
+      "name": "反推提示词",
+      "value": "未来感极简美妆 CG 风格，倾斜悬浮特写：雅诗兰黛 Re-Nutriv 护肤水（通透金色玻璃瓶身 + 浅金色液体，配金色金属瓶盖，印'Re-Nutriv''ESTÉE LAUDER'字样）被透明流体光带环绕托举；背景浅灰渐变，色调清冷奢华，突出瓶身通透感、瓶盖金属质感与流体光带动态纹理，营造科技感高端护肤氛围",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p1.s2.inputs[0]",
+      "name": "反推提示词",
+      "value": "未来感极简美妆 CG 风格，倾斜悬浮特写：雅诗兰黛 Re-Nutriv 护肤水（通透金色玻璃瓶身 + 浅金色液体，配金色金属瓶盖，印'Re-Nutriv''ESTÉE LAUDER'字样）被透明流体光带环绕托举；背景浅灰渐变，色调清冷奢华，突出瓶身通透感、瓶盖金属质感与流体光带动态纹理，营造科技感高端护肤氛围",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p1.s2.inputs[1]",
+      "name": "目标产品图",
+      "value": "<雅诗兰黛Re-Nutriv护肤水官方图：瓶身为透明金色玻璃材质（内含浅金色液体），搭配金色金属质感瓶盖；瓶身印有'Re-Nutriv''ESTÉE LAUDER'字样及产品说明，风格雍容高贵>",
+      "type": "产品图",
+      "related_images": []
+    },
+    {
+      "path": "p1.s2.outputs[0]",
+      "name": "生图配置",
+      "value": "<Jimeng生图配置：反推提示词已粘贴，目标产品图已上传，画面比例2:3，参考强度100%，图片4.0模型，高清2K>",
+      "type": "工具配置",
+      "related_images": []
+    },
+    {
+      "path": "p1.s3.inputs[0]",
+      "name": "生图配置",
+      "value": "<Jimeng生图配置：反推提示词已粘贴，目标产品图已上传，画面比例2:3，参考强度100%，图片4.0模型，高清2K>",
+      "type": "工具配置",
+      "related_images": []
+    },
+    {
+      "path": "p1.s3.outputs[0]",
+      "name": "产品海报",
+      "value": "<雅诗兰黛Re-Nutriv护肤水高端产品海报：倾斜悬浮CG风格，透明流体光带环绕瓶身，背景浅灰渐变底色，清冷金色调（金色+透明+浅灰），高清2K，竖版（2:3比例）>",
+      "type": "图像",
+      "related_images": []
+    }
+  ]
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/_source.json

@@ -0,0 +1,15 @@
+{
+  "title": "如何让AI抄参考图？（附指令词）",
+  "link": "https://www.xiaohongshu.com/explore/6968aa4d000000000a03329c",
+  "body_text": "一、反推提示词\n拿到参考图后，拆解元素并替换产品：\n💡指令参考：”分析并描述图1的主体内容、场景设定、风格参考、色彩色调、构图视角、细节等信息。描述图2的产品外观、材质结构。并将图2的产品信息添加到图1，然后生成一段用于文生图的提示词。“\n示例提示词：未来感极简美妆 CG 风格，倾斜悬浮特写：雅诗兰黛 Re-Nutriv 护肤水（通透金色玻璃瓶身 + 浅金色液体，配金色金属瓶盖，印 “Re-Nutriv”“ESTÉE LAUDER” 字样）被透明流体光带环绕托举；背景浅灰渐变，色调清冷奢华，突出瓶身通透感、瓶盖金属质感与流体光带动态纹理，营造科技感高端护肤氛围\n二、AI 工具操作\n打开 jimeng：\n粘贴提示词，上传产品图\n调整画面比例为 2:3（适配竖版海报）\n点击产品图，将参考强度设为100%\n三、生成选图\n点击 “生成”，多次抽卡后选择符合以下特征的海报：\n流体光带的动态自然不生硬\n产品细节（瓶身通透感、文字标识）清晰\n色调贴合 “清冷奢华” 的科技感\n这类海报属于高端护肤的科技风视觉，通过 “流体光带+悬浮构图” 突出产品的渗透力与修护功效，适合用于品牌高端线宣传、电商详情页，能快速传递 “科技护肤” 的专业感。\n#电商设计[话题]# #作图神器[话题]# #设计技巧[话题]# #ai工具[话题]# #产品拍摄[话题]# #ai关键词[话题]# #AI教程[话题]# #海报设计[话题]# #小红书设计周[话题]# #设计教程[话题]#",
+  "images": [
+    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k031rbt112s0m6g5q8b87gdoctnikl0c48?imageView2/2/w/1080/format/webp",
+    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831rbt1153n4d05q8b87gdoctn9tgdt3o?imageView2/2/w/1080/format/webp",
+    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831rbt117p0m8g5q8b87gdoctn3ebhbvg?imageView2/2/w/1080/format/webp",
+    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831rbt117p0m905q8b87gdoctn3iapdd8?imageView2/2/w/1080/format/webp",
+    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831rbt117p0m9g5q8b87gdoctnehicamo?imageView2/2/w/1080/format/webp",
+    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k831rbt117p0mag5q8b87gdoctnu2bq7k8?imageView2/2/w/1080/format/webp"
+  ],
+  "publish_timestamp": "2026-01-15 16:50:21",
+  "channel_account_name": "xhs"
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/_trace.md

@@ -0,0 +1,533 @@
+
+
+---
+
+## ▶ Fresh @ 2026-06-03 17:07:20
+
+- case: `A_xhs_6968aa4d`
+- source: `C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\runs_full\q0018\procedures\A_xhs_6968aa4d\_source.json`
+- spec: `spec`
+- images: `6`
+- model: `claude-sonnet-4-6`
+- max_turns: `300`
+
+_session=240b8a33 · model='claude-sonnet-4-6' · apiKeySource='none'_
+
+
+### Turn 1 · 17:07:27
+
+
+### Turn 2 · 17:07:29
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/README.md`
+
+
+### Turn 3 · 17:07:30
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/_source.json`
+
+
+### Turn 4 · 17:07:35
+
+
+### Turn 5 · 17:07:35
+> Now let me read all Phase 1 spec files in parallel:
+
+
+
+### Turn 6 · 17:07:37
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools.md`
+
+
+### Turn 7 · 17:07:37
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/syntax.md`
+
+
+### Turn 8 · 17:07:38
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/fields.md`
+
+
+### Turn 9 · 17:07:44
+
+
+### Turn 10 · 17:07:46
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/control-flow.md`
+
+
+### Turn 11 · 17:07:50
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/phase1-skeleton.md`
+
+
+### Turn 12 · 17:07:50
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/format/md-structure.md`
+
+
+### Turn 13 · 17:07:54
+
+
+### Turn 14 · 17:07:54
+> Now let me read the workflow template:
+
+
+
+### Turn 15 · 17:07:56
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/templates/workflow.template.json`
+
+
+### Turn 16 · 17:15:31
+
+
+### Turn 17 · 17:15:31
+> Now I have all the Phase 1 information I need. Let me write the understanding.md and workflow.json simultaneously:
+
+
+
+### Turn 18 · 17:15:57
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/understanding.md` (1,683 chars)
+
+
+### Turn 19 · 17:16:05
+
+
+### Turn 20 · 17:16:05
+> Now write the workflow.json skeleton:
+
+
+
+### Turn 21 · 17:16:44
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/workflow.json` (5,357 chars)
+
+
+### Turn 22 · 17:17:06
+
+
+### Turn 23 · 17:17:06
+> Now Phase 1.3 — add anchors via wf-patch. First write the anchor patch list to _scratch:
+
+
+
+### Turn 24 · 17:17:09
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/_scratch/anchors.json` (473 chars)
+
+
+### Turn 25 · 17:17:13
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/wf-patch.py --workflow "C:/Users/11304/gitlab/cybertogether/Agent/example...`
+
+
+### Turn 26 · 17:17:24
+
+
+### Turn 27 · 17:17:24
+> All 8 anchors set. Now run `--resolve-passthrough` to auto-fill transparent value transfers:
+
+
+
+### Turn 28 · 17:17:26
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/wf-patch.py --workflow "C:/Users/11304/gitlab/cybertogether/Agent/example...`
+
+
+### Turn 29 · 17:17:33
+
+
+### Turn 30 · 17:17:33
+> Good — 2 passthrough fills. The `← 工序输入` for s2.inputs[1] wasn't auto-resolved (it points to declarations, not a step). Let me set it manually:
+
+
+
+### Turn 31 · 17:17:37
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/wf-patch.py --workflow "C:/Users/11304/gitlab/cybertogether/Agent/example...`
+
+
+### Turn 32 · 17:18:02
+
+
+### Turn 33 · 17:18:02
+> Phase 1 complete. Now load Phase 2 spec:
+
+
+
+### Turn 34 · 17:18:03
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/phase2-normalize.md`
+
+
+### Turn 35 · 17:18:04
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/taxonomy/type_suggestions.md`
+
+
+### Turn 36 · 17:18:23
+
+
+### Turn 37 · 17:18:23
+> Now I need to run the prepare-subtask.py script to generate sub-task files. Let me check if it exists:
+
+
+
+### Turn 38 · 17:18:24
+- `Bash` — `ls "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/"`
+
+
+### Turn 39 · 17:18:28
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/prepare-subtask.py --workflow "C:/Users/11304/gitlab/cybertogether/Agent/...`
+
+
+### Turn 40 · 17:18:36
+
+
+### Turn 41 · 17:18:36
+> Sub-task files generated (3 steps, 8 IO items). Now dispatch Phase 2A and 2B sub-agents in parallel:
+
+
+
+### Turn 42 · 17:18:52
+- `Agent` — `Phase 2A: normalize effect/action/type for A_xhs_6968aa4d` [subagent=phase-2a-normalizer]
+
+
+### Turn 43 · 17:18:54
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/_scratch/task_2a.json`
+
+
+### Turn 44 · 17:18:56
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/taxonomy/effect.json`
+
+
+### Turn 45 · 17:18:56
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/taxonomy/action.json`
+
+
+### Turn 46 · 17:18:56
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/taxonomy/type.json`
+
+
+### Turn 47 · 17:19:08
+- `Agent` — `Phase 2B: substance/form taxonomy matching for A_xhs_6968aa4d` [subagent=phase-2b-matcher]
+
+
+### Turn 48 · 17:19:10
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/_scratch/task_2b.json`
+
+
+### Turn 49 · 17:19:15
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "产品图 美妆 护肤水 参考风格 CG风格"`
+
+
+### Turn 50 · 17:19:17
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "产品图 CG风格 倾斜悬浮 科技感 精华水"`
+
+
+### Turn 51 · 17:19:20
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "产品图 护肤品 高端美妆 雅诗兰黛 实物"`
+
+
+### Turn 52 · 17:19:21
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "产品图 实物拍摄 官方图 玻璃瓶"`
+
+
+### Turn 53 · 17:19:25
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "文生图提示词 图像描述 AI生图 场景描述"`
+
+
+### Turn 54 · 17:19:26
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "提示词 文本 指令文本 AI提示"`
+
+
+### Turn 55 · 17:19:29
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "提示词 AI指令 图像生成描述 prompt"`
+
+
+### Turn 56 · 17:19:31
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "AI工具 生图工具 AIGC 图像生成"`
+
+
+### Turn 57 · 17:19:35
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "AI生图 Prompt 文生图 图像描述文本"`
+
+
+### Turn 58 · 17:19:36
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "产品海报 AI生成图 高端美妆 护肤品广告"`
+
+
+### Turn 59 · 17:19:40
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --narrow "产品海报 AI生成图 高端美妆 护肤品广告"`
+
+
+### Turn 60 · 17:19:41
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/_scratch/patch_2a.json` (1,212 chars)
+
+
+### Turn 61 · 17:19:41
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "产品海报 AI生成图 CG风格 竖版海报 高清"`
+
+
+### Turn 62 · 17:19:44
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "产品海报 商业宣传 高端护肤 美妆广告图"`
+
+
+### Turn 63 · 17:19:46
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "CG渲染 AI生成图像 数字艺术 三维渲染"`
+
+
+### Turn 64 · 17:19:50
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "海报 宣传物料 商业美术 美妆"`
+
+
+### Turn 65 · 17:19:51
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "AI生成图 合成图像 数字渲染 科技感画面"`
+
+
+### Turn 66 · 17:19:54
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "护肤品 美妆 产品宣传图"`
+
+
+### Turn 67 · 17:19:55
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "后期合成 图像合成 特效制作"`
+
+
+### Turn 68 · 17:19:58
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "AI生成 AIGC图像 AI合成"`
+
+
+### Turn 69 · 17:20:00
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "参考图 风格迁移 美妆图片 CG产品图"`
+
+
+### Turn 70 · 17:20:03
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "护肤品 产品图片 美妆护肤"`
+
+
+### Turn 71 · 17:20:04
+- `Bash` — `cd "C:/Users/11304/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "倾斜悬浮构图 透明流体 光带 科技感"`
+
+
+### Turn 72 · 17:20:07
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "倾斜悬浮构图 透明流体 光带 科技感"`
+
+
+### Turn 73 · 17:20:09
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "光效 流体特效 粒子光效 动态纹理"`
+
+
+### Turn 74 · 17:20:12
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "悬浮构图 倾斜角度 特写镜头"`
+
+
+### Turn 75 · 17:20:13
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "AI提示词 文生图指令 图像生成指令"`
+
+
+### Turn 76 · 17:20:17
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "机器学习 AI应用 提示工程 AIGC"`
+
+
+### Turn 77 · 17:20:18
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --narrow "提示词 prompt 文生图指令"`
+
+
+### Turn 78 · 17:20:21
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --subtree "/理念/知识/商业/前沿技术/AI智能"`
+
+
+### Turn 79 · 17:20:22
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --subtree "/表象/视觉/符号/文字符号/标注"`
+
+
+### Turn 80 · 17:20:25
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "/理念/知识/商业/前沿技术/AI智能/机器学习"`
+
+
+### Turn 81 · 17:20:27
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"`
+
+
+### Turn 82 · 17:20:30
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --help 2>&1 | head -50`
+
+
+### Turn 83 · 17:20:35
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python -c "
+import subprocess, sys
+
+# Test that --match works normally
+result = subprocess....`
+
+
+### Turn 84 · 17:20:41
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "AI应用 AIGC 图像生成 创意工具"`
+
+
+### Turn 85 · 17:20:42
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "纯文本 文字描述 段落文本"`
+
+
+### Turn 86 · 17:20:45
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --list-l2`
+
+
+### Turn 87 · 17:20:47
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --list-l2`
+
+
+### Turn 88 · 17:20:51
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "美妆护肤 产品图片 商品展示"`
+
+
+### Turn 89 · 17:20:52
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "产品特写 悬浮展示 科技未来风 渐变背景"`
+
+
+### Turn 90 · 17:20:56
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "竖版海报 产品海报 版式设计"`
+
+
+### Turn 91 · 17:20:57
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "产品广告 商业宣传 奢侈品美妆"`
+
+
+### Turn 92 · 17:21:00
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --narrow "产品广告 商业宣传 奢侈品"`
+
+
+### Turn 93 · 17:21:01
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "科技未来 未来感 极简风格 奢华"`
+
+
+### Turn 94 · 17:21:06
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "AI生成合成 AIGC图像"`
+
+
+### Turn 95 · 17:21:07
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "CG渲染 三维建模 数字绘画"`
+
+
+### Turn 96 · 17:21:10
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "AI风格 数字合成 未来感视觉"`
+
+
+### Turn 97 · 17:21:11
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "产品悬浮展示 实物呈现"`
+
+
+### Turn 98 · 17:21:40
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/_scratch/patch_2b.json` (1,448 chars)
+
+
+### Turn 99 · 17:22:03
+
+
+### Turn 100 · 17:22:03
+> Both sub-agents completed. Now apply both patches:
+
+
+
+### Turn 101 · 17:22:04
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/wf-patch.py --workflow "C:/Users/11304/gitlab/cybertogether/Agent/example...`
+
+
+### Turn 102 · 17:22:08
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/wf-patch.py --workflow "C:/Users/11304/gitlab/cybertogether/Agent/example...`
+
+
+### Turn 103 · 17:22:13
+
+
+### Turn 104 · 17:22:14
+> Phase 2 complete. Now load Phase 3 spec files:
+
+
+
+### Turn 105 · 17:22:15
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/phase3-finalize.md`
+
+
+### Turn 106 · 17:22:15
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/format/case-data.schema.json`
+
+
+### Turn 107 · 17:23:32
+
+
+### Turn 108 · 17:23:32
+> Now update the intent sentences (referencing normalized types/actions) and run lint:
+
+
+
+### Turn 109 · 17:23:36
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/wf-patch.py --workflow "C:/Users/11304/gitlab/cybertogether/Agent/example...`
+
+
+### Turn 110 · 17:23:41
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/lint-case.py --workflow "C:/Users/11304/gitlab/cybertogether/Agent/exampl...`
+
+
+### Turn 111 · 17:23:51
+
+
+### Turn 112 · 17:23:51
+> Lint passes clean. Now render the HTML:
+
+
+
+### Turn 113 · 17:23:54
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/render-case.py --workflow "C:/Users/11304/gitlab/cybertogether/Agent/exam...`
+
+
+### Turn 114 · 17:24:25
+
+
+### Turn 115 · 17:24:25
+> Now write the .md DSL text version:
+
+
+
+### Turn 116 · 17:25:28
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/case-A_xhs_6968aa4d-ai-haibao-fengge-qianyi.md` (4,607 chars)
+
+
+### Turn 117 · 17:25:41
+
+
+### Turn 118 · 17:25:55
+> 所有阶段完成。以下是完整总结：
+>
+> ---
+>
+> ## 工序梗概
+>
+> 本 case 是一个 **3 步线性工序**——「AI 产品海报生成（参考图风格迁移）」：
+>
+> 1. **s1 反推提示词**：用多模态 AI 同时接收参考图（魔法世家蓝色精华水 CG 风格）和目标产品图（雅诗兰黛 Re-Nutriv 护肤水），一次指令完成"分析参考风格 + 提取产品外观 + 合成文生图提示词"
+> 2. **s2 Jimeng 配置**：粘贴提示词、上传产品图、设画面比例 2:3、参考强度 100%（关键参数——保持产品形态而非风格化）
+> 3. **s3 生成选图**：Jimeng 文生图多次抽卡，人工挑选符合"流体光带自然、产品细节清晰、清冷奢华调"的海报
+>
+> ---
+>
+> ## 输出文件路径
+>
+> ```
+> C:/Users/11304/gitlab/.../runs_full/q0018/procedures/A_xhs_6968aa4d/
+> ├── understanding.md                              ← Phase 1.1 心智模型
+> ├── workflow.json                                 ← 唯一中间产物（fully evolved）
+> ├── case-A_xhs_6968aa4d-ai-haibao-fengge-qianyi.html  ← Phase 3 HTML 可视化
+> ├── case-A_xhs_6968aa4d-ai-haibao-fengge-qianyi.md    ← Phase 3 DSL 文本版
+> └── _scratch/
+>     ├── anchors.json / task_2a.json / task_2b.json
+>     └── patch_2a.json (18 fields) / patch_2b.json (16 fields)
+> ```
+>
+> ---
+>
+> ## 对 DSL 的关键发现
+>
+> 1. **"双图合一反推"是新型提示词范式**：两张图（参考风格 + 目标产品）同时输入单次 LLM 调用，在一次 `提取/化学提取/反推` 动作内完成三个语义操作。这类多输入反推在产品海报工序中将频繁出现。
+>
+> 2. **`产品图` 与 `参考图` 语义分裂**：两者都是图像，但工序角色完全不同——前者是"被迁移风格的目标实体"（同时被分析 + 上传作 Jimeng reference），后者是"风格来源"。`产品图 extends 参考图` 的 type_registry 注册捕获了这一分裂。
+>
+> 3. **参考强度 100% = 形态保留，非风格化**：Jimeng 参数"参考强度"决定的不是风格强度而是产品外形的保真度，形成与提示词的双约束（风格约束 + 形态约束），当前 DSL 仅能通过 `config` 列隐含表达。
+>
+> 4. **"工具配置"作为 inferred 中间产物**：UI 类工具的配置状态必须显式建模为中间产物，才能保持 DSL anchor 链条闭合——这是 UI 驱动工序的通用建模挑战，inferred 标记是诚实记录这一补全的正确方式。
+
+
+
+### ◀ Result · 17:25:56
+
+- subtype: `success` · is_error: `False`
+- num_turns: `30` · duration: `1113800ms` · wall: `1116.1s`
+- tokens: in=24 out=54,330 cache_w=140,341 cache_r=2,202,785
+- cost: sdk=$2.7602, est_if_api=$2.0021

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/case-A_xhs_6968aa4d-ai-haibao-fengge-qianyi.md

@@ -0,0 +1,145 @@
+# Case A_xhs_6968aa4d: AI产品海报生成（参考图风格迁移）
+
+**Source**: https://www.xiaohongshu.com/explore/6968aa4d000000000a03329c
+**标题**: 如何让AI抄参考图？（附指令词）
+**平台**: 小红书 (xhs) · 2026-01-15
+**主题**: 用多模态 AI 反推参考图视觉风格提示词，替换产品信息，在 Jimeng 中生成高端产品海报
+
+---
+
+## 工序梗概 (人话)
+
+三步实现 AI 产品海报风格迁移：
+
+1. **反推提示词**：拿到具有目标风格的参考图（本例：魔法世家蓝色精华水 CG 悬浮构图）和目标产品图（雅诗兰黛 Re-Nutriv 护肤水），用多模态 AI 一次性完成"分析参考图风格 + 提取目标产品外观 + 合成文生图提示词"。核心指令：让 AI 分析图1的主体/场景/风格/色调/构图/细节，描述图2的产品外观与材质，再将图2产品信息替换入图1结构，生成提示词。
+2. **Jimeng 配置**：打开即梦 AI（Jimeng），粘贴反推提示词，上传目标产品图，设画面比例 2:3（竖版海报），将产品图参考强度设为 100% 并保存。参考强度 100% 是保持产品形态不失真的关键。
+3. **生成并选图**：点击生成，多次抽卡，人工选取最符合"流体光带动态自然、产品细节清晰、清冷奢华色调"的版本，得到高端护肤品科技风产品海报。
+
+---
+
+## 引用的类型 (stdlib + 自定义)
+
+### stdlib 类型 (直接引用)
+
+| 类型名 | 在工序中的用途 |
+|---|---|
+| `参考图` | 具有目标视觉风格的产品广告参考图 (s1.inputs[0], declarations.inputs[0]) |
+| `提示词` | 多模态 AI 反推生成的文生图提示词文本 (s1.outputs[0], s2.inputs[0]) |
+| `成品图` | Jimeng 文生图生成的最终产品海报 (s3.outputs[0]) |
+
+### case-specific 类型 (type_registry)
+
+| 类型名 | extends | desc |
+|---|---|---|
+| `产品图` | `参考图` | 本次工序中待迁移风格的目标产品图（雅诗兰黛 Re-Nutriv 护肤水官方图），被 AI 分析以提取目标产品的外观特征，并在生图时作为图像参考上传到 Jimeng；角色是"被模仿/被迁移风格的目标实体"，而非单纯的风格范本 |
+| `工具配置` | `生成参数` | Jimeng 文生图工具的综合生图配置集合（已粘贴提示词 + 已上传产品参考图 + 画面比例 2:3 + 参考强度 100% + 图片4.0 + 高清2K），超出单一生成参数范畴故作 case-specific 注册 |
+
+---
+
+## L1 外部函数库
+
+| 工具 | 描述 | 在工序中的角色 |
+|---|---|---|
+| `(多模态AI)` | 接受多图输入 + 文本指令的多模态 LLM（原文未指名具体工具，可为 Claude、GPT-4V、Kimi 等） | s1 · 反推参考图风格提示词 |
+| `jimeng` | 即梦 AI（Bytedance 出品），支持文生图、文本+图片联合输入、画面比例控制、参考强度控制 | s2 · 配置生图参数；s3 · 文生图生成 |
+
+---
+
+## L2 抽象动作 + impl 关系
+
+| 动作路径 | impl via | 步骤 |
+|---|---|---|
+| `提取/化学提取/反推` | `(多模态AI)` | s1 · 从参考图+产品图反推文生图提示词 |
+| `获取/录入/键入` | `jimeng` | s2 · 向 Jimeng 键入提示词并配置参数 |
+| `生成/元素生成` | `jimeng` | s3 · 文生图生成产品海报 |
+
+---
+
+## L3 工序模板
+
+**工序名**: AI产品海报生成（参考图风格迁移）
+**类别**: 产物创造
+**purpose**: 用AI分析参考图视觉风格并替换为目标产品信息，在Jimeng中生成高端产品海报
+
+```
+procedure AI产品海报生成（参考图风格迁移）
+  inputs:
+    参考图: 参考图         -- 具有目标视觉风格的产品广告参考图
+    目标产品图: 产品图     -- 需要生成海报的目标产品实拍或官方图
+  returns: 成品图
+
+  s1 [预处理] via (多模态AI), action 提取/化学提取/反推, feature 随机
+    directive: "分析并描述图1的主体内容、场景设定、风格参考、色彩色调、构图视角、细节等信息。
+               描述图2的产品外观、材质结构。并将图2的产品信息添加到图1，然后生成一段用于文生图的提示词。"
+    inputs:  参考图 ← 工序输入
+             目标产品图 ← 工序输入
+    outputs: 反推提示词 (提示词) → s2
+
+  s2 [预处理] via jimeng, action 获取/录入/键入, feature 幂等
+    directive: "将提示词添加到jimeng并上传产品图，调整画面比例为2:3，点击产品图，将参考强度设为100%，点击保存"
+    config: { 比例: 2:3, 图片4.0, 高清2K, 参考强度: 100% }
+    inputs:  反推提示词 (提示词) ← s1.反推提示词
+             目标产品图 (产品图) ← 工序输入
+    outputs: 生图配置 (工具配置) → s3  [inferred]
+
+  s3 [主体生成] via jimeng, action 生成/元素生成, feature 随机
+    decorator: @采样(pick=人工)
+    directive: "点击'生成'，多次抽卡后选择符合以下特征的海报：流体光带的动态自然不生硬，
+               产品细节（瓶身通透感、文字标识）清晰，色调贴合'清冷奢华'的科技感"
+    inputs:  生图配置 (工具配置) ← s2.生图配置  [inferred]
+    outputs: 产品海报 (成品图) → 返回 产品海报
+```
+
+---
+
+## L4 工序实例
+
+### inputs (本次实际值)
+
+| 变量名 | 类型 | 实际值 |
+|---|---|---|
+| 参考图 | 参考图 | `<魔法世家蓝色精华水产品CG风格图：倾斜悬浮特写构图，被透明流体光带环绕，背景浅灰渐变，科技感美妆氛围，银色金属瓶盖>` |
+| 目标产品图 | 产品图 | `<雅诗兰黛Re-Nutriv护肤水官方图：瓶身为透明金色玻璃材质（内含浅金色液体），搭配金色金属质感瓶盖；瓶身印有'Re-Nutriv''ESTÉE LAUDER'字样及产品说明，风格雍容高贵>` |
+
+### bindings (本次工具选择)
+
+| 步骤 | via | 备注 |
+|---|---|---|
+| s1 | `(多模态AI)` | 原文未指名具体工具（inferred），可为任意支持多图分析的多模态 LLM |
+| s2 | `jimeng` | 即梦 AI，Bytedance 出品 |
+| s3 | `jimeng` | 同上 |
+
+### extracted_values (本次中间产物 — 逐字回填)
+
+**s1.outputs[0] — 反推提示词** (type: 提示词)
+
+> 未来感极简美妆 CG 风格，倾斜悬浮特写：雅诗兰黛 Re-Nutriv 护肤水（通透金色玻璃瓶身 + 浅金色液体，配金色金属瓶盖，印'Re-Nutriv''ESTÉE LAUDER'字样）被透明流体光带环绕托举；背景浅灰渐变，色调清冷奢华，突出瓶身通透感、瓶盖金属质感与流体光带动态纹理，营造科技感高端护肤氛围
+
+**s2.outputs[0] — 生图配置** (type: 工具配置，inferred)
+
+> `<Jimeng生图配置：反推提示词已粘贴，目标产品图已上传，画面比例2:3，参考强度100%，图片4.0模型，高清2K>`
+
+**s3.outputs[0] — 产品海报** (type: 成品图)
+
+> `<雅诗兰黛Re-Nutriv护肤水高端产品海报：倾斜悬浮CG风格，透明流体光带环绕瓶身，背景浅灰渐变底色，清冷金色调（金色+透明+浅灰），高清2K，竖版（2:3比例）>`
+
+**选图标准** (来自 s3 directive):
+- 流体光带的动态自然不生硬
+- 产品细节（瓶身通透感、文字标识）清晰
+- 色调贴合"清冷奢华"的科技感
+
+### trace
+
+(暂无 timing / cost / retry 数据记录)
+
+---
+
+## 这个 case 对 DSL 设计的关键启发
+
+1. **"双图合一反推"是新型提示词工程范式**：传统反推是单图 → 提示词；本 case 是双图（参考风格图 + 目标产品图）→ 提示词。AI 在单次调用内完成三个语义操作：分析参考图风格、提取目标产品特征、合并写出提示词。这是一个典型的 `提取/化学提取/反推` 动作携带多输入的例子，两个输入的 substance 不同（都是美妆护肤，但一是风格参考、一是目标实体）。
+
+2. **`产品图` vs `参考图` 的语义区分值得收入 type_registry**：两者在工序中的角色完全不同——`参考图` 是风格来源（被分析但不作为生图 reference），`产品图` 是被迁移风格的对象（同时被 AI 分析 + 上传到 Jimeng 作图参考）。这种"同是图，角色不同"的情况在产品海报类工序中将频繁出现，`产品图 extends 参考图` 的注册方式是合理的。
+
+3. **Jimeng 的"参考强度"参数分裂了 action 的语义**：100% 参考强度不是"风格化"而是"形态保留"——它让生图工具优先保持产品图的外形结构，而非自由发挥。这意味着 s3 的 `生成/元素生成` action 实际上有两个并行约束：提示词驱动的风格约束 + 产品图参考强度约束。当前 DSL 没有表达这层"双约束"的字段，只能通过 config `参考强度: 100%` 隐含传达。
+
+4. **"工具配置"中间产物的合法性**：s2 → s3 之间传递的"生图配置"是一个 UI 状态量，不是真正的数据文件。这是工序 DSL 在建模 UI 类工具时的常见困境——工具的配置状态必须作为中间产物显式化（inferred 标记），才能让 DSL 的 anchor 链条闭合。如果不引入这个中间产物，s2 就成了"没有 output 的步骤"，违反 DSL 规范。

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/understanding.md

@@ -0,0 +1,60 @@
+# Case A_xhs_6968aa4d · 心智模型
+
+**Case 原文**: 如何让AI抄参考图？（附指令词）
+**Source**: https://www.xiaohongshu.com/explore/6968aa4d000000000a03329c
+**Channel**: xhs (小红书)
+**发布时间**: 2026-01-15
+
+---
+
+## 工序数量判断
+
+**本 case 含 1 个独立工序**。
+
+**判断依据**:
+- **唯一终态产物**: 高端产品海报（以雅诗兰黛 Re-Nutriv 护肤水为例）
+- **完整输入→输出链条**: 参考图（魔法世家精华水风格图）+ 目标产品图（Re-Nutriv 护肤水）→ 产品海报
+- **可操作做法**: 三步工艺（反推提示词 → Jimeng 配置 → 生成选图）
+- 无并行子流程，无多终态产物分支，无独立输入→输出链条的变体工序
+
+---
+
+## 工序一：AI产品海报生成（参考图风格迁移）
+
+**目的**: 用多模态 AI 反推参考图的视觉风格提示词，替换为目标产品信息，在 Jimeng 中生成高端产品海报
+
+**工艺类型**: 图像生成 / 风格迁移（Style Transfer via Prompt Reverse Engineering）
+
+**终态产物**: 产品海报（高端护肤品 CG 风格竖版图，2:3 比例，高清 2K）
+
+**关键外部工具依赖**:
+- **多模态 AI 工具**（原文只说"AI 软件"，未指名具体工具；需支持多图输入+文本分析）
+- **Jimeng（即梦 AI）**（Bytedance 出品图像生成工具；用于文生图+参考强度控制）
+
+**步骤数**: 3 步（s1 → s2 → s3，线性链条，无分支/循环/并行）
+
+| # | 步骤名 | 核心动作 | 关键 IO |
+|---|--------|----------|---------|
+| s1 | 反推参考图提示词 | 多模态 AI 同时分析参考图（风格）和产品图（形态），合成含产品替换的文生图提示词 | IN: 参考图 + 目标产品图 → OUT: 反推提示词 |
+| s2 | Jimeng 配置生图参数 | 粘贴提示词、上传产品图、设比例 2:3、参考强度 100%、保存 | IN: 反推提示词 + 目标产品图 → OUT: 生图配置（推断中间产物） |
+| s3 | 生成并人工选图 | Jimeng 文生图 + 多次抽卡人工筛选 | IN: 生图配置 → OUT: 产品海报 |
+
+---
+
+## 隐含工艺规约与推断补全说明
+
+**核心技法**: "分析图1风格 + 替换图2产品信息" 是此类参考图风格迁移的标准范式。AI 需同时理解参考图的构图/色调/风格（保留）和目标产品的形态特征（替换入）。
+
+**Jimeng 参考强度 = 100%**: 确保目标产品图的外形、主体结构被严格保留，生成图不会改变产品的基本形态（非风格化），是保证产品辨识度的关键参数。
+
+**推断补全说明**:
+1. **s2 输出"生图配置"是推断中间产物** — 原文仅描述配置操作（粘贴提示词、上传图、调比例等），未明示存在独立中间产物；但 s3 的生成动作依赖 s2 的配置结果，DSL 需明确此数据流，故补出该中间状态
+2. **s3 输入"生图配置"是推断** — 同上，原文未将其列为 s3 的独立输入
+3. **s1 的 via（多模态 AI 工具）是推断** — 原文只说"AI 软件"，未指名具体工具（可能是 Claude、GPT-4V、Kimi 等任意支持多图分析的 LLM）
+
+**多次抽卡质量控制**: 原文"多次抽卡直到满意"是随机生成工具的标准质量控制手段，映射为 `@采样(pick=人工)` 装饰器
+
+**选图标准（来自源文）**:
+- 流体光带的动态自然不生硬
+- 产品细节（瓶身通透感、文字标识）清晰
+- 色调贴合"清冷奢华"的科技感

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6968aa4d/workflow.json

@@ -0,0 +1,219 @@
+{
+  "source": {
+    "platform": "xhs",
+    "author": "xhs博主",
+    "date": "2026-01-15",
+    "url": "https://www.xiaohongshu.com/explore/6968aa4d000000000a03329c",
+    "title": "如何让AI抄参考图？（附指令词）",
+    "excerpt": "三步实现AI产品海报风格迁移：用多模态AI反推参考图视觉风格提示词、替换目标产品信息，再在Jimeng中生成高端产品海报。"
+  },
+  "procedures": [
+    {
+      "id": "p1",
+      "name": "AI产品海报生成（参考图风格迁移）",
+      "purpose": "用AI分析参考图视觉风格并替换为目标产品信息，在Jimeng中生成高端产品海报",
+      "category": "产物创造",
+      "platform": "xhs",
+      "author": "xhs博主",
+      "declarations": {
+        "inputs": [
+          {
+            "type": "参考图",
+            "name": "参考图",
+            "desc": "具有目标视觉风格的产品广告参考图片（本例为魔法世家蓝色精华水CG风格图）"
+          },
+          {
+            "type": "产品图",
+            "name": "目标产品图",
+            "desc": "需要生成海报的目标产品实拍或官方图（本例为雅诗兰黛Re-Nutriv护肤水）"
+          }
+        ],
+        "resources": [],
+        "returns": {
+          "type": "图像"
+        }
+      },
+      "steps": [
+        {
+          "id": "s1",
+          "kind": "step",
+          "effect": "预处理",
+          "via": "(多模态AI)",
+          "action": "提取/化学提取/反推",
+          "feature": "随机",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "分析并描述图1的主体内容、场景设定、风格参考、色彩色调、构图视角、细节等信息。描述图2的产品外观、材质结构。并将图2的产品信息添加到图1，然后生成一段用于文生图的提示词。"
+            ]
+          ],
+          "inputs": [
+            {
+              "substance": "/表象/视觉/实体/物品/起居物品/美妆护肤",
+              "form": [
+                "/呈现/视觉/视觉风格/视觉艺术风格/现代气质/科技未来",
+                "/呈现/视觉/视觉制作/后期处理/图像合成处理/三维建模合成"
+              ],
+              "type": "参考图",
+              "name": "参考图",
+              "value": "<魔法世家蓝色精华水产品CG风格图：倾斜悬浮特写构图，被透明流体光带环绕，背景浅灰渐变，科技感美妆氛围，银色金属瓶盖>",
+              "anchor": "← 工序输入"
+            },
+            {
+              "substance": "/表象/视觉/实体/物品/起居物品/美妆护肤",
+              "form": "/呈现/视觉/视觉制作/实景拍摄/拍摄主体/实物呈现",
+              "type": "产品图",
+              "name": "目标产品图",
+              "value": "<雅诗兰黛Re-Nutriv护肤水官方图：瓶身为透明金色玻璃材质（内含浅金色液体），搭配金色金属质感瓶盖；瓶身印有'Re-Nutriv''ESTÉE LAUDER'字样及产品说明，风格雍容高贵>",
+              "anchor": "← 工序输入"
+            }
+          ],
+          "outputs": [
+            {
+              "substance": "/理念/知识/商业/前沿技术/AI智能/AI应用",
+              "form": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构",
+              "type": "提示词",
+              "name": "反推提示词",
+              "value": "未来感极简美妆 CG 风格，倾斜悬浮特写：雅诗兰黛 Re-Nutriv 护肤水（通透金色玻璃瓶身 + 浅金色液体，配金色金属瓶盖，印'Re-Nutriv''ESTÉE LAUDER'字样）被透明流体光带环绕托举；背景浅灰渐变，色调清冷奢华，突出瓶身通透感、瓶盖金属质感与流体光带动态纹理，营造科技感高端护肤氛围",
+              "anchor": "→ s2"
+            }
+          ],
+          "intent": "用 {via:(多模态AI)} {act:反推} {in-type:参考图}+{in-type:产品图} 得 {out-type:提示词}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "in-type-1",
+            "out-type-0"
+          ],
+          "inferred_marks": {
+            "via": "原文只说'使用AI软件输入指令反推提示词'，未指名具体多模态AI工具"
+          }
+        },
+        {
+          "id": "s2",
+          "kind": "step",
+          "effect": "预处理",
+          "via": "jimeng",
+          "action": "获取/录入/键入",
+          "feature": "幂等",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "将提示词添加到jimeng并上传产品图，调整画面比例为2:3（适配竖版海报），点击产品图，将参考强度设为100%，点击保存"
+            ],
+            [
+              "config",
+              "{ 比例: 2:3, 图片4.0, 高清2K, 参考强度: 100% }"
+            ]
+          ],
+          "inputs": [
+            {
+              "substance": "/理念/知识/商业/前沿技术/AI智能/AI应用",
+              "form": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构",
+              "type": "提示词",
+              "name": "反推提示词",
+              "value": "未来感极简美妆 CG 风格，倾斜悬浮特写：雅诗兰黛 Re-Nutriv 护肤水（通透金色玻璃瓶身 + 浅金色液体，配金色金属瓶盖，印'Re-Nutriv''ESTÉE LAUDER'字样）被透明流体光带环绕托举；背景浅灰渐变，色调清冷奢华，突出瓶身通透感、瓶盖金属质感与流体光带动态纹理，营造科技感高端护肤氛围",
+              "anchor": "← s1.反推提示词"
+            },
+            {
+              "substance": "/表象/视觉/实体/物品/起居物品/美妆护肤",
+              "form": "/呈现/视觉/视觉制作/实景拍摄/拍摄主体/实物呈现",
+              "type": "产品图",
+              "name": "目标产品图",
+              "value": "<雅诗兰黛Re-Nutriv护肤水官方图：瓶身为透明金色玻璃材质（内含浅金色液体），搭配金色金属质感瓶盖；瓶身印有'Re-Nutriv''ESTÉE LAUDER'字样及产品说明，风格雍容高贵>",
+              "anchor": "← 工序输入"
+            }
+          ],
+          "outputs": [
+            {
+              "substance": null,
+              "form": null,
+              "type": "工具配置",
+              "name": "生图配置",
+              "value": "<Jimeng生图配置：反推提示词已粘贴，目标产品图已上传，画面比例2:3，参考强度100%，图片4.0模型，高清2K>",
+              "anchor": "→ s3",
+              "inferred": true,
+              "inferred_reason": "原文描述了s2的配置操作但未明示存在'生图配置'这一中间产物；工艺上s3（生成）需要s2的配置结果作为前置状态，故补出此中间产物以显式化数据流"
+            }
+          ],
+          "intent": "{via:jimeng} {act:键入} {in-type:提示词}+{in-type:产品图}，产出 {out-type:工具配置}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "in-type-1",
+            "out-type-0"
+          ]
+        },
+        {
+          "id": "s3",
+          "kind": "step",
+          "effect": "主体生成",
+          "via": "jimeng",
+          "action": "生成/元素生成",
+          "feature": "随机",
+          "control": "-",
+          "instruction": [
+            [
+              "decorator",
+              "@采样(pick=人工)"
+            ],
+            [
+              "directive",
+              "点击'生成'，多次抽卡后选择符合以下特征的海报：流体光带的动态自然不生硬，产品细节（瓶身通透感、文字标识）清晰，色调贴合'清冷奢华'的科技感"
+            ]
+          ],
+          "inputs": [
+            {
+              "substance": null,
+              "form": null,
+              "type": "工具配置",
+              "name": "生图配置",
+              "value": "<Jimeng生图配置：反推提示词已粘贴，目标产品图已上传，画面比例2:3，参考强度100%，图片4.0模型，高清2K>",
+              "anchor": "← s2.生图配置",
+              "inferred": true,
+              "inferred_reason": "s2输出的生图配置是s3生成动作的前置输入，原文未将其明确列为s3的独立输入参数"
+            }
+          ],
+          "outputs": [
+            {
+              "substance": "/表象/视觉/实体/物品/起居物品/美妆护肤",
+              "form": [
+                "/呈现/视觉/视觉制作/后期处理/图像合成处理/AI生成合成",
+                "/呈现/视觉/视觉风格/视觉艺术风格/现代气质/科技未来",
+                "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"
+              ],
+              "type": "成品图",
+              "name": "产品海报",
+              "value": "<雅诗兰黛Re-Nutriv护肤水高端产品海报：倾斜悬浮CG风格，透明流体光带环绕瓶身，背景浅灰渐变底色，清冷金色调（金色+透明+浅灰），高清2K，竖版（2:3比例）>",
+              "anchor": "→ 返回 产品海报"
+            }
+          ],
+          "intent": "{via:jimeng} {act:元素生成} 由 {in-type:工具配置} 生成 {out-type:成品图}",
+          "focus": [
+            "via",
+            "action",
+            "out-type-0"
+          ]
+        }
+      ],
+      "return_row": {
+        "arrow": "↩",
+        "text": "<span class=\"kw\">返回</span> <span class=\"chip\" data-type=\"图像\">图像</span>"
+      },
+      "type_registry": {
+        "产品图": {
+          "extends": "参考图",
+          "desc": "本次工序中待迁移风格的目标产品图（雅诗兰黛 Re-Nutriv 护肤水官方图），被 AI 分析以提取目标产品的外观特征，并在生图时作为图像参考上传到 Jimeng；其角色是'被模仿/被迁移风格的目标实体'，而非单纯的风格范本，因此以 case-specific 类型独立标注。"
+        },
+        "工具配置": {
+          "extends": "生成参数",
+          "desc": "Jimeng 文生图工具的综合生图配置集合，包含：已粘贴的反推提示词、已上传的目标产品参考图、画面比例（2:3）、参考强度（100%）、使用模型（图片 4.0）、输出规格（高清 2K）；属于本次生图任务的完整参数束，超出单一生成参数范畴故作 case-specific 注册。"
+        }
+      }
+    }
+  ]
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/.session_id

@@ -0,0 +1 @@
+6044e329-0183-4696-abde-c32738c3fae4

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/_meta.json

@@ -0,0 +1,8 @@
+{
+  "case_id": "xhs_6a16c1d0000000003601a30f",
+  "from_q": "q0018",
+  "form": "A",
+  "score": 9.0,
+  "score_norm": 0.9,
+  "started_at": "2026-06-03T16:18:38"
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/_scratch/anchors.json

@@ -0,0 +1,18 @@
+[
+  {"path": "p1.s1.inputs[0].anchor",  "value": "← 字面量"},
+  {"path": "p1.s1.outputs[0].anchor", "value": "→ s3.图片生成配置"},
+  {"path": "p1.s2.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p1.s2.outputs[0].anchor", "value": "→ s3.儿童节主题提示词"},
+  {"path": "p1.s3.inputs[0].anchor",  "value": "← s2.儿童节主题提示词"},
+  {"path": "p1.s3.inputs[1].anchor",  "value": "← s1.图片生成配置"},
+  {"path": "p1.s3.outputs[0].anchor", "value": "→ 返回 图像"},
+  {"path": "p2.s1.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p2.s1.inputs[1].anchor",  "value": "← 字面量"},
+  {"path": "p2.s1.outputs[0].anchor", "value": "→ 返回 图像"},
+  {"path": "p3.s1.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p3.s1.outputs[0].anchor", "value": "→ s2.视频生成配置"},
+  {"path": "p3.s2.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p3.s2.inputs[1].anchor",  "value": "← 字面量"},
+  {"path": "p3.s2.inputs[2].anchor",  "value": "← s1.视频生成配置"},
+  {"path": "p3.s2.outputs[0].anchor", "value": "→ 返回 视频"}
+]

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/_scratch/patch_2a.json

@@ -0,0 +1,43 @@
+[
+  {"path": "p1.s1.effect", "value": "预处理"},
+  {"path": "p1.s1.action", "value": "获取/录入/键入"},
+  {"path": "p1.s1.inputs[0].type", "value": "平台工具"},
+  {"path": "p1.s1.outputs[0].type", "value": "工具配置"},
+
+  {"path": "p1.s2.effect", "value": "预处理"},
+  {"path": "p1.s2.action", "value": "生成/元素生成"},
+  {"path": "p1.s2.inputs[0].type", "value": "描述"},
+  {"path": "p1.s2.outputs[0].type", "value": "提示词"},
+
+  {"path": "p1.s3.effect", "value": "主体生成"},
+  {"path": "p1.s3.action", "value": "生成/元素生成"},
+  {"path": "p1.s3.inputs[0].type", "value": "提示词"},
+  {"path": "p1.s3.inputs[1].type", "value": "工具配置"},
+  {"path": "p1.s3.outputs[0].type", "value": "底图"},
+
+  {"path": "p2.s1.effect", "value": "主体生成"},
+  {"path": "p2.s1.action", "value": "修改/变/风格化"},
+  {"path": "p2.s1.inputs[0].type", "value": "底图"},
+  {"path": "p2.s1.inputs[1].type", "value": "提示词"},
+  {"path": "p2.s1.outputs[0].type", "value": "底图"},
+
+  {"path": "p3.s1.effect", "value": "预处理"},
+  {"path": "p3.s1.action", "value": "获取/录入/键入"},
+  {"path": "p3.s1.inputs[0].type", "value": "参考图"},
+  {"path": "p3.s1.outputs[0].type", "value": "工具配置"},
+
+  {"path": "p3.s2.effect", "value": "主体生成"},
+  {"path": "p3.s2.action", "value": "生成/元素生成"},
+  {"path": "p3.s2.inputs[0].type", "value": "参考图"},
+  {"path": "p3.s2.inputs[1].type", "value": "提示词"},
+  {"path": "p3.s2.inputs[2].type", "value": "工具配置"},
+  {"path": "p3.s2.outputs[0].type", "value": "视频成品"},
+
+  {"path": "p1.type_registry.平台工具.extends", "value": "生成参数"},
+  {"path": "p1.type_registry.平台工具.desc", "value": "操作入口所在的平台及工具信息，说明使用哪个 Web 平台访问哪个 AI 工具"},
+  {"path": "p1.type_registry.工具配置.extends", "value": "生成参数"},
+  {"path": "p1.type_registry.工具配置.desc", "value": "工具与模型的综合配置项，含平台模型选型、风格模型选择及规格参数等"},
+
+  {"path": "p3.type_registry.工具配置.extends", "value": "生成参数"},
+  {"path": "p3.type_registry.工具配置.desc", "value": "工具与模型的综合配置项，含平台模型选型、参考模式、画幅规格、时长及配音等"}
+]

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/_scratch/patch_2b.json

@@ -0,0 +1,34 @@
+[
+  {"path": "p1.s1.inputs[0].substance", "value": "__null__"},
+  {"path": "p1.s1.inputs[0].form", "value": "__null__"},
+  {"path": "p1.s1.outputs[0].substance", "value": "__null__"},
+  {"path": "p1.s1.outputs[0].form", "value": "__null__"},
+  {"path": "p1.s2.inputs[0].substance", "value": ["/表象/视觉/时间/节日/现代节日", "/理念/知识/文化/节气节日/节日"]},
+  {"path": "p1.s2.inputs[0].form", "value": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"},
+  {"path": "p1.s2.outputs[0].substance", "value": "/表象/视觉/符号/文字符号/标注/提示引导"},
+  {"path": "p1.s2.outputs[0].form", "value": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"},
+  {"path": "p1.s3.inputs[0].substance", "value": "/表象/视觉/符号/文字符号/标注/提示引导"},
+  {"path": "p1.s3.inputs[0].form", "value": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"},
+  {"path": "p1.s3.inputs[1].substance", "value": "__null__"},
+  {"path": "p1.s3.inputs[1].form", "value": "__null__"},
+  {"path": "p1.s3.outputs[0].substance", "value": "/表象/视觉/实体/物品/影像物品/艺术创作"},
+  {"path": "p1.s3.outputs[0].form", "value": "/呈现/视觉/视觉风格/视觉艺术风格/绘画风格/插画 + /呈现/视觉/视觉制作/构图编排/版面设计/版面结构"},
+  {"path": "p2.s1.inputs[0].substance", "value": "/表象/视觉/实体/物品/影像物品/艺术创作"},
+  {"path": "p2.s1.inputs[0].form", "value": "/呈现/视觉/视觉风格/视觉艺术风格/绘画风格/插画 + /呈现/视觉/视觉制作/构图编排/版面设计/版面结构"},
+  {"path": "p2.s1.inputs[1].substance", "value": "/表象/视觉/符号/文字符号/标注/提示引导"},
+  {"path": "p2.s1.inputs[1].form", "value": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"},
+  {"path": "p2.s1.outputs[0].substance", "value": "/表象/视觉/实体/物品/影像物品/艺术创作"},
+  {"path": "p2.s1.outputs[0].form", "value": "/呈现/视觉/视觉制作/后期处理/图像合成处理/三维建模合成 + /呈现/视觉/视觉制作/构图编排/版面设计/版面结构"},
+  {"path": "p3.s1.inputs[0].substance", "value": "/表象/视觉/实体/物品/影像物品/艺术创作"},
+  {"path": "p3.s1.inputs[0].form", "value": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"},
+  {"path": "p3.s1.outputs[0].substance", "value": "__null__"},
+  {"path": "p3.s1.outputs[0].form", "value": "__null__"},
+  {"path": "p3.s2.inputs[0].substance", "value": "/表象/视觉/实体/物品/影像物品/艺术创作"},
+  {"path": "p3.s2.inputs[0].form", "value": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"},
+  {"path": "p3.s2.inputs[1].substance", "value": "/表象/视觉/符号/文字符号/标注/提示引导"},
+  {"path": "p3.s2.inputs[1].form", "value": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"},
+  {"path": "p3.s2.inputs[2].substance", "value": "__null__"},
+  {"path": "p3.s2.inputs[2].form", "value": "__null__"},
+  {"path": "p3.s2.outputs[0].substance", "value": "/表象/视觉/实体/物品/影像物品/艺术创作"},
+  {"path": "p3.s2.outputs[0].form", "value": "/呈现/视觉/视觉制作/后期处理/动态特效处理/运动变换 + /呈现/视觉/视觉制作/后期处理/图像合成处理/AI生成合成"}
+]

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/_scratch/task_2a.json

@@ -0,0 +1,191 @@
+{
+  "case_id": "A_xhs_6a16c1d0",
+  "task": "Phase 2A (effect/action/type normalization)",
+  "image_url_list": [],
+  "steps": [
+    {
+      "path": "p1.s1",
+      "name": "",
+      "action": "选择工具与风格配置",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p1.s1.inputs[0]",
+          "name": "LibLib平台",
+          "value": "liblib.art web端（需浏览器访问，免登录可用部分额度）",
+          "type": "平台工具",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p1.s1.outputs[0]",
+          "name": "图片生成配置",
+          "value": "LibLib AI 智能图片V2 · 风格模型：小红书风海报手绘涂鸦插画",
+          "type": "工具配置",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p1.s2",
+      "name": "",
+      "action": "撰写主题提示词",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p1.s2.inputs[0]",
+          "name": "海报主题构思",
+          "value": "儿童节主题选定：复古橘子汽水瓶 / 街机游戏机 / 泡泡糖扭蛋机 / 零食包装袋墙（小卖部）等童年元素",
+          "type": "提示词",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p1.s2.outputs[0]",
+          "name": "儿童节主题提示词",
+          "value": "巨大复古橘子汽水玻璃瓶作为画面中心框架，瓶内是夏日童年场景（蓝天、白云、彩虹、气泡、小卖部、冰棒、西瓜），屏幕中跳出卡通小孩角色：穿彩色背带裤、手拿橘子汽水、戴小熊帽子，表情开心兴奋；另一位小朋友蹲在旁边玩泡泡机；道具元素：玻璃弹珠、吸管、橘子切片、冰块、爆米花、贴纸、气泡；纸张质感，高饱和撞色：红、蓝、黄、绿、黑色粗线条勾边，漫画感强烈，手绘涂鸦插画，复古童年元素与现代卡通结合，美式漫画爆炸贴效果，活泼动感，年轻潮流",
+          "type": "提示词",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p1.s3",
+      "name": "",
+      "action": "文生图生成插画海报",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p1.s3.inputs[0]",
+          "name": "儿童节主题提示词",
+          "value": "巨大复古橘子汽水玻璃瓶作为画面中心框架，瓶内是夏日童年场景（蓝天、白云、彩虹、气泡、小卖部、冰棒、西瓜），屏幕中跳出卡通小孩角色：穿彩色背带裤、手拿橘子汽水、戴小熊帽子，表情开心兴奋；另一位小朋友蹲在旁边玩泡泡机；道具元素：玻璃弹珠、吸管、橘子切片、冰块、爆米花、贴纸、气泡；纸张质感，高饱和撞色：红、蓝、黄、绿、黑色粗线条勾边，漫画感强烈，手绘涂鸦插画，复古童年元素与现代卡通结合，美式漫画爆炸贴效果，活泼动感，年轻潮流",
+          "type": "提示词",
+          "related_images": []
+        },
+        {
+          "path": "p1.s3.inputs[1]",
+          "name": "图片生成配置",
+          "value": "LibLib AI 智能图片V2 · 风格模型：小红书风海报手绘涂鸦插画",
+          "type": "工具配置",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p1.s3.outputs[0]",
+          "name": "2D涂鸦风插画海报",
+          "value": "<生成的2D平面涂鸦风儿童节主题插画海报，3:4竖版，高饱和漫画感，含主体容器+童年场景+卡通角色+手绘涂鸦风格+爆炸贴效果>",
+          "type": "图像",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p2.s1",
+      "name": "",
+      "action": "图生图3D立体效果转换",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p2.s1.inputs[0]",
+          "name": "2D涂鸦风插画海报",
+          "value": "<P1生成的2D平面涂鸦风儿童节主题插画海报（如橘味汽水主题）>",
+          "type": "图像",
+          "related_images": []
+        },
+        {
+          "path": "p2.s1.inputs[1]",
+          "name": "3D效果提示词",
+          "value": "3d效果",
+          "type": "提示词",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p2.s1.outputs[0]",
+          "name": "3D立体效果海报",
+          "value": "<3D立体效果版儿童节主题海报，在2D涂鸦插画基础上增加立体质感和3D渲染效果，人物和元素呈立体膨胀感>",
+          "type": "图像",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p3.s1",
+      "name": "",
+      "action": "选择参考图并配置视频生成参数",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p3.s1.inputs[0]",
+          "name": "静态海报图",
+          "value": "<选定的静态涂鸦风插画海报（P1或P2产出）>",
+          "type": "图像",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p3.s1.outputs[0]",
+          "name": "视频生成配置",
+          "value": "Seedance 2.0 VIP · 全能参考 · 3:4 · 1080p · 6s · 有配音",
+          "type": "工具配置",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p3.s2",
+      "name": "",
+      "action": "图生视频生成动态海报",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p3.s2.inputs[0]",
+          "name": "静态海报图",
+          "value": "<静态涂鸦风插画海报（橘味汽水主题，P1产出）>",
+          "type": "图像",
+          "related_images": []
+        },
+        {
+          "path": "p3.s2.inputs[1]",
+          "name": "动态效果提示词",
+          "value": "瓶内夏日童年场景（蓝天、白云、彩虹、小卖部、冰棒、西瓜），卡通小孩从瓶中跳出挥舞橘子汽水，喷出大量透明泡泡；气泡持续上升，冰块碰撞晃动，橘子切片旋转，爆米花弹跳，贴纸与漫画元素漂浮；镜头轻微拉远镜，整体具有循环动画感",
+          "type": "提示词",
+          "related_images": []
+        },
+        {
+          "path": "p3.s2.inputs[2]",
+          "name": "视频生成配置",
+          "value": "Seedance 2.0 VIP · 全能参考 · 3:4 · 1080p · 6s · 有配音",
+          "type": "工具配置",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p3.s2.outputs[0]",
+          "name": "动态视频海报",
+          "value": "<6秒循环动态视频海报，3:4竖版1080p，有配音，涂鸦风元素动态运动（气泡上升、冰块晃动、爆米花弹跳），适合节日氛围传播>",
+          "type": "视频",
+          "related_images": []
+        }
+      ]
+    }
+  ]
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/_scratch/task_2b.json

@@ -0,0 +1,119 @@
+{
+  "case_id": "A_xhs_6a16c1d0",
+  "task": "Phase 2B (substance/form matching)",
+  "image_url_list": [],
+  "items_to_match": [
+    {
+      "path": "p1.s1.inputs[0]",
+      "name": "LibLib平台",
+      "value": "liblib.art web端（需浏览器访问，免登录可用部分额度）",
+      "type": "平台工具",
+      "related_images": []
+    },
+    {
+      "path": "p1.s1.outputs[0]",
+      "name": "图片生成配置",
+      "value": "LibLib AI 智能图片V2 · 风格模型：小红书风海报手绘涂鸦插画",
+      "type": "工具配置",
+      "related_images": []
+    },
+    {
+      "path": "p1.s2.inputs[0]",
+      "name": "海报主题构思",
+      "value": "儿童节主题选定：复古橘子汽水瓶 / 街机游戏机 / 泡泡糖扭蛋机 / 零食包装袋墙（小卖部）等童年元素",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p1.s2.outputs[0]",
+      "name": "儿童节主题提示词",
+      "value": "巨大复古橘子汽水玻璃瓶作为画面中心框架，瓶内是夏日童年场景（蓝天、白云、彩虹、气泡、小卖部、冰棒、西瓜），屏幕中跳出卡通小孩角色：穿彩色背带裤、手拿橘子汽水、戴小熊帽子，表情开心兴奋；另一位小朋友蹲在旁边玩泡泡机；道具元素：玻璃弹珠、吸管、橘子切片、冰块、爆米花、贴纸、气泡；纸张质感，高饱和撞色：红、蓝、黄、绿、黑色粗线条勾边，漫画感强烈，手绘涂鸦插画，复古童年元素与现代卡通结合，美式漫画爆炸贴效果，活泼动感，年轻潮流",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p1.s3.inputs[0]",
+      "name": "儿童节主题提示词",
+      "value": "巨大复古橘子汽水玻璃瓶作为画面中心框架，瓶内是夏日童年场景（蓝天、白云、彩虹、气泡、小卖部、冰棒、西瓜），屏幕中跳出卡通小孩角色：穿彩色背带裤、手拿橘子汽水、戴小熊帽子，表情开心兴奋；另一位小朋友蹲在旁边玩泡泡机；道具元素：玻璃弹珠、吸管、橘子切片、冰块、爆米花、贴纸、气泡；纸张质感，高饱和撞色：红、蓝、黄、绿、黑色粗线条勾边，漫画感强烈，手绘涂鸦插画，复古童年元素与现代卡通结合，美式漫画爆炸贴效果，活泼动感，年轻潮流",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p1.s3.inputs[1]",
+      "name": "图片生成配置",
+      "value": "LibLib AI 智能图片V2 · 风格模型：小红书风海报手绘涂鸦插画",
+      "type": "工具配置",
+      "related_images": []
+    },
+    {
+      "path": "p1.s3.outputs[0]",
+      "name": "2D涂鸦风插画海报",
+      "value": "<生成的2D平面涂鸦风儿童节主题插画海报，3:4竖版，高饱和漫画感，含主体容器+童年场景+卡通角色+手绘涂鸦风格+爆炸贴效果>",
+      "type": "图像",
+      "related_images": []
+    },
+    {
+      "path": "p2.s1.inputs[0]",
+      "name": "2D涂鸦风插画海报",
+      "value": "<P1生成的2D平面涂鸦风儿童节主题插画海报（如橘味汽水主题）>",
+      "type": "图像",
+      "related_images": []
+    },
+    {
+      "path": "p2.s1.inputs[1]",
+      "name": "3D效果提示词",
+      "value": "3d效果",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p2.s1.outputs[0]",
+      "name": "3D立体效果海报",
+      "value": "<3D立体效果版儿童节主题海报，在2D涂鸦插画基础上增加立体质感和3D渲染效果，人物和元素呈立体膨胀感>",
+      "type": "图像",
+      "related_images": []
+    },
+    {
+      "path": "p3.s1.inputs[0]",
+      "name": "静态海报图",
+      "value": "<选定的静态涂鸦风插画海报（P1或P2产出）>",
+      "type": "图像",
+      "related_images": []
+    },
+    {
+      "path": "p3.s1.outputs[0]",
+      "name": "视频生成配置",
+      "value": "Seedance 2.0 VIP · 全能参考 · 3:4 · 1080p · 6s · 有配音",
+      "type": "工具配置",
+      "related_images": []
+    },
+    {
+      "path": "p3.s2.inputs[0]",
+      "name": "静态海报图",
+      "value": "<静态涂鸦风插画海报（橘味汽水主题，P1产出）>",
+      "type": "图像",
+      "related_images": []
+    },
+    {
+      "path": "p3.s2.inputs[1]",
+      "name": "动态效果提示词",
+      "value": "瓶内夏日童年场景（蓝天、白云、彩虹、小卖部、冰棒、西瓜），卡通小孩从瓶中跳出挥舞橘子汽水，喷出大量透明泡泡；气泡持续上升，冰块碰撞晃动，橘子切片旋转，爆米花弹跳，贴纸与漫画元素漂浮；镜头轻微拉远镜，整体具有循环动画感",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p3.s2.inputs[2]",
+      "name": "视频生成配置",
+      "value": "Seedance 2.0 VIP · 全能参考 · 3:4 · 1080p · 6s · 有配音",
+      "type": "工具配置",
+      "related_images": []
+    },
+    {
+      "path": "p3.s2.outputs[0]",
+      "name": "动态视频海报",
+      "value": "<6秒循环动态视频海报，3:4竖版1080p，有配音，涂鸦风元素动态运动（气泡上升、冰块晃动、爆米花弹跳），适合节日氛围传播>",
+      "type": "视频",
+      "related_images": []
+    }
+  ]
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/_source.json

@@ -0,0 +1,18 @@
+{
+  "title": "AI一键生成六一儿童节插画海报｜附提示词🔥",
+  "link": "https://www.xiaohongshu.com/explore/6a16c1d0000000003601a30f",
+  "body_text": "嗨👋这里是布偶～\n马上六一了，这次我直接用 AI 做了一组小红书涂鸦风儿童节主题插画海报～\n主题全是小时候的快乐回忆：\n🍊橘子汽水🎮复古游戏机🍬泡泡糖\n整个画面做成了复古漫画 + 涂鸦海报的感觉\n真的越做越上头🥹\n而且这次让我很惊喜的是——\n海报排版和文字都稳定很多‼️\n\t\n✨我这次用的是：\nliblib AI 最新的「智能图片V2模型」\n相比之前的 nanobanana这次真的明显感觉到：\n✔画面丰富度更高\n✔文字生成更精准\n✔构图完成度更强\n✔人物动作自然很多\n✔AI味淡了！\n\t\n[看R]它出的完整度真的很高，画图没灵感也真的可以从这里找到很多元素[萌萌哒R]\n很多图几乎不用怎么后期。\n这次我还顺手生成了动态海报💕\n真的很适合节日氛围海报✨\n完整教程和提示词[星R][星R]\n我放在后面啦👇\n#liblib[话题]##AI插画[话题]##AI海报[话题]##AI教程[话题]##插画[话题]##小红书海报[话题]##小红书插画[话题]##六一儿童节[话题]##儿童节插画[话题]##矢量插画[话题]##扁平插画[话题]##人物海报[话题]##AIGC[话题]#",
+  "images": [
+    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320lt4iufmeog5nckb9lg88r2qhs0bso?imageView2/2/w/1080/format/webp",
+    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320lt4iufmep05nckb9lg88r2un1ei90?imageView2/2/w/1080/format/webp",
+    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320lt4iufmepg5nckb9lg88r2nl6q768?imageView2/2/w/1080/format/webp",
+    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320lt4iufmeq05nckb9lg88r2rk6e8m0?imageView2/2/w/1080/format/webp",
+    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320lt4iufmeqg5nckb9lg88r2qmcp3ig?imageView2/2/w/1080/format/webp",
+    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320lt4iufmer05nckb9lg88r2is9tkk0?imageView2/2/w/1080/format/webp",
+    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k8320lt4iufmerg5nckb9lg88r2oqfdug8?imageView2/2/w/1080/format/webp",
+    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k0320lt93r5me705nckb9lg88r2emmpkn0?imageView2/2/w/1080/format/webp",
+    "https://ci.xiaohongshu.com/notes_pre_post/1040g3k0320lt93r5me7g5nckb9lg88r24vqc1mg?imageView2/2/w/1080/format/webp"
+  ],
+  "publish_timestamp": "2026-05-27 20:39:43",
+  "channel_account_name": "xhs"
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/_trace.md

@@ -0,0 +1,464 @@
+
+
+---
+
+## ▶ Fresh @ 2026-06-03 16:18:46
+
+- case: `A_xhs_6a16c1d0`
+- source: `C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\runs_full\q0018\procedures\A_xhs_6a16c1d0\_source.json`
+- spec: `spec`
+- images: `9`
+- model: `claude-sonnet-4-6`
+- max_turns: `300`
+
+_session=6044e329 · model='claude-sonnet-4-6' · apiKeySource='none'_
+
+
+### Turn 1 · 16:18:54
+
+
+### Turn 2 · 16:18:56
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/README.md`
+
+
+### Turn 3 · 16:18:56
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools.md`
+
+
+### Turn 4 · 16:18:58
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/_source.json`
+
+
+### Turn 5 · 16:19:01
+
+
+### Turn 6 · 16:19:01
+> Now let me read the Phase 1 spec files in parallel.
+
+
+
+### Turn 7 · 16:19:02
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/syntax.md`
+
+
+### Turn 8 · 16:19:03
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/fields.md`
+
+
+### Turn 9 · 16:19:04
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/control-flow.md`
+
+
+### Turn 10 · 16:19:05
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/phase1-skeleton.md`
+
+
+### Turn 11 · 16:19:05
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/format/md-structure.md`
+
+
+### Turn 12 · 16:19:13
+
+
+### Turn 13 · 16:19:13
+> Now let me read the workflow template and Phase 2 spec file in parallel.
+
+
+
+### Turn 14 · 16:19:15
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/templates/workflow.template.json`
+
+
+### Turn 15 · 16:19:15
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/phase2-normalize.md`
+
+
+### Turn 16 · 16:23:36
+
+
+### Turn 17 · 16:23:36
+> Now I have all Phase 1 specs. Let me analyze the source and write the understanding.md and workflow.json.
+
+
+
+### Turn 18 · 16:24:16
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/understanding.md` (2,572 chars)
+
+
+### Turn 19 · 16:26:22
+
+
+### Turn 20 · 16:26:22
+> Now let me write the workflow.json skeleton:
+
+
+
+### Turn 21 · 16:27:37
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/workflow.json` (10,107 chars)
+
+
+### Turn 22 · 16:28:05
+
+
+### Turn 23 · 16:28:05
+> Now Phase 1.3 — add all IO anchors via wf-patch. Let me prepare the anchors patch file:
+
+
+
+### Turn 24 · 16:28:10
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/_scratch/anchors.json` (965 chars)
+
+
+### Turn 25 · 16:28:19
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/wf-patch.py --workflow "C:/Users/11304/gitlab/cybertogether/Agent/example...`
+
+
+### Turn 26 · 16:28:24
+
+
+### Turn 27 · 16:28:25
+> All 16 anchors applied. Now Phase 2 — run the prepare-subtask script, then launch sub-agents in parallel:
+
+
+
+### Turn 28 · 16:28:29
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/prepare-subtask.py --workflow "C:/Users/11304/gitlab/cybertogether/Agent/...`
+
+
+### Turn 29 · 16:28:33
+
+
+### Turn 30 · 16:28:34
+> Now launching Phase 2A and 2B sub-agents in parallel:
+
+
+
+### Turn 31 · 16:28:45
+- `Agent` — `Phase 2A: effect/action/type normalization for A_xhs_6a16c1d0` [subagent=phase-2a-normalizer]
+
+
+### Turn 32 · 16:28:48
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/_scratch/task_2a.json`
+
+
+### Turn 33 · 16:28:48
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/taxonomy/effect.json`
+
+
+### Turn 34 · 16:28:49
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/taxonomy/action.json`
+
+
+### Turn 35 · 16:28:49
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/taxonomy/type.json`
+
+
+### Turn 36 · 16:28:55
+- `Agent` — `Phase 2B: substance/form taxonomy matching for A_xhs_6a16c1d0` [subagent=phase-2b-matcher]
+
+
+### Turn 37 · 16:28:56
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/wf-patch.py` (offset=0, limit=80)
+
+
+### Turn 38 · 16:28:57
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/_scratch/task_2b.json`
+
+
+### Turn 39 · 16:28:58
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/workflow.json` (offset=0, limit=50)
+
+
+### Turn 40 · 16:29:04
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "儿童节主题海报构思 童年元素 创意主题选定"`
+
+
+### Turn 41 · 16:29:05
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "儿童节主题海报构思 童年元素 创意主题选定"`
+
+
+### Turn 42 · 16:29:08
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "创意构思 主题选定"`
+
+
+### Turn 43 · 16:29:09
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "图像生成提示词 涂鸦风插画"`
+
+
+### Turn 44 · 16:29:12
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "创意"`
+
+
+### Turn 45 · 16:29:12
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "涂鸦插画 图像"`
+
+
+### Turn 46 · 16:29:16
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "提示词 prompt 文字描述"`
+
+
+### Turn 47 · 16:29:16
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "提示词 文字"`
+
+
+### Turn 48 · 16:29:22
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "儿童节 节日 童年场景"`
+
+
+### Turn 49 · 16:29:24
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "海报 插画 涂鸦风 儿童"`
+
+
+### Turn 50 · 16:29:29
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "插画 海报 涂鸦风 手绘 平面设计"`
+
+
+### Turn 51 · 16:29:30
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "3D立体效果 海报"`
+
+
+### Turn 52 · 16:29:36
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "3D立体效果 膨胀感 三维渲染"`
+
+
+### Turn 53 · 16:29:37
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "动态视频 海报 节日氛围"`
+
+
+### Turn 54 · 16:29:42
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "动态视频 循环动画 动效"`
+
+
+### Turn 55 · 16:29:43
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "3D 立体 膨胀 渲染"`
+
+
+### Turn 56 · 16:29:50
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "插画 海报 平面作品"`
+
+
+### Turn 57 · 16:29:51
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "视频 动态 动画"`
+
+
+### Turn 58 · 16:29:55
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/_scratch/patch_2a.json` (1,865 chars)
+
+
+### Turn 59 · 16:29:56
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "AI生成视频 动态海报"`
+
+
+### Turn 60 · 16:29:57
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "视频 动态海报 竖版"`
+
+
+### Turn 61 · 16:30:02
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/wf-patch.py --workflow "../runs_full/q0018/procedures/A_xhs_6a16c1d0/work...`
+
+
+### Turn 62 · 16:30:03
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "动态提示词 文字指令 场景描述"`
+
+
+### Turn 63 · 16:30:04
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "3D效果 提示词 立体转换"`
+
+
+### Turn 64 · 16:30:10
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "文字 指令描述 纯文本"`
+
+
+### Turn 65 · 16:30:11
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "静态图片 海报 图像"`
+
+
+### Turn 66 · 16:30:16
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "静态图 海报 竖版图片"`
+
+
+### Turn 67 · 16:30:17
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "主题构思 创意设计方向"`
+
+
+### Turn 68 · 16:30:22
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --narrow "儿童节主题 童年元素 节日选题"`
+
+
+### Turn 69 · 16:30:24
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "图片 平面 静态"`
+
+
+### Turn 70 · 16:30:29
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --narrow "节日主题 童年场景 文化节日"`
+
+
+### Turn 71 · 16:30:30
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --narrow "涂鸦风 插画 卡通 艺术创作"`
+
+
+### Turn 72 · 16:30:34
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --subtree "/表象/视觉/实体/物品/影像物品"`
+
+
+### Turn 73 · 16:30:35
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "/表象/视觉/实体/物品/影像物品/艺术创作"`
+
+
+### Turn 74 · 16:30:39
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "表象/视觉/实体/物品/影像物品/艺术创作"`
+
+
+### Turn 75 · 16:30:42
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "表象/视觉/时间/节日/现代节日"`
+
+
+### Turn 76 · 16:30:43
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "理念/知识/文化/节气节日/节日"`
+
+
+### Turn 77 · 16:30:46
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "表象/视觉/实体/物品/影像物品/生活影像"`
+
+
+### Turn 78 · 16:30:48
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "呈现/视觉/视觉风格/视觉艺术风格/绘画风格/插画"`
+
+
+### Turn 79 · 16:30:54
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "呈现/视觉/视觉制作/构图编排/版面设计/版面结构"`
+
+
+### Turn 80 · 16:30:55
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "呈现/视觉/视觉制作/后期处理/图像合成处理/三维建模合成"`
+
+
+### Turn 81 · 16:30:59
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "呈现/视觉/视觉制作/后期处理/动态特效处理/运动变换"`
+
+
+### Turn 82 · 16:31:00
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "呈现/视觉/视觉制作/后期处理/图像合成处理/AI生成合成"`
+
+
+### Turn 83 · 16:31:04
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "表象/视觉/符号/文字符号/标注/提示引导"`
+
+
+### Turn 84 · 16:31:05
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "理念/知识/商业/前沿技术/AI智能/机器学习"`
+
+
+### Turn 85 · 16:31:09
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "呈现/视觉/视觉制作/构图编排/版面设计/字体标题"`
+
+
+### Turn 86 · 16:31:10
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "纯文字 段落 描述文字"`
+
+
+### Turn 87 · 16:31:15
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "节日 儿童节 现代节日"`
+
+
+### Turn 88 · 16:31:17
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "呈现/视觉/视觉制作/构图编排/版面设计/版面结构"`
+
+
+### Turn 89 · 16:31:46
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/_scratch/patch_2b.json` (2,415 chars)
+
+
+### Turn 90 · 16:31:51
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/wf-patch.py --workflow "C:/Users/11304/gitlab/cybertogether/Agent/example...`
+
+
+### Turn 91 · 16:32:36
+
+
+### Turn 92 · 16:32:36
+> Phase 2A and 2B both complete. Now Phase 3 — lint then render:
+
+
+
+### Turn 93 · 16:32:39
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/lint-case.py --workflow "C:/Users/11304/gitlab/cybertogether/Agent/exampl...`
+
+
+### Turn 94 · 16:32:44
+
+
+### Turn 95 · 16:32:44
+> Lint passes clean. Now render the HTML:
+
+
+
+### Turn 96 · 16:32:48
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/render-case.py --workflow "C:/Users/11304/gitlab/cybertogether/Agent/exam...`
+
+
+### Turn 97 · 16:33:10
+
+
+### Turn 98 · 16:33:10
+> HTML rendered (3 procedures, 6 steps). Now write the .md deliverable:
+
+
+
+### Turn 99 · 16:34:44
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/case-A_xhs_6a16c1d0-liblib-ai-children-poster.md` (5,964 chars)
+
+
+### Turn 100 · 16:34:53
+
+
+### Turn 101 · 16:35:09
+> 所有产物已生成完毕。以下是本次处理的完整总结：
+>
+> ---
+>
+> ## 工序梗概
+>
+> **原帖**：小红书「布偶」教程——用 LibLib AI 智能图片V2（image 2 模型）制作六一儿童节涂鸦风插画海报。
+>
+> 判定 **3 个独立工序**：
+>
+> | 工序 | 工具 | 输入 → 产出 | 步骤数 |
+> |---|---|---|---|
+> | **P1** 涂鸦风2D平面插画海报生成 | LibLib AI 智能图片V2（涂鸦插画风格） | 主题构思 → 2D涂鸦风插画海报（无需垫图） | 3 |
+> | **P2** 3D立体效果海报生成 | LibLib AI 智能图片V2（3D模型变体） | 2D插画 + "3d效果" → 3D立体效果海报 | 1 |
+> | **P3** 动态视频海报生成 | Seedance 2.0 VIP | 静态海报 + 动态提示词 → 6s动态视频 | 2 |
+>
+> ---
+>
+> ## 输出文件路径
+>
+> | 产物 | 路径 |
+> |---|---|
+> | 心智模型 | `.../A_xhs_6a16c1d0/understanding.md` |
+> | 工序定义 (主产物) | `.../A_xhs_6a16c1d0/workflow.json` |
+> | HTML 可视化 | `.../A_xhs_6a16c1d0/case-A_xhs_6a16c1d0-liblib-ai-children-poster.html` |
+> | DSL 文本版 | `.../A_xhs_6a16c1d0/case-A_xhs_6a16c1d0-liblib-ai-children-poster.md` |
+>
+> ---
+>
+> ## 关键 DSL 发现
+>
+> 1. **提示词5段式模板可工序化**：博主的4套主题提示词共享「容器框架/场景/角色/道具/风格词」结构，属于可抽象为 resource 的跨主题模板，当前以单步 human 撰写的形式被捕获。
+>
+> 2. **"产出物接力"≠ 合并工序**：P1→P2→P3 依次接力，但工具/方法/产出物维度均有实质差异，必须拆为 3 个 procedure，不能以"流水线连续"为由合并。
+>
+> 3. **极简 directive 也需保留独立工序**：P2 的整个 directive 只有 3 个字（"3d效果"），但它改变了产出物的视觉维度（2D→3D），是独立的工艺价值节点。
+>
+> 4. **动态提示词策略属于跨主题可复用 trick**："按元素描述运动 + 镜头运动 + 循环感兜底"这一结构知识被记在 `instruction[memo]` 而非 `intent`，准确区分了 caller 描述与工艺经验。
+>
+> 5. **Lint 全绿**：type 完整性 OK，value 自包含 OK；新增 2 个 case-specific 类型（`平台工具`/`工具配置`，均 extends `生成参数`）已自动 record 到 `type_suggestions.md`。
+
+
+
+### ◀ Result · 16:35:10
+
+- subtype: `success` · is_error: `False`
+- num_turns: `21` · duration: `982020ms` · wall: `984.2s`
+- tokens: in=17 out=47,047 cache_w=113,593 cache_r=1,111,949
+- cost: sdk=$2.4189, est_if_api=$1.4653

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/case-A_xhs_6a16c1d0-liblib-ai-children-poster.md

@@ -0,0 +1,239 @@
+# Case A_xhs_6a16c1d0: LibLib AI 儿童节涂鸦风插画海报全流程
+
+**Source**: 小红书 · @一只布偶偶 · 2026-05-27  
+**URL**: https://www.xiaohongshu.com/explore/6a16c1d0000000003601a30f  
+**主题**: 用 LibLib AI 智能图片V2 制作六一儿童节涂鸦风插画海报，含 2D 生成、3D 转换、动态视频三条工序
+
+---
+
+## 工序梗概 (人话)
+
+### P1 · 涂鸦风2D平面插画海报生成
+
+打开 LibLib AI 的「智能图片V2」模型，选择「小红书风海报手绘涂鸦插画」风格，撰写一段结构化提示词（以橘味汽水为例：复古大瓶子 + 童年场景 + 卡通小孩 + 道具细节 + 高饱和漫画感风格词），直接提交生成。**无需垫图，一键直出**，输出 3:4 竖版 2D 平面涂鸦风插画海报。
+
+博主提供了 4 套主题提示词模板（橘味汽水/复古街机/泡泡糖机/儿时小卖部），结构完全一致，只替换主体容器和场景元素。
+
+### P2 · 3D立体效果海报生成
+
+在同一 LibLib 智能图片V2 界面切换选择 **3D 模型变体**，上传 P1 产出的 2D 涂鸦插画作为垫图，输入简短提示词「3d效果」，生成具有立体质感和膨胀渲染效果的 3D 版海报。输入极简，工具负责2D→3D风格迁移。
+
+### P3 · 动态视频海报生成
+
+在 LibLib 平台切换到「视频生成」，选择 **Seedance 2.0** 模型，上传静态海报图，输入动态效果提示词（描述图中元素的运动方式：气泡上升、冰块晃动、爆米花弹跳、镜头轻微拉远 + 循环动画感），6 秒动态视频海报自动生成，1080p · 有配音。
+
+---
+
+## 引用的类型 (stdlib + 自定义)
+
+### stdlib 类型 (直接引用)
+
+| type | 在工序中出现 | 说明 |
+|---|---|---|
+| 提示词 | P1.s2输出, P1.s3/P2.s1/P3.s2输入 | 向生成模型发出的正向提示词文本 |
+| 图像 | P1.s3输出 → P2/P3输入，P2.s1输出 | 2D涂鸦插画海报、3D立体效果海报 |
+| 参考图 | P3.s1/s2输入 | 作为视频生成视觉参考的静态海报图 |
+| 底图 | P1.s3输出, P2.s1输入输出 | 供图生图改造的中间态2D/3D插画 |
+| 描述 | P1.s2输入 | 对主题元素的文字刻画 |
+| 视频成品 | P3.s2输出 | 6秒完整动态视频 |
+
+### case-specific 类型 (type_registry)
+
+| type | extends | 出现工序 | desc |
+|---|---|---|---|
+| 平台工具 | 生成参数 | P1 | 操作入口所在的平台及工具信息，说明使用哪个 Web 平台访问哪个 AI 工具 |
+| 工具配置 | 生成参数 | P1, P3 | 工具与模型的综合配置项，含平台模型选型、风格模型选择及规格参数等 |
+
+---
+
+## L1 外部函数库
+
+| L1 函数 | 绑定工具 | 平台 | 用途 |
+|---|---|---|---|
+| `human` | 用户手动操作 | — | UI 导航 / 提示词撰写 / 参数配置 |
+| `liblib_ai_智能图片v2` | LibLib AI 智能图片V2 (image 2 模型) | liblib.art | 图像生成（文生图 / 图生图 / 2D→3D 转换） |
+| `seedance_2_0` | Seedance 2.0 VIP | liblib.art 视频生成 | 图生视频，生成动态海报 |
+
+---
+
+## L2 抽象动作 + impl 关系
+
+| L2 抽象动作 | impl via | 出现步骤 |
+|---|---|---|
+| 获取/录入/键入 | `human` | P1.s1 (配置工具), P3.s1 (配置视频参数) |
+| 生成/元素生成 | `human` | P1.s2 (撰写提示词) |
+| 生成/元素生成 | `liblib_ai_智能图片v2` | P1.s3 (文生图) |
+| 修改/变/风格化 | `liblib_ai_智能图片v2` | P2.s1 (2D→3D 风格化) |
+| 生成/元素生成 | `seedance_2_0` | P3.s2 (图生视频) |
+
+---
+
+## L3 工序模板
+
+### 模板 A：涂鸦风2D平面插画海报生成 (对应 P1)
+
+```
+输入: 海报主题构思 (描述)
+  s1 [human, 幂等]:  获取/录入/键入  → 图片生成配置 (工具配置)
+  s2 [human, 人工]:  生成/元素生成   海报主题构思 → 儿童节主题提示词 (提示词)
+  s3 [liblib_ai_智能图片v2, 随机]: 生成/元素生成  提示词 + 工具配置 → 2D涂鸦风插画海报 (底图)
+返回: 图像 (2D涂鸦风插画海报)
+```
+
+**工序特点**: 无需垫图的纯文生图流程；提示词结构模板化，5段式（容器框架 + 场景 + 角色 + 道具 + 风格词）可迁移到不同主题。
+
+### 模板 B：3D立体效果海报生成 (对应 P2)
+
+```
+输入: 2D涂鸦风插画海报 (底图)
+  s1 [liblib_ai_智能图片v2, 随机]: 修改/变/风格化  底图 + 3D效果提示词 → 3D立体效果海报 (底图)
+返回: 图像 (3D立体效果海报)
+```
+
+**工序特点**: 单步图生图，依赖 P1 的底图；3D提示词极简（"3d效果"），模型内化3D转换逻辑；需在智能图片V2中切换3D模型变体。
+
+### 模板 C：动态视频海报生成 (对应 P3)
+
+```
+输入: 静态海报图 (参考图)
+  s1 [human, 幂等]:  获取/录入/键入  静态海报图 → 视频生成配置 (工具配置)
+  s2 [seedance_2_0, 随机]: 生成/元素生成  静态海报图 + 动态效果提示词 + 视频生成配置 → 动态视频海报 (视频成品)
+返回: 视频 (动态视频海报)
+```
+
+**工序特点**: 图生视频；动态提示词核心策略是逐元素描述运动方式 + 镜头运动 + "循环动画感"兜底；与 P1/P2 共享同一平台 (LibLib)。
+
+---
+
+## L4 工序实例
+
+### P1 实例 — 橘味汽水主题插画海报 (2D)
+
+**inputs (本次实际值)**
+- `海报主题构思` → 儿童节主题选定：复古橘子汽水瓶 / 街机游戏机 / 泡泡糖扭蛋机 / 零食包装袋墙（小卖部）等童年元素
+
+**bindings (本次工具选择)**
+- `s1.via = human`
+- `s2.via = human`
+- `s3.via = liblib_ai_智能图片v2` (智能图片V2, 小红书风海报手绘涂鸦插画风格, image 2 模型)
+
+**extracted_values (本次中间产物 — 原文回填)**
+
+_s1 — 工具配置_
+```
+LibLib AI 智能图片V2 · 风格模型：小红书风海报手绘涂鸦插画
+```
+
+_s2 — 儿童节主题提示词（橘味汽水主题，STEP 3 图）_
+```
+巨大复古橘子汽水玻璃瓶作为画面中心框架，瓶内是夏日童年场景（蓝天、白云、彩虹、气泡、
+小卖部、冰棒、西瓜），屏幕中跳出卡通小孩角色：穿彩色背带裤、手拿橘子汽水、戴小熊帽子，
+表情开心兴奋；另一位小朋友蹲在旁边玩泡泡机；道具元素：玻璃弹珠、吸管、橘子切片、冰块、
+爆米花、贴纸、气泡；纸张质感，高饱和撞色：红、蓝、黄、绿、黑色粗线条勾边，漫画感强烈，
+手绘涂鸦插画，复古童年元素与现代卡通结合，美式漫画爆炸贴效果，活泼动感，年轻潮流
+```
+
+_附录：其他3套主题提示词（提示词分享图）_
+
+**超级游戏世界·复古街机**（图7）:
+```
+巨大复古街机游戏机作为画面中心框架，屏幕内：像素风游戏场景（蓝天、白云、金币、砖块、
+管道、爱心）；屏幕中跳出卡通小孩角色：戴彩色鸭舌帽、穿宽松背带裤、手拿游戏手柄，表情
+夸张兴奋；另一位小朋友坐在地上操作街机摇杆；道具元素：游戏币、街机摇杆、彩色按钮、爆
+米花桶、金币、问号方块、游戏手柄；纸张质感，高饱和撞色：红、蓝、黄、绿、黑色粗线条勾
+边，漫画感强烈，手绘涂鸦插画，复古游戏像素风与现代卡通结合，美式漫画爆炸贴效果，活泼
+动感，年轻潮流
+```
+
+**甜到冒泡的快乐·泡泡糖机**（图8）:
+```
+巨大复古泡泡糖机作为画面中心框架，透明糖果球内部是梦幻童年场景（彩虹、云朵、星星、棉
+花糖、糖果店），卡通小孩角色从泡泡糖球中跳出：穿黑色卫衣、戴兔耳帽、吹着巨大泡泡糖，
+表情搞怪开心；另一位小朋友抱着糖果罐坐在旁边；道具元素：棒棒糖、QQ糖、糖纸、气球、
+爆米花、漫画贴纸；纸张质感高饱和撞色：红、蓝、黄、绿、黑色粗线条勾边，漫画感强烈，手
+绘涂鸦插画，复古印刷质感，复古童年元素与现代卡通结合，美式漫画爆炸贴效果，活泼动感，
+年轻潮流
+```
+
+**儿时零食大满贯·小卖部**（图9）:
+```
+巨大童年零食包装袋墙作为画面中心框架，内部是复古小卖部场景（辣条、泡泡糖、汽水、薯片、
+彩虹糖），卡通小孩角色背着小书包从零食包装中跳出，表情兴奋；另一位小朋友在地上拼零食；
+道具元素：零食包装、游戏币、贴纸、漫画爆炸框、星星；纸张质感高饱和和红蓝黄绿橙色，黑
+色粗线条勾边，漫画感强烈，手绘涂鸦插画，美式漫画爆炸贴效果，活泼动感，年轻潮流
+```
+
+_s3 — 生成配置_
+```
+智能图片V2 · 风格：小红书风海报手绘涂鸦插画 · 比例：3:4 · 数量：4张 · 无需垫图
+```
+
+**trace** (可空，原文未提供耗时/费用)
+
+---
+
+### P2 实例 — 橘味汽水3D立体效果海报
+
+**inputs (本次实际值)**
+- `2D涂鸦风插画海报` → P1.s3 产出的橘味汽水主题2D涂鸦风插画海报
+
+**bindings (本次工具选择)**
+- `s1.via = liblib_ai_智能图片v2` (3D模型变体，区别于P1的涂鸦插画风格模型)
+
+**extracted_values (本次中间产物 — 原文回填)**
+
+_s1 — 3D效果提示词（推断，原文极简）_
+```
+3d效果
+```
+_(注: 原文仅说"输入插画和提示词：3d效果即可生成"，以上为推断值，inferred: true)_
+
+_s1 — 生成配置_
+```
+智能图片V2 · 3D模型变体 · 比例：3:4 · 数量：4张
+```
+
+**trace** (可空，原文未提供)
+
+---
+
+### P3 实例 — 橘味汽水动态视频海报
+
+**inputs (本次实际值)**
+- `静态海报图` → P1/P2 产出的涂鸦风插画海报（P1 2D版或P2 3D版均可）
+
+**bindings (本次工具选择)**
+- `s1.via = human`
+- `s2.via = seedance_2_0` (Seedance 2.0 VIP)
+
+**extracted_values (本次中间产物 — 原文回填)**
+
+_s1 — 视频生成配置_
+```
+Seedance 2.0 VIP · 全能参考 · 比例：3:4 · 分辨率：1080p · 时长：6s · 有配音
+```
+
+_s2 — 动态效果提示词（STEP 1 图）_
+```
+瓶内夏日童年场景（蓝天、白云、彩虹、小卖部、冰棒、西瓜），卡通小孩从瓶中跳出挥舞橘子
+汽水，喷出大量透明泡泡；气泡持续上升，冰块碰撞晃动，橘子切片旋转，爆米花弹跳，贴纸与
+漫画元素漂浮；镜头轻微拉远镜，整体具有循环动画感
+```
+
+**trace**
+- 输出图尺寸: 1248×1664 (由 Seedance VIP 推算)
+- 时长: 6s
+
+---
+
+## 这个 case 对 DSL 设计的关键启发
+
+1. **提示词模板工序化**：博主的4套主题提示词不是随机的，而是共享一个5段式结构模板（容器框架/场景/角色/道具/风格词）。这种"主题替换"模式在 DSL 中应表现为：s2（撰写提示词）的 intent 可用 `{act:按模板撰写}` 语义，或在 declarations 中声明一个"提示词模板"类型的 resource（跨主题复用）。当前因仅示范一主题，简化为单步撰写。
+
+2. **三条独立工序的边界判定**：P1 (文生图) → P2 (图生图/风格转换) → P3 (图生视频) 是天然流水线，每条工序的输入依赖前一条输出，但三者在 **工具/method/产出物类型** 上都有本质差异，因此拆为3个 procedure 而非拍平成单一流程。这说明"产出物接力"不能作为合并工序的理由。
+
+3. **P2 极简工序的 DSL 表达**：P2 只有1个步骤，且核心 directive 只有3个字（"3d效果"），大量信息被内化进模型能力（2D→3D风格迁移）。DSL 中这种"极简调用但重语义"的 step 要保留而不合并进 P1，因为它改变了产出物的视觉维度（2D→3D），是独立的工艺价值节点。
+
+4. **动态提示词的结构知识**：P3.s2 的动态提示词体现了一套 "按元素描述运动 + 镜头运动 + 循环感兜底" 的 prompt 工程知识。这类 trick 应记入 step.instruction[memo] 而非 intent，因为它是跨主题可复用的工艺经验，不是本次调用的 caller-side 描述。
+
+5. **inferred 补全的合理边界**：P1.s1 的 LibLib平台 input 和 P2.s1 的 3D提示词都被标为 inferred，但置信度不同——前者是工艺必经（高置信，标A），后者是原文极简导致的文本推断（中等置信，value 填了合理猜测但标注了 inferred_reason）。两种 inferred 在 DSL 中应有相同标记，但 reviewer 的处理优先级不同。

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/understanding.md

@@ -0,0 +1,86 @@
+# Phase 1.1 · 心智模型 — A_xhs_6a16c1d0
+
+## 原文概要
+
+**标题**: AI一键生成六一儿童节插画海报｜附提示词🔥  
+**作者**: 布偶（@一只布偶偶）  
+**平台**: 小红书 (xhs)  
+**发布**: 2026-05-27  
+**URL**: https://www.xiaohongshu.com/explore/6a16c1d0000000003601a30f
+
+博主分享利用 LibLib AI「智能图片V2」模型（即新出的 image 2 模型）制作六一儿童节主题涂鸦风插画海报的完整流程，并顺手生成动态海报。内容包含：①2D平面涂鸦插画海报生成、②3D立体效果海报生成（同工具切换3D模型）、③动态视频海报生成（Seedance 2.0），附多主题提示词分享（橘味汽水、超级游戏世界、泡泡糖机、儿时零食小卖部）。
+
+---
+
+## 多工序判断
+
+**判定：3 个独立工序**
+
+以原文中明确展示的终态产物为扫描单元：
+
+| 产出物 | 对应段落 | 做法 | 工具 |
+|---|---|---|---|
+| 2D涂鸦风插画海报（4张主题） | ①图片生成 STEP 1-3 | 文生图，涂鸦插画风格模型 | LibLib AI 智能图片V2（涂鸦风格） |
+| 3D立体效果海报 | ①图片生成 STEP 4 | 图生图，切换3D模型变体 | LibLib AI 智能图片V2（3D模型） |
+| 动态视频海报 | ③动态海报生成 STEP 1-2 | 图生视频，Seedance 2.0 | LibLib AI Seedance 2.0 |
+
+**工序识别推理**：
+
+- **P1 vs P2**：同为"图片生成"大类，但 P1 是"文生图"（无参考图，直接从文字提示词生成），P2 是"图生图"（以P1产物为参考，切换3D模型，加"3D效果"提示词）。输入不同、模型不同、产出物维度（2D vs 3D）不同 → **算两个工序**。
+- **P2 vs P3**：P3 完全换用视频生成工具（Seedance 2.0），产出物从图像变成视频，方法截然不同 → **明确两个工序**。
+- **提示词share（④⑤⑥⑦）**：附录式内容，展示4套主题提示词，没有引入新做法 → **不额外计工序**，其提示词文本作为 P1 步骤的 value 补充。
+
+---
+
+## 各工序梗概
+
+### P1 · 涂鸦风2D平面插画海报生成
+
+**终态产物**: 小红书风手绘涂鸦插画海报（2D平面，4:3竖版）  
+**核心创新**: LibLib AI 智能图片V2 + 小红书风海报手绘涂鸦插画风格模型，无需垫图一键直出  
+**大致步骤数**: 3步（进入平台配置 → 撰写提示词 → AI生成）  
+**关键提示词示例**（橘味汽水主题，来自STEP 3图）:  
+> 巨大复古橘子汽水玻璃瓶作为画面中心框架，瓶内是夏日童年场景（蓝天、白云、彩虹、气泡、小卖部、冰棒、西瓜），屏幕中跳出卡通小孩角色：穿彩色背带裤、手拿橘子汽水、戴小熊帽子，表情开心兴奋；另一位小朋友蹲在旁边玩泡泡机；道具元素：玻璃弹珠、吸管、橘子切片、冰块、爆米花、贴纸、气泡；纸张质感，高饱和撞色：红、蓝、黄、绿、黑色粗线条勾边，漫画感强烈，手绘涂鸦插画，复古童年元素与现代卡通结合，美式漫画爆炸贴效果，活泼动感，年轻潮流
+
+### P2 · 3D立体效果海报生成
+
+**终态产物**: 3D立体效果插画海报（以P1输出作为基础）  
+**核心创新**: 同一平台（LibLib 智能图片V2）切换为3D模型变体，实现2D→3D转换  
+**大致步骤数**: 1步（上传2D插画+简短3D提示词 → AI生成3D效果）  
+**工艺要点**: 不需要新的详细提示词，只需选择3D模型并输入"3d效果"即可
+
+### P3 · 动态视频海报生成
+
+**终态产物**: 动态视频海报（6秒，1080p，3:4，有配音）  
+**核心创新**: LibLib 平台内置 Seedance 2.0 视频生成，将静态海报图转为循环动态效果  
+**大致步骤数**: 2步（选图+配置 → AI生成视频）  
+**关键动态提示词**（来自图5 STEP 1）:  
+> 瓶内夏日童年场景（蓝天、白云、彩虹、小卖部、冰棒、西瓜），卡通小孩从瓶中跳出挥舞橘子汽水，喷出大量透明泡泡；气泡持续上升，冰块碰撞晃动，橘子切片旋转，爆米花弹跳，贴纸与漫画元素漂浮；镜头轻微拉远镜，整体具有循环动画感
+
+---
+
+## 工具链
+
+| 工具 | 用途 | 特点 |
+|---|---|---|
+| LibLib AI 智能图片V2（涂鸦插画风格） | P1 文生图 | 画面丰富度高，文字生成精准，AI味淡，不需垫图 |
+| LibLib AI 智能图片V2（3D模型变体） | P2 图生图（2D→3D） | 同平台，切换模型，输入原图即可 |
+| LibLib AI Seedance 2.0 VIP | P3 图生视频 | 6s循环动态，全能参考模式 |
+
+## 提示词架构分析（P1）
+
+4套主题提示词（来自提示词分享图）结构高度一致：
+1. **主体元素**：复古大型容器（瓶/机器/包装袋）作为画面中心框架
+2. **场景内容**：容器内或周围的童年场景元素（小卖部、零食、玩具）
+3. **角色动作**：卡通小孩从容器中跳出，表情夸张兴奋
+4. **道具细节**：具体小物件列举（弹珠、贴纸、零食包装等）
+5. **风格修饰**：统一的视觉风格词——纸张质感、高饱和撞色、漫画感强烈、手绘涂鸦、美式漫画爆炸贴效果、活泼动感、年轻潮流
+
+---
+
+## 关键发现
+
+1. **提示词模板化**：4套主题共享相同的提示词结构模板，只替换主体容器和场景元素，可泛化为可复用模板
+2. **零垫图一键直出**：P1 完全文生图，无需参考图，降低了创作门槛
+3. **工序串联**：P1→P2→P3 是天然流水线，每步都以上一步的产出物作为输入
+4. **工具集成**：所有工序在LibLib一个平台内完成，减少了工具切换成本

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_xhs_6a16c1d0/workflow.json

@@ -0,0 +1,447 @@
+{
+  "source": {
+    "platform": "xhs",
+    "author": "一只布偶偶",
+    "date": "2026-05-27",
+    "url": "https://www.xiaohongshu.com/explore/6a16c1d0000000003601a30f",
+    "title": "AI一键生成六一儿童节插画海报｜附提示词🔥",
+    "excerpt": "用 LibLib AI 智能图片V2 一键生成儿童节涂鸦风插画海报（无需垫图）；同工具切换3D模型生成立体效果；再用 Seedance 2.0 转成动态视频海报。附4套童年主题提示词分享。"
+  },
+  "procedures": [
+    {
+      "id": "p1",
+      "name": "涂鸦风2D平面插画海报生成",
+      "purpose": "利用 LibLib AI 智能图片V2（涂鸦插画风格模型），从主题提示词无需垫图直接生成儿童节小红书风手绘涂鸦插画海报",
+      "category": "产物创造",
+      "platform": "xhs",
+      "author": "一只布偶偶",
+      "declarations": {
+        "inputs": [
+          {
+            "type": "提示词",
+            "name": "海报主题构思",
+            "desc": "儿童节主题灵感，如：橘味汽水、复古游戏机、泡泡糖机、儿时零食小卖部等童年元素"
+          }
+        ],
+        "resources": [],
+        "returns": {
+          "type": "图像"
+        }
+      },
+      "steps": [
+        {
+          "id": "s1",
+          "kind": "step",
+          "effect": "预处理",
+          "via": "human",
+          "action": "获取/录入/键入",
+          "feature": "幂等",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "打开 liblib.art web 端，点击左侧菜单「图片生成」，在模型区选择「智能图片V2」，风格模型选择「小红书风海报手绘涂鸦插画」（即新出的 image 2 模型）"
+            ],
+            [
+              "memo",
+              "智能图片V2 = LibLib 最新 image 2 模型，相比旧版 nanobanana：画面丰富度更高、文字生成更精准、构图完成度更强、人物动作更自然、AI味淡"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "平台工具",
+              "name": "LibLib平台",
+              "value": "liblib.art web端（需浏览器访问，免登录可用部分额度）",
+              "anchor": "← 字面量",
+              "inferred": true,
+              "inferred_reason": "原文说'打开liblib web端'，隐含需要浏览器访问此平台，为工艺必要前提",
+              "substance": null,
+              "form": null
+            }
+          ],
+          "outputs": [
+            {
+              "type": "工具配置",
+              "name": "图片生成配置",
+              "value": "LibLib AI 智能图片V2 · 风格模型：小红书风海报手绘涂鸦插画",
+              "anchor": "→ s3.图片生成配置",
+              "substance": null,
+              "form": null
+            }
+          ],
+          "intent": "用 {via:human} {act:配置工具} 在 {in-type:平台工具} 中选定 {out-type:工具配置}（智能图片V2 + 涂鸦插画风格）",
+          "focus": [
+            "via",
+            "action",
+            "out-type-0"
+          ]
+        },
+        {
+          "id": "s2",
+          "kind": "step",
+          "effect": "预处理",
+          "via": "human",
+          "action": "生成/元素生成",
+          "feature": "人工",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "根据儿童节主题构思，按模板结构撰写中文提示词：①大型复古主体容器作画面中心框架 ②容器内童年场景元素 ③卡通小孩角色（服装、动作、表情） ④道具细节列举 ⑤统一视觉风格词（纸张质感、高饱和撞色、黑色粗线条、漫画感、手绘涂鸦、美式漫画爆炸贴效果、活泼动感、年轻潮流）"
+            ],
+            [
+              "memo",
+              "4套主题提示词（图中附件分享）：①橘味汽水 ②超级游戏世界·街机 ③甜到冒泡的快乐·泡泡糖机 ④儿时零食大满贯·小卖部 — 结构高度一致，只替换主体容器和场景元素即可泛化"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "描述",
+              "name": "海报主题构思",
+              "value": "儿童节主题选定：复古橘子汽水瓶 / 街机游戏机 / 泡泡糖扭蛋机 / 零食包装袋墙（小卖部）等童年元素",
+              "anchor": "← 工序输入",
+              "substance": [
+                "/表象/视觉/时间/节日/现代节日",
+                "/理念/知识/文化/节气节日/节日"
+              ],
+              "form": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "提示词",
+              "name": "儿童节主题提示词",
+              "value": "巨大复古橘子汽水玻璃瓶作为画面中心框架，瓶内是夏日童年场景（蓝天、白云、彩虹、气泡、小卖部、冰棒、西瓜），屏幕中跳出卡通小孩角色：穿彩色背带裤、手拿橘子汽水、戴小熊帽子，表情开心兴奋；另一位小朋友蹲在旁边玩泡泡机；道具元素：玻璃弹珠、吸管、橘子切片、冰块、爆米花、贴纸、气泡；纸张质感，高饱和撞色：红、蓝、黄、绿、黑色粗线条勾边，漫画感强烈，手绘涂鸦插画，复古童年元素与现代卡通结合，美式漫画爆炸贴效果，活泼动感，年轻潮流",
+              "anchor": "→ s3.儿童节主题提示词",
+              "substance": "/表象/视觉/符号/文字符号/标注/提示引导",
+              "form": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"
+            }
+          ],
+          "intent": "用 {via:human} 人工 {act:撰写提示词}，从 {in-type:提示词}（主题构思）生成详细的 {out-type:提示词}（涂鸦风格画面描述）",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ]
+        },
+        {
+          "id": "s3",
+          "kind": "step",
+          "effect": "主体生成",
+          "via": "liblib_ai_智能图片v2",
+          "action": "生成/元素生成",
+          "feature": "随机",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "在智能图片V2界面将提示词粘贴到输入框，选择比例3:4，生成数量4张，点击生成，无需垫图"
+            ],
+            [
+              "config",
+              "智能图片V2 · 风格：小红书风海报手绘涂鸦插画 · 比例：3:4 · 数量：4张"
+            ],
+            [
+              "memo",
+              "无需垫图，一键直出。很多图几乎不用怎么后期处理。"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "提示词",
+              "name": "儿童节主题提示词",
+              "value": "巨大复古橘子汽水玻璃瓶作为画面中心框架，瓶内是夏日童年场景（蓝天、白云、彩虹、气泡、小卖部、冰棒、西瓜），屏幕中跳出卡通小孩角色：穿彩色背带裤、手拿橘子汽水、戴小熊帽子，表情开心兴奋；另一位小朋友蹲在旁边玩泡泡机；道具元素：玻璃弹珠、吸管、橘子切片、冰块、爆米花、贴纸、气泡；纸张质感，高饱和撞色：红、蓝、黄、绿、黑色粗线条勾边，漫画感强烈，手绘涂鸦插画，复古童年元素与现代卡通结合，美式漫画爆炸贴效果，活泼动感，年轻潮流",
+              "anchor": "← s2.儿童节主题提示词",
+              "substance": "/表象/视觉/符号/文字符号/标注/提示引导",
+              "form": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"
+            },
+            {
+              "type": "工具配置",
+              "name": "图片生成配置",
+              "value": "LibLib AI 智能图片V2 · 风格模型：小红书风海报手绘涂鸦插画",
+              "anchor": "← s1.图片生成配置",
+              "substance": null,
+              "form": null
+            }
+          ],
+          "outputs": [
+            {
+              "type": "底图",
+              "name": "2D涂鸦风插画海报",
+              "value": "<生成的2D平面涂鸦风儿童节主题插画海报，3:4竖版，高饱和漫画感，含主体容器+童年场景+卡通角色+手绘涂鸦风格+爆炸贴效果>",
+              "anchor": "→ 返回 图像",
+              "substance": "/表象/视觉/实体/物品/影像物品/艺术创作",
+              "form": [
+                "/呈现/视觉/视觉风格/视觉艺术风格/绘画风格/插画",
+                "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"
+              ]
+            }
+          ],
+          "intent": "用 {via:liblib_ai_智能图片v2} {act:文生图} 从 {in-type:提示词} 直接生成 {out-type:图像}（2D涂鸦风插画海报）",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ]
+        }
+      ],
+      "return_row": {
+        "arrow": "↩",
+        "text": "<span class=\"kw\">返回</span> <span class=\"chip\" data-type=\"图像\">图像</span>（2D涂鸦风插画海报）"
+      },
+      "type_registry": {
+        "平台工具": {
+          "extends": "生成参数",
+          "desc": "操作入口所在的平台及工具信息，说明使用哪个 Web 平台访问哪个 AI 工具"
+        },
+        "工具配置": {
+          "extends": "生成参数",
+          "desc": "工具与模型的综合配置项，含平台模型选型、风格模型选择及规格参数等"
+        }
+      }
+    },
+    {
+      "id": "p2",
+      "name": "3D立体效果海报生成",
+      "purpose": "以P1生成的2D涂鸦插画为参考图，在LibLib AI 智能图片V2中切换3D模型，输入简短3D提示词，生成3D立体效果版插画海报",
+      "category": "产物创造",
+      "platform": "xhs",
+      "author": "一只布偶偶",
+      "declarations": {
+        "inputs": [
+          {
+            "type": "图像",
+            "name": "2D涂鸦风插画海报",
+            "desc": "P1生成的2D平面涂鸦插画海报，作为3D转换的参考垫图"
+          }
+        ],
+        "resources": [],
+        "returns": {
+          "type": "图像"
+        }
+      },
+      "steps": [
+        {
+          "id": "s1",
+          "kind": "step",
+          "effect": "主体生成",
+          "via": "liblib_ai_智能图片v2",
+          "action": "修改/变/风格化",
+          "feature": "随机",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "在智能图片V2中切换选择「3D效果」或合适的3D模型，上传2D涂鸦插画作为垫图参考，输入提示词，点击生成"
+            ],
+            [
+              "config",
+              "智能图片V2 · 3D模型变体 · 比例：3:4 · 数量：4张"
+            ],
+            [
+              "memo",
+              "新出的 image 2 模型（智能图片V2）3D效果生成明显优于旧版 nanobanana，图像丰富度和文字精准度均提升一个 level，相较 nanobanana 更全面提升，真做到去AI味"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "底图",
+              "name": "2D涂鸦风插画海报",
+              "value": "<P1生成的2D平面涂鸦风儿童节主题插画海报（如橘味汽水主题）>",
+              "anchor": "← 工序输入",
+              "substance": "/表象/视觉/实体/物品/影像物品/艺术创作",
+              "form": [
+                "/呈现/视觉/视觉风格/视觉艺术风格/绘画风格/插画",
+                "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"
+              ]
+            },
+            {
+              "type": "提示词",
+              "name": "3D效果提示词",
+              "value": "3d效果",
+              "anchor": "← 字面量",
+              "inferred": true,
+              "inferred_reason": "原文仅说'输入插画和提示词：3d效果即可生成'，实际提示词推断为简短的3D效果描述词，具体文本以原文为准",
+              "substance": "/表象/视觉/符号/文字符号/标注/提示引导",
+              "form": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "底图",
+              "name": "3D立体效果海报",
+              "value": "<3D立体效果版儿童节主题海报，在2D涂鸦插画基础上增加立体质感和3D渲染效果，人物和元素呈立体膨胀感>",
+              "anchor": "→ 返回 图像",
+              "substance": "/表象/视觉/实体/物品/影像物品/艺术创作",
+              "form": [
+                "/呈现/视觉/视觉制作/后期处理/图像合成处理/三维建模合成",
+                "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"
+              ]
+            }
+          ],
+          "intent": "用 {via:liblib_ai_智能图片v2}（3D模型）{act:图生图} 将 {in-type:图像}（2D涂鸦插画）转换为 {out-type:图像}（3D立体效果海报）",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ]
+        }
+      ],
+      "return_row": {
+        "arrow": "↩",
+        "text": "<span class=\"kw\">返回</span> <span class=\"chip\" data-type=\"图像\">图像</span>（3D立体效果海报）"
+      }
+    },
+    {
+      "id": "p3",
+      "name": "动态视频海报生成",
+      "purpose": "将静态涂鸦风插画海报通过 LibLib AI 内置 Seedance 2.0 视频生成功能，转为带动画效果的循环动态视频海报，增强节日氛围感",
+      "category": "产物创造",
+      "platform": "xhs",
+      "author": "一只布偶偶",
+      "declarations": {
+        "inputs": [
+          {
+            "type": "图像",
+            "name": "静态海报图",
+            "desc": "P1或P2生成的静态涂鸦风插画海报，作为视频生成的参考图"
+          }
+        ],
+        "resources": [],
+        "returns": {
+          "type": "视频"
+        }
+      },
+      "steps": [
+        {
+          "id": "s1",
+          "kind": "step",
+          "effect": "预处理",
+          "via": "human",
+          "action": "获取/录入/键入",
+          "feature": "幂等",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "在 LibLib 平台点击「视频生成」，在标签栏选择「图片」上传静态海报图，选择模型 Seedance 2.0，配置输出参数"
+            ],
+            [
+              "config",
+              "Seedance 2.0 VIP · 全能参考 · 比例：3:4 · 分辨率：1080p · 时长：6s · 有配音"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "参考图",
+              "name": "静态海报图",
+              "value": "<选定的静态涂鸦风插画海报（P1或P2产出）>",
+              "anchor": "← 工序输入",
+              "substance": "/表象/视觉/实体/物品/影像物品/艺术创作",
+              "form": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "工具配置",
+              "name": "视频生成配置",
+              "value": "Seedance 2.0 VIP · 全能参考 · 3:4 · 1080p · 6s · 有配音",
+              "anchor": "→ s2.视频生成配置",
+              "substance": null,
+              "form": null
+            }
+          ],
+          "intent": "用 {via:human} {act:配置视频生成} 选择 {in-type:图像}（静态海报）并设置 {out-type:工具配置}（Seedance 2.0视频参数）",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ]
+        },
+        {
+          "id": "s2",
+          "kind": "step",
+          "effect": "主体生成",
+          "via": "seedance_2_0",
+          "action": "生成/元素生成",
+          "feature": "随机",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "输入动态效果提示词，点击生成"
+            ],
+            [
+              "config",
+              "Seedance 2.0 VIP · 全能参考 · 3:4 · 1080p · 6s · 有配音"
+            ],
+            [
+              "memo",
+              "动态提示词策略：描述图中各元素的运动方式（弹跳/漂浮/旋转/碰撞），加入镜头运动（如'轻微拉远'），末尾加'整体具有循环动画感'确保循环流畅。适合节日氛围海报。生成后可去 LibTV 做更完善的创作剪辑。"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "参考图",
+              "name": "静态海报图",
+              "value": "<静态涂鸦风插画海报（橘味汽水主题，P1产出）>",
+              "anchor": "← 工序输入",
+              "substance": "/表象/视觉/实体/物品/影像物品/艺术创作",
+              "form": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"
+            },
+            {
+              "type": "提示词",
+              "name": "动态效果提示词",
+              "value": "瓶内夏日童年场景（蓝天、白云、彩虹、小卖部、冰棒、西瓜），卡通小孩从瓶中跳出挥舞橘子汽水，喷出大量透明泡泡；气泡持续上升，冰块碰撞晃动，橘子切片旋转，爆米花弹跳，贴纸与漫画元素漂浮；镜头轻微拉远镜，整体具有循环动画感",
+              "anchor": "← 字面量",
+              "substance": "/表象/视觉/符号/文字符号/标注/提示引导",
+              "form": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"
+            },
+            {
+              "type": "工具配置",
+              "name": "视频生成配置",
+              "value": "Seedance 2.0 VIP · 全能参考 · 3:4 · 1080p · 6s · 有配音",
+              "anchor": "← s1.视频生成配置",
+              "substance": null,
+              "form": null
+            }
+          ],
+          "outputs": [
+            {
+              "type": "视频成品",
+              "name": "动态视频海报",
+              "value": "<6秒循环动态视频海报，3:4竖版1080p，有配音，涂鸦风元素动态运动（气泡上升、冰块晃动、爆米花弹跳），适合节日氛围传播>",
+              "anchor": "→ 返回 视频",
+              "substance": "/表象/视觉/实体/物品/影像物品/艺术创作",
+              "form": [
+                "/呈现/视觉/视觉制作/后期处理/动态特效处理/运动变换",
+                "/呈现/视觉/视觉制作/后期处理/图像合成处理/AI生成合成"
+              ]
+            }
+          ],
+          "intent": "用 {via:seedance_2_0} {act:图生视频} 从 {in-type:图像}（静态海报）和 {in-type:提示词}（动态效果描述）生成 {out-type:视频}（动态视频海报）",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ]
+        }
+      ],
+      "return_row": {
+        "arrow": "↩",
+        "text": "<span class=\"kw\">返回</span> <span class=\"chip\" data-type=\"视频\">视频</span>（动态视频海报）"
+      },
+      "type_registry": {
+        "工具配置": {
+          "extends": "生成参数",
+          "desc": "工具与模型的综合配置项，含平台模型选型、参考模式、画幅规格、时长及配音等"
+        }
+      }
+    }
+  ]
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/.session_id

@@ -0,0 +1 @@
+b706877c-b568-414d-8166-ce8d6bf67502

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/_meta.json

@@ -0,0 +1,8 @@
+{
+  "case_id": "zhihu_2038237278795981638",
+  "from_q": "q0018",
+  "form": "A",
+  "score": 9.2,
+  "score_norm": 0.92,
+  "started_at": "2026-06-03T14:09:07"
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/_scratch/anchors.json

@@ -0,0 +1,63 @@
+[
+  {"path": "p1.s1.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p1.s1.outputs[0].anchor", "value": "→ s2"},
+  {"path": "p1.s2.inputs[0].anchor",  "value": "← s1.参数变量清单"},
+  {"path": "p1.s2.outputs[0].anchor", "value": "→ s3"},
+  {"path": "p1.s3.inputs[0].anchor",  "value": "← s2.参数化提示词模板"},
+  {"path": "p1.s3.outputs[0].anchor", "value": "→ 返回 商品主图"},
+
+  {"path": "p2.s1.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p2.s1.outputs[0].anchor", "value": "→ s2"},
+  {"path": "p2.s2.inputs[0].anchor",  "value": "← s1.JSON结构化提示词"},
+  {"path": "p2.s2.outputs[0].anchor", "value": "→ 返回 多区域广告图"},
+
+  {"path": "p3.s1.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p3.s1.outputs[0].anchor", "value": "→ s2"},
+  {"path": "p3.s2.inputs[0].anchor",  "value": "← s1.摄影要素清单"},
+  {"path": "p3.s2.outputs[0].anchor", "value": "→ s3"},
+  {"path": "p3.s3.inputs[0].anchor",  "value": "← s2.摄影指令提示词"},
+  {"path": "p3.s3.outputs[0].anchor", "value": "→ 返回 电影感人像照"},
+
+  {"path": "p4.s1.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p4.s1.outputs[0].anchor", "value": "→ s2"},
+  {"path": "p4.s2.inputs[0].anchor",  "value": "← s1.摄影要素清单"},
+  {"path": "p4.s2.outputs[0].anchor", "value": "→ s3"},
+  {"path": "p4.s3.inputs[0].anchor",  "value": "← s2.摄影指令提示词"},
+  {"path": "p4.s3.outputs[0].anchor", "value": "→ 返回 温泉旅馆人像"},
+
+  {"path": "p5.s1.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p5.s1.outputs[0].anchor", "value": "→ s2"},
+  {"path": "p5.s2.inputs[0].anchor",  "value": "← s1.基础构图提示词"},
+  {"path": "p5.s2.outputs[0].anchor", "value": "→ s3, s4"},
+  {"path": "p5.s3.inputs[0].anchor",  "value": "← s2.基础汉堡图"},
+  {"path": "p5.s3.inputs[1].anchor",  "value": "← 工序输入"},
+  {"path": "p5.s3.outputs[0].anchor", "value": "→ s4"},
+  {"path": "p5.s4.inputs[0].anchor",  "value": "← s3.叠加效果提示词"},
+  {"path": "p5.s4.inputs[1].anchor",  "value": "← s2.基础汉堡图"},
+  {"path": "p5.s4.outputs[0].anchor", "value": "→ 返回 精修商品图"},
+
+  {"path": "p6.s1.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p6.s1.outputs[0].anchor", "value": "→ s2"},
+  {"path": "p6.s2.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p6.s2.inputs[1].anchor",  "value": "← s1.转换指令提示词"},
+  {"path": "p6.s2.outputs[0].anchor", "value": "→ 返回 动漫风格图"},
+
+  {"path": "p7.s1.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p7.s1.outputs[0].anchor", "value": "→ s2"},
+  {"path": "p7.s2.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p7.s2.inputs[1].anchor",  "value": "← s1.角色设定提示词"},
+  {"path": "p7.s2.outputs[0].anchor", "value": "→ 返回 角色设定卡"},
+
+  {"path": "p8.s1.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p8.s1.outputs[0].anchor", "value": "→ s2"},
+  {"path": "p8.s2.inputs[0].anchor",  "value": "← s1.多语言混合提示词"},
+  {"path": "p8.s2.outputs[0].anchor", "value": "→ 返回 本地化内容图"},
+
+  {"path": "p9.s1.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p9.s1.outputs[0].anchor", "value": "→ s2"},
+  {"path": "p9.s2.inputs[0].anchor",  "value": "← s1.结构测试提示词"},
+  {"path": "p9.s2.outputs[0].anchor", "value": "→ s3"},
+  {"path": "p9.s3.inputs[0].anchor",  "value": "← s2.结构验证图"},
+  {"path": "p9.s3.inputs[1].anchor",  "value": "← 工序输入"},
+  {"path": "p9.s3.outputs[0].anchor", "value": "→ 返回 验证结果"}
+]

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/_scratch/patch_2a.json

@@ -0,0 +1,128 @@
+[
+  {"path": "p1.s1.effect", "value": "预处理"},
+  {"path": "p1.s1.action", "value": "提取/化学提取/解构"},
+  {"path": "p1.s1.inputs[0].type", "value": "描述"},
+  {"path": "p1.s1.outputs[0].type", "value": "参数列表"},
+  {"path": "p1.type_registry.参数列表.extends", "value": "描述"},
+  {"path": "p1.type_registry.参数列表.desc", "value": "从需求描述中提取出的结构化变量参数清单，列出可替换参数与固定元素，供后续提示词模板组合使用"},
+
+  {"path": "p1.s2.effect", "value": "预处理"},
+  {"path": "p1.s2.action", "value": "生成/关系生成/结构生成"},
+  {"path": "p1.s2.inputs[0].type", "value": "参数列表"},
+  {"path": "p1.s2.outputs[0].type", "value": "提示词"},
+
+  {"path": "p1.s3.effect", "value": "主体生成"},
+  {"path": "p1.s3.action", "value": "生成/元素生成"},
+  {"path": "p1.s3.inputs[0].type", "value": "提示词"},
+  {"path": "p1.s3.outputs[0].type", "value": "成品图"},
+
+  {"path": "p2.s1.effect", "value": "预处理"},
+  {"path": "p2.s1.action", "value": "生成/关系生成/结构生成"},
+  {"path": "p2.s1.inputs[0].type", "value": "描述"},
+  {"path": "p2.s1.outputs[0].type", "value": "提示词"},
+
+  {"path": "p2.s2.effect", "value": "主体生成"},
+  {"path": "p2.s2.action", "value": "生成/元素生成"},
+  {"path": "p2.s2.inputs[0].type", "value": "提示词"},
+  {"path": "p2.s2.outputs[0].type", "value": "成品图"},
+
+  {"path": "p3.s1.effect", "value": "预处理"},
+  {"path": "p3.s1.action", "value": "提取/化学提取/解构"},
+  {"path": "p3.s1.inputs[0].type", "value": "描述"},
+  {"path": "p3.s1.outputs[0].type", "value": "提示词"},
+
+  {"path": "p3.s2.effect", "value": "预处理"},
+  {"path": "p3.s2.action", "value": "生成/关系生成/结构生成"},
+  {"path": "p3.s2.inputs[0].type", "value": "提示词"},
+  {"path": "p3.s2.outputs[0].type", "value": "提示词"},
+
+  {"path": "p3.s3.effect", "value": "主体生成"},
+  {"path": "p3.s3.action", "value": "生成/元素生成"},
+  {"path": "p3.s3.inputs[0].type", "value": "提示词"},
+  {"path": "p3.s3.outputs[0].type", "value": "成品图"},
+
+  {"path": "p4.s1.effect", "value": "预处理"},
+  {"path": "p4.s1.action", "value": "提取/化学提取/解构"},
+  {"path": "p4.s1.inputs[0].type", "value": "描述"},
+  {"path": "p4.s1.outputs[0].type", "value": "提示词"},
+
+  {"path": "p4.s2.effect", "value": "预处理"},
+  {"path": "p4.s2.action", "value": "生成/关系生成/结构生成"},
+  {"path": "p4.s2.inputs[0].type", "value": "提示词"},
+  {"path": "p4.s2.outputs[0].type", "value": "提示词"},
+
+  {"path": "p4.s3.effect", "value": "主体生成"},
+  {"path": "p4.s3.action", "value": "生成/元素生成"},
+  {"path": "p4.s3.inputs[0].type", "value": "提示词"},
+  {"path": "p4.s3.outputs[0].type", "value": "成品图"},
+
+  {"path": "p5.s1.effect", "value": "预处理"},
+  {"path": "p5.s1.action", "value": "生成/关系生成/结构生成"},
+  {"path": "p5.s1.inputs[0].type", "value": "描述"},
+  {"path": "p5.s1.outputs[0].type", "value": "提示词"},
+
+  {"path": "p5.s2.effect", "value": "主体生成"},
+  {"path": "p5.s2.action", "value": "生成/元素生成"},
+  {"path": "p5.s2.inputs[0].type", "value": "提示词"},
+  {"path": "p5.s2.outputs[0].type", "value": "底图"},
+
+  {"path": "p5.s3.effect", "value": "预处理"},
+  {"path": "p5.s3.action", "value": "生成/关系生成/结构生成"},
+  {"path": "p5.s3.inputs[0].type", "value": "底图"},
+  {"path": "p5.s3.inputs[1].type", "value": "描述"},
+  {"path": "p5.s3.outputs[0].type", "value": "提示词"},
+
+  {"path": "p5.s4.effect", "value": "主体生成"},
+  {"path": "p5.s4.action", "value": "生成/元素生成"},
+  {"path": "p5.s4.inputs[0].type", "value": "提示词"},
+  {"path": "p5.s4.inputs[1].type", "value": "底图"},
+  {"path": "p5.s4.outputs[0].type", "value": "成品图"},
+
+  {"path": "p6.s1.effect", "value": "预处理"},
+  {"path": "p6.s1.action", "value": "生成/关系生成/结构生成"},
+  {"path": "p6.s1.inputs[0].type", "value": "描述"},
+  {"path": "p6.s1.outputs[0].type", "value": "提示词"},
+
+  {"path": "p6.s2.effect", "value": "主体生成"},
+  {"path": "p6.s2.action", "value": "生成/元素生成"},
+  {"path": "p6.s2.inputs[0].type", "value": "参考图"},
+  {"path": "p6.s2.inputs[1].type", "value": "提示词"},
+  {"path": "p6.s2.outputs[0].type", "value": "成品图"},
+
+  {"path": "p7.s1.effect", "value": "预处理"},
+  {"path": "p7.s1.action", "value": "生成/关系生成/结构生成"},
+  {"path": "p7.s1.inputs[0].type", "value": "描述"},
+  {"path": "p7.s1.outputs[0].type", "value": "提示词"},
+
+  {"path": "p7.s2.effect", "value": "主体生成"},
+  {"path": "p7.s2.action", "value": "生成/元素生成"},
+  {"path": "p7.s2.inputs[0].type", "value": "参考图"},
+  {"path": "p7.s2.inputs[1].type", "value": "提示词"},
+  {"path": "p7.s2.outputs[0].type", "value": "成品图"},
+
+  {"path": "p8.s1.effect", "value": "预处理"},
+  {"path": "p8.s1.action", "value": "生成/关系生成/结构生成"},
+  {"path": "p8.s1.inputs[0].type", "value": "描述"},
+  {"path": "p8.s1.outputs[0].type", "value": "提示词"},
+
+  {"path": "p8.s2.effect", "value": "主体生成"},
+  {"path": "p8.s2.action", "value": "生成/元素生成"},
+  {"path": "p8.s2.inputs[0].type", "value": "提示词"},
+  {"path": "p8.s2.outputs[0].type", "value": "成品图"},
+
+  {"path": "p9.s1.effect", "value": "预处理"},
+  {"path": "p9.s1.action", "value": "生成/关系生成/结构生成"},
+  {"path": "p9.s1.inputs[0].type", "value": "描述"},
+  {"path": "p9.s1.outputs[0].type", "value": "提示词"},
+
+  {"path": "p9.s2.effect", "value": "主体生成"},
+  {"path": "p9.s2.action", "value": "生成/元素生成"},
+  {"path": "p9.s2.inputs[0].type", "value": "提示词"},
+  {"path": "p9.s2.outputs[0].type", "value": "样图"},
+
+  {"path": "p9.s3.effect", "value": "检验"},
+  {"path": "p9.s3.action", "value": "提取/化学提取/识别"},
+  {"path": "p9.s3.inputs[0].type", "value": "样图"},
+  {"path": "p9.s3.inputs[1].type", "value": "描述"},
+  {"path": "p9.s3.outputs[0].type", "value": "评语"}
+]

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/_scratch/patch_2b.json

@@ -0,0 +1,160 @@
+[
+  {"path": "p1.s1.inputs[0].substance", "value": "/表象/视觉/实体/物品/起居物品/美妆护肤"},
+  {"path": "p1.s1.inputs[0].form", "value": "/架构/修辞/语调风格/平实表达/直白陈述"},
+
+  {"path": "p1.s1.outputs[0].substance", "value": "/理念/知识/商业/产品服务/产品特征/规格参数"},
+  {"path": "p1.s1.outputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/逐条罗列/通用列举"},
+
+  {"path": "p1.s2.inputs[0].substance", "value": "/理念/知识/商业/产品服务/产品特征/规格参数"},
+  {"path": "p1.s2.inputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/逐条罗列/通用列举"},
+
+  {"path": "p1.s2.outputs[0].substance", "value": "/表象/视觉/实体/物品/起居物品/美妆护肤"},
+  {"path": "p1.s2.outputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+
+  {"path": "p1.s3.inputs[0].substance", "value": "/表象/视觉/实体/物品/起居物品/美妆护肤"},
+  {"path": "p1.s3.inputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+
+  {"path": "p1.s3.outputs[0].substance", "value": "/表象/视觉/实体/物品/起居物品/美妆护肤"},
+  {"path": "p1.s3.outputs[0].form", "value": "/呈现/视觉/视觉制作/实景拍摄"},
+
+  {"path": "p2.s1.inputs[0].substance", "value": "/理念/事件/商业事件/推广活动"},
+  {"path": "p2.s1.inputs[0].form", "value": "/架构/修辞/语调风格/平实表达/直白陈述"},
+
+  {"path": "p2.s1.outputs[0].substance", "value": "/理念/事件/商业事件/推广活动"},
+  {"path": "p2.s1.outputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+
+  {"path": "p2.s2.inputs[0].substance", "value": "/理念/事件/商业事件/推广活动"},
+  {"path": "p2.s2.inputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+
+  {"path": "p2.s2.outputs[0].substance", "value": "/理念/事件/商业事件/推广活动"},
+  {"path": "p2.s2.outputs[0].form", "value": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"},
+
+  {"path": "p3.s1.inputs[0].substance", "value": "/表象/视觉/空间/空间场所/生活场景/商业场所"},
+  {"path": "p3.s1.inputs[0].form", "value": "/架构/修辞/语调风格/平实表达/直白陈述"},
+
+  {"path": "p3.s1.outputs[0].substance", "value": "/表象/视觉/实体/物品/影像物品/艺术创作"},
+  {"path": "p3.s1.outputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/逐条罗列/通用列举"},
+
+  {"path": "p3.s2.inputs[0].substance", "value": "/表象/视觉/实体/物品/影像物品/艺术创作"},
+  {"path": "p3.s2.inputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/逐条罗列/通用列举"},
+
+  {"path": "p3.s2.outputs[0].substance", "value": "/表象/视觉/实体/人物/人物特征/形象呈现"},
+  {"path": "p3.s2.outputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+
+  {"path": "p3.s3.inputs[0].substance", "value": "/表象/视觉/实体/人物/人物特征/形象呈现"},
+  {"path": "p3.s3.inputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+
+  {"path": "p3.s3.outputs[0].substance", "value": "/表象/视觉/实体/人物/人物特征/形象呈现"},
+  {"path": "p3.s3.outputs[0].form", "value": "/呈现/视觉/视觉风格/视觉艺术风格/影视风格/胶片摄影"},
+
+  {"path": "p4.s1.inputs[0].substance", "value": "/表象/视觉/空间/空间场所"},
+  {"path": "p4.s1.inputs[0].form", "value": "/架构/修辞/语调风格/平实表达/直白陈述"},
+
+  {"path": "p4.s1.outputs[0].substance", "value": "/表象/视觉/实体/物品/影像物品/艺术创作"},
+  {"path": "p4.s1.outputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/逐条罗列/通用列举"},
+
+  {"path": "p4.s2.inputs[0].substance", "value": "/表象/视觉/实体/物品/影像物品/艺术创作"},
+  {"path": "p4.s2.inputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/逐条罗列/通用列举"},
+
+  {"path": "p4.s2.outputs[0].substance", "value": "/表象/视觉/实体/人物/人物特征/形象呈现"},
+  {"path": "p4.s2.outputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+
+  {"path": "p4.s3.inputs[0].substance", "value": "/表象/视觉/实体/人物/人物特征/形象呈现"},
+  {"path": "p4.s3.inputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+
+  {"path": "p4.s3.outputs[0].substance", "value": "/表象/视觉/实体/人物/人物特征/形象呈现"},
+  {"path": "p4.s3.outputs[0].form", "value": "/呈现/视觉/视觉风格/视觉艺术风格/影视风格/胶片摄影"},
+
+  {"path": "p5.s1.inputs[0].substance", "value": "/表象/视觉/实体/物品/起居物品/食物"},
+  {"path": "p5.s1.inputs[0].form", "value": "/架构/修辞/语调风格/平实表达/直白陈述"},
+
+  {"path": "p5.s1.outputs[0].substance", "value": "/表象/视觉/实体/物品/起居物品/食物"},
+  {"path": "p5.s1.outputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+
+  {"path": "p5.s2.inputs[0].substance", "value": "/表象/视觉/实体/物品/起居物品/食物"},
+  {"path": "p5.s2.inputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+
+  {"path": "p5.s2.outputs[0].substance", "value": "/表象/视觉/实体/物品/起居物品/食物"},
+  {"path": "p5.s2.outputs[0].form", "value": "/呈现/视觉/视觉制作/实景拍摄"},
+
+  {"path": "p5.s3.inputs[0].substance", "value": "/表象/视觉/实体/物品/起居物品/食物"},
+  {"path": "p5.s3.inputs[0].form", "value": "/呈现/视觉/视觉制作/实景拍摄"},
+
+  {"path": "p5.s3.inputs[1].substance", "value": "/表象/视觉/符号/文字符号/水印"},
+  {"path": "p5.s3.inputs[1].form", "value": "/架构/修辞/语调风格/平实表达/直白陈述"},
+
+  {"path": "p5.s3.outputs[0].substance", "value": "/表象/视觉/实体/物品/起居物品/食物"},
+  {"path": "p5.s3.outputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+
+  {"path": "p5.s4.inputs[0].substance", "value": "/表象/视觉/实体/物品/起居物品/食物"},
+  {"path": "p5.s4.inputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+
+  {"path": "p5.s4.inputs[1].substance", "value": "/表象/视觉/实体/物品/起居物品/食物"},
+  {"path": "p5.s4.inputs[1].form", "value": "/呈现/视觉/视觉制作/实景拍摄"},
+
+  {"path": "p5.s4.outputs[0].substance", "value": "/表象/视觉/实体/物品/起居物品/食物"},
+  {"path": "p5.s4.outputs[0].form", "value": "/呈现/视觉/视觉制作/实景拍摄"},
+
+  {"path": "p6.s1.inputs[0].substance", "value": "/表象/视觉/实体/人物/人物特征/形象呈现"},
+  {"path": "p6.s1.inputs[0].form", "value": "/架构/修辞/语调风格/平实表达/直白陈述"},
+
+  {"path": "p6.s1.outputs[0].substance", "value": "/表象/视觉/实体/人物/人物特征/形象呈现"},
+  {"path": "p6.s1.outputs[0].form", "value": "/架构/修辞/语调风格/平实表达/直白陈述"},
+
+  {"path": "p6.s2.inputs[0].substance", "value": "/表象/视觉/实体/人物/人物特征/形象呈现"},
+  {"path": "p6.s2.inputs[0].form", "value": "/呈现/视觉/视觉制作/实景拍摄"},
+
+  {"path": "p6.s2.inputs[1].substance", "value": "/表象/视觉/实体/人物/人物特征/形象呈现"},
+  {"path": "p6.s2.inputs[1].form", "value": "/架构/修辞/语调风格/平实表达/直白陈述"},
+
+  {"path": "p6.s2.outputs[0].substance", "value": "/表象/视觉/实体/人物/人物特征/形象呈现"},
+  {"path": "p6.s2.outputs[0].form", "value": "/呈现/视觉/视觉风格/视觉艺术风格/绘画风格/卡通"},
+
+  {"path": "p7.s1.inputs[0].substance", "value": "/表象/视觉/实体/人物/人物特征/形象呈现"},
+  {"path": "p7.s1.inputs[0].form", "value": "/架构/修辞/语调风格/平实表达/直白陈述"},
+
+  {"path": "p7.s1.outputs[0].substance", "value": "/表象/视觉/实体/人物/人物特征/形象呈现"},
+  {"path": "p7.s1.outputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+
+  {"path": "p7.s2.inputs[0].substance", "value": "/表象/视觉/实体/人物/个体/虚构人物"},
+  {"path": "p7.s2.inputs[0].form", "value": "/呈现/视觉/视觉风格/视觉艺术风格/绘画风格/卡通"},
+
+  {"path": "p7.s2.inputs[1].substance", "value": "/表象/视觉/实体/人物/人物特征/形象呈现"},
+  {"path": "p7.s2.inputs[1].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+
+  {"path": "p7.s2.outputs[0].substance", "value": "/表象/视觉/实体/人物/人物特征/形象呈现"},
+  {"path": "p7.s2.outputs[0].form", "value": "/呈现/视觉/视觉风格/视觉艺术风格/绘画风格/卡通"},
+
+  {"path": "p8.s1.inputs[0].substance", "value": ["/表象/视觉/实体/物品/起居物品/食物", "/表象/视觉/实体/物品/器物物品/道具"]},
+  {"path": "p8.s1.inputs[0].form", "value": "/架构/修辞/语调风格/平实表达/直白陈述"},
+
+  {"path": "p8.s1.outputs[0].substance", "value": ["/表象/视觉/实体/物品/起居物品/食物", "/表象/视觉/实体/物品/器物物品/道具"]},
+  {"path": "p8.s1.outputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+
+  {"path": "p8.s2.inputs[0].substance", "value": ["/表象/视觉/实体/物品/起居物品/食物", "/表象/视觉/实体/物品/器物物品/道具"]},
+  {"path": "p8.s2.inputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+
+  {"path": "p8.s2.outputs[0].substance", "value": ["/表象/视觉/实体/物品/起居物品/食物", "/表象/视觉/实体/物品/器物物品/道具"]},
+  {"path": "p8.s2.outputs[0].form", "value": "/呈现/视觉/视觉风格/视觉艺术风格/绘画风格/插画"},
+
+  {"path": "p9.s1.inputs[0].substance", "value": "/表象/视觉/实体/物品/起居物品/家居"},
+  {"path": "p9.s1.inputs[0].form", "value": "/架构/修辞/语调风格/平实表达/直白陈述"},
+
+  {"path": "p9.s1.outputs[0].substance", "value": "/表象/视觉/实体/物品/起居物品/家居"},
+  {"path": "p9.s1.outputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+
+  {"path": "p9.s2.inputs[0].substance", "value": "/表象/视觉/实体/物品/起居物品/家居"},
+  {"path": "p9.s2.inputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+
+  {"path": "p9.s2.outputs[0].substance", "value": "/表象/视觉/实体/物品/起居物品/家居"},
+  {"path": "p9.s2.outputs[0].form", "value": "/呈现/视觉/视觉制作/实景拍摄"},
+
+  {"path": "p9.s3.inputs[0].substance", "value": "/表象/视觉/实体/物品/起居物品/家居"},
+  {"path": "p9.s3.inputs[0].form", "value": "/呈现/视觉/视觉制作/实景拍摄"},
+
+  {"path": "p9.s3.inputs[1].substance", "value": "/表象/视觉/实体/物品/起居物品/家居"},
+  {"path": "p9.s3.inputs[1].form", "value": "/架构/修辞/语调风格/平实表达/直白陈述"},
+
+  {"path": "p9.s3.outputs[0].substance", "value": "/理念/知识/商业/前沿技术/AI智能/AI应用"},
+  {"path": "p9.s3.outputs[0].form", "value": "/架构/逻辑/信息编排逻辑/归纳论证/归纳提炼/概括浓缩"}
+]

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/_scratch/patch_intent.json

@@ -0,0 +1,34 @@
+[
+  {"path": "p1.s1.intent", "value": "用 {via:human} {act:解构} 从 {in-type:描述} 梳理出可参数化的 {out-type:参数列表}"},
+  {"path": "p1.s2.intent", "value": "用 {via:human} {act:结构生成} 将 {in-type:参数列表} 嵌入固定风格骨架，生成可复用的 {out-type:提示词}"},
+  {"path": "p1.s3.intent", "value": "用 {via:AI生图工具} {act:元素生成} 从 {in-type:提示词} 生成 {out-type:成品图}"},
+
+  {"path": "p2.s1.intent", "value": "用 {via:human} {act:结构生成} 从 {in-type:描述} 编写精确的 {out-type:提示词}"},
+  {"path": "p2.s2.intent", "value": "用 {via:AI生图工具} {act:元素生成} 将 {in-type:提示词} 渲染为精确布局的 {out-type:成品图}"},
+
+  {"path": "p3.s1.intent", "value": "用 {via:human} {act:解构} 从 {in-type:描述} 提炼五要素得到 {out-type:提示词}"},
+  {"path": "p3.s2.intent", "value": "用 {via:human} {act:结构生成} 将 {in-type:提示词} 拼接为完整摄影指令 {out-type:提示词}"},
+  {"path": "p3.s3.intent", "value": "用 {via:AI生图工具} {act:元素生成} 从 {in-type:提示词} 生成 {out-type:成品图}"},
+
+  {"path": "p4.s1.intent", "value": "用 {via:human} {act:解构} 从 {in-type:描述} 提炼温泉场景五要素 {out-type:提示词}"},
+  {"path": "p4.s2.intent", "value": "用 {via:human} {act:结构生成} 将温泉场景 {in-type:提示词} 拼接为完整 {out-type:提示词}"},
+  {"path": "p4.s3.intent", "value": "用 {via:AI生图工具} {act:元素生成} 从 {in-type:提示词} 生成 {out-type:成品图}"},
+
+  {"path": "p5.s1.intent", "value": "用 {via:human} {act:结构生成} 从 {in-type:描述} 撰写聚焦主体构图的 {out-type:提示词}"},
+  {"path": "p5.s2.intent", "value": "用 {via:AI生图工具} {act:元素生成} 从 {in-type:提示词} 生成主体 {out-type:底图}"},
+  {"path": "p5.s3.intent", "value": "用 {via:human} {act:结构生成} 从 {in-type:底图} 和 {in-type:描述} 撰写 {out-type:提示词}"},
+  {"path": "p5.s4.intent", "value": "用 {via:AI生图工具} {act:元素生成} 按 {in-type:提示词} 对 {in-type:底图} 叠加细节生成 {out-type:成品图}"},
+
+  {"path": "p6.s1.intent", "value": "用 {via:human} {act:结构生成} 从 {in-type:描述} 撰写极简 {out-type:提示词}"},
+  {"path": "p6.s2.intent", "value": "用 {via:AI生图工具} {act:元素生成} 将 {in-type:参考图} 按 {in-type:提示词} 转换为 {out-type:成品图}"},
+
+  {"path": "p7.s1.intent", "value": "用 {via:human} {act:结构生成} 从 {in-type:描述} 撰写日文 {out-type:提示词}"},
+  {"path": "p7.s2.intent", "value": "用 {via:AI生图工具} {act:元素生成} 从 {in-type:参考图} 和 {in-type:提示词} 生成 {out-type:成品图}"},
+
+  {"path": "p8.s1.intent", "value": "用 {via:human} {act:结构生成} 从 {in-type:描述} 构建中英协同的 {out-type:提示词}"},
+  {"path": "p8.s2.intent", "value": "用 {via:AI生图工具} {act:元素生成} 将 {in-type:提示词} 渲染为精细的 {out-type:成品图}"},
+
+  {"path": "p9.s1.intent", "value": "用 {via:human} {act:结构生成} 从 {in-type:描述} 撰写精确计数的 {out-type:提示词}"},
+  {"path": "p9.s2.intent", "value": "用 {via:AI生图工具} {act:元素生成} 从 {in-type:提示词} 生成 {out-type:样图}"},
+  {"path": "p9.s3.intent", "value": "用 {via:human} {act:识别} 核验 {in-type:样图} 计数准确性，输出 {out-type:评语}"}
+]

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/_scratch/task_2a.json

@@ -0,0 +1,666 @@
+{
+  "case_id": "A_zhihu_20382372",
+  "task": "Phase 2A (effect/action/type normalization)",
+  "image_url_list": [],
+  "steps": [
+    {
+      "path": "p1.s1",
+      "name": "",
+      "action": "提取/信息提取/解构",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p1.s1.inputs[0]",
+          "name": "产品品类描述",
+          "value": "电商香水瓶产品主图，需要生成一套可替换品牌名和装饰花材的系列图，保持luxurious cinematic风格不变",
+          "type": "文本",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p1.s1.outputs[0]",
+          "name": "参数变量清单",
+          "value": "参数变量：brand label（默认 N°5 CHANEL PARIS PARFUM）、flowers（默认 chamomile blossoms）；固定元素：glossy black marble surface with white veining / cinematic product photograph / dramatic side lighting / ultra-detailed glass texture / photorealistic 8K",
+          "type": "参数列表",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p1.s2",
+      "name": "",
+      "action": "生成/关系生成/组合",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p1.s2.inputs[0]",
+          "name": "参数变量清单",
+          "value": "参数变量：brand label（默认 N°5 CHANEL PARIS PARFUM）、flowers（默认 chamomile blossoms）；固定元素：glossy black marble surface with white veining / cinematic product photograph / dramatic side lighting / ultra-detailed glass texture / photorealistic 8K",
+          "type": "参数列表",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p1.s2.outputs[0]",
+          "name": "参数化提示词模板",
+          "value": "A luxurious cinematic product photograph of a classic rectangular perfume bottle inspired by {argument name=\"brand label\" default=\"N°5 CHANEL PARIS PARFUM\"}, placed upright on a glossy black marble surface with white veining, surrounded by scattered {argument name=\"flowers\" default=\"chamomile blossoms\"}, dramatic side lighting creating specular highlights on the glass, ultra-detailed glass texture, photorealistic rendering, 8K, luxury advertising style",
+          "type": "提示词",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p1.s3",
+      "name": "",
+      "action": "生成/图像生成/文生图",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p1.s3.inputs[0]",
+          "name": "参数化提示词模板",
+          "value": "A luxurious cinematic product photograph of a classic rectangular perfume bottle inspired by {argument name=\"brand label\" default=\"N°5 CHANEL PARIS PARFUM\"}, placed upright on a glossy black marble surface with white veining, surrounded by scattered {argument name=\"flowers\" default=\"chamomile blossoms\"}, dramatic side lighting creating specular highlights on the glass, ultra-detailed glass texture, photorealistic rendering, 8K, luxury advertising style",
+          "type": "提示词",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p1.s3.outputs[0]",
+          "name": "商品主图",
+          "value": "<AI生成的香水瓶商品主图：Jimmy Choo香水瓶，深色渐变玻璃瓶身+金色JC标志，置于黑色大理石台面，金属花瓶和水晶碗陪衬，高端奢华广告风格，Pollo.ai水印>",
+          "type": "图像",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p2.s1",
+      "name": "",
+      "action": "生成/关系生成/组合",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p2.s1.inputs[0]",
+          "name": "布局需求描述",
+          "value": "4格日系广告Banner（2x2网格），各象限主题分别为：旅行（冲绳旅行3日 39,800日元）、护肤（透明感水光肌 初回限定78%OFF 1,980日元）、美食（黑毛和牛A4/A5豪华牛排 期间限定4,980日元）、学习（在线资格讲座 10万人受讲 20%OFF）",
+          "type": "文本",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p2.s1.outputs[0]",
+          "name": "JSON结构化提示词",
+          "value": "{\n  \"type\": \"2x2 grid of Japanese digital advertisement banners\",\n  \"layout\": {\n    \"structure\": \"4 equal quadrants\",\n    \"quadrants\": [\n      {\n        \"position\": \"top-left\",\n        \"theme\": \"Travel\",\n        \"subject\": \"A couple holding hands on a white sand beach with turquoise water in Okinawa, bright tropical sky\",\n        \"text_labels\": [\"今年こそ、解き放て。\", \"沖縄旅行\", \"3日間の癒やし旅\", \"39,800円〜\", \"航空券+ホテル\", \"往復航空券\", \"人気ホテル宿泊\", \"レンタカー付き\", \"絶景、グルメ、体験ぜんぶ叶う！\"]\n      },\n      {\n        \"position\": \"top-right\",\n        \"theme\": \"Skincare\",\n        \"subject\": \"Close-up of a young woman's face applying glowing skincare cream, skin radiant and clear\",\n        \"text_labels\": [\"毛穴・くすみ卒業!\", \"透明感あふれる水光肌へ\", \"新感覚スキンケア\", \"毛穴ケア\", \"高保湿\", \"ハリ・ツヤ\", \"初回限定 78%OFF\", \"1,980円\"]\n      },\n      {\n        \"position\": \"bottom-left\",\n        \"theme\": \"Food\",\n        \"subject\": \"Premium A5 Wagyu beef steak sizzling on a grill, close-up, steam and fat marbling visible\",\n        \"text_labels\": [\"とろける旨さ!\", \"黒毛和牛 A4 A5等級 贅沢ステーキ\", \"期間限定 特別価格\", \"通価格8,980円 4,980円\"]\n      },\n      {\n        \"position\": \"bottom-right\",\n        \"theme\": \"Education\",\n        \"subject\": \"Young man studying efficiently with smartphone at a bright modern desk\",\n        \"text_labels\": [\"スキマ時間で最短合格!\", \"オンライン資格講座\", \"スマホで完結\", \"効率学習で差がつく!\", \"受講者数10万人突破!\", \"今だけ! 受講料20%OFF\"]\n      }\n    ]\n  }\n}",
+          "type": "提示词",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p2.s2",
+      "name": "",
+      "action": "生成/图像生成/文生图",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p2.s2.inputs[0]",
+          "name": "JSON结构化提示词",
+          "value": "{\n  \"type\": \"2x2 grid of Japanese digital advertisement banners\",\n  \"layout\": {\n    \"structure\": \"4 equal quadrants\",\n    \"quadrants\": [\n      {\"position\": \"top-left\", \"theme\": \"Travel\", \"subject\": \"A couple holding hands on a white sand beach with turquoise water in Okinawa, bright tropical sky\", \"text_labels\": [\"今年こそ、解き放て。\", \"沖縄旅行\", \"3日間の癒やし旅\", \"39,800円〜\", \"航空券+ホテル\", \"往復航空券\", \"人気ホテル宿泊\", \"レンタカー付き\", \"絶景、グルメ、体験ぜんぶ叶う！\"]},\n      {\"position\": \"top-right\", \"theme\": \"Skincare\", \"subject\": \"Close-up of a young woman's face applying glowing skincare cream, skin radiant and clear\", \"text_labels\": [\"毛穴・くすみ卒業!\", \"透明感あふれる水光肌へ\", \"初回限定 78%OFF\", \"1,980円\"]},\n      {\"position\": \"bottom-left\", \"theme\": \"Food\", \"subject\": \"Premium A5 Wagyu beef steak sizzling on a grill, close-up, steam and fat marbling visible\", \"text_labels\": [\"黒毛和牛 A4 A5等級 贅沢ステーキ\", \"期間限定 特別価格\", \"4,980円\"]},\n      {\"position\": \"bottom-right\", \"theme\": \"Education\", \"subject\": \"Young man studying efficiently with smartphone at a bright modern desk\", \"text_labels\": [\"スキマ時間で最短合格!\", \"受講者数10万人突破!\", \"受講料20%OFF\"]}\n    ]\n  }\n}",
+          "type": "提示词",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p2.s2.outputs[0]",
+          "name": "多区域广告图",
+          "value": "<AI生成的4格日系广告Banner：左上冲绳旅行（海滩情侣背影+日文旅行信息+价格39,800円），右上护肤品（水光肌特写+品牌LUMIÈRE+折扣信息1,980円），左下黑毛和牛牛排（特写冒烟+价格4,980円），右下在线资格讲座（学习男生+10万人突破+20%OFF）>",
+          "type": "图像",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p3.s1",
+      "name": "",
+      "action": "提取/信息提取/解构",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p3.s1.inputs[0]",
+          "name": "拍摄场景描述",
+          "value": "便利店夜晚场景，女性中景，电影感街拍风格，有霓虹灯光效，胶片质感",
+          "type": "文本",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p3.s1.outputs[0]",
+          "name": "摄影要素清单",
+          "value": "胶片格式: 35mm film photography / 光线: harsh convenience store fluorescent lighting mixed with colorful neon signs from outside / 质感: authentic film grain, high contrast, slight color cast / 构图景别: intimate medium shot / 后期风格: cinematic street editorial style",
+          "type": "提示词",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p3.s2",
+      "name": "",
+      "action": "生成/关系生成/组合",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p3.s2.inputs[0]",
+          "name": "摄影要素清单",
+          "value": "胶片格式: 35mm film photography / 光线: harsh convenience store fluorescent lighting mixed with colorful neon signs from outside / 质感: authentic film grain, high contrast, slight color cast / 构图景别: intimate medium shot / 后期风格: cinematic street editorial style",
+          "type": "提示词",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p3.s2.outputs[0]",
+          "name": "摄影指令提示词",
+          "value": "35mm film photography with harsh convenience store fluorescent lighting mixed with colorful neon signs from outside, authentic film grain, high contrast, slight color cast, cinematic street editorial style, intimate medium shot, young Asian woman in casual white button-down shirt and black mini skirt, leaning against glass refrigerator door, holding a drink bottle, messy bun hairstyle, natural makeup, photorealistic",
+          "type": "提示词",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p3.s3",
+      "name": "",
+      "action": "生成/图像生成/文生图",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p3.s3.inputs[0]",
+          "name": "摄影指令提示词",
+          "value": "35mm film photography with harsh convenience store fluorescent lighting mixed with colorful neon signs from outside, authentic film grain, high contrast, slight color cast, cinematic street editorial style, intimate medium shot, young Asian woman in casual white button-down shirt and black mini skirt, leaning against glass refrigerator door, holding a drink bottle, messy bun hairstyle, natural makeup, photorealistic",
+          "type": "提示词",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p3.s3.outputs[0]",
+          "name": "电影感人像照",
+          "value": "<AI生成的便利店霓虹人像：年轻亚裔女性，白色半透明衬衫+黑色短裙，倚着便利店冰柜玻璃门，手持棕色饮料瓶，荧光灯+霓虹灯混合光效，35mm胶片颗粒感，高对比度电影街拍风格>",
+          "type": "图像",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p4.s1",
+      "name": "",
+      "action": "提取/信息提取/解构",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p4.s1.inputs[0]",
+          "name": "拍摄场景描述",
+          "value": "日式温泉旅馆（onsen ryokan）人像，暖色木质灯笼光源，editorial高大气感风格，胶片质感，女性",
+          "type": "文本",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p4.s1.outputs[0]",
+          "name": "摄影要素清单",
+          "value": "胶片格式: 35mm film photography / 光线: warm vintage Japanese onsen ryokan aesthetic, soft ambient wooden lantern lighting mixed with gentle natural window light / 质感: subtle film grain, gentle color shift / 构图景别: high atmosphere editorial style / 后期风格: vintage Japanese aesthetic",
+          "type": "提示词",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p4.s2",
+      "name": "",
+      "action": "生成/关系生成/组合",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p4.s2.inputs[0]",
+          "name": "摄影要素清单",
+          "value": "胶片格式: 35mm film photography / 光线: warm vintage Japanese onsen ryokan aesthetic, soft ambient wooden lantern lighting mixed with gentle natural window light / 质感: subtle film grain, gentle color shift / 构图景别: high atmosphere editorial style / 后期风格: vintage Japanese aesthetic",
+          "type": "提示词",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p4.s2.outputs[0]",
+          "name": "摄影指令提示词",
+          "value": "35mm film photography, warm vintage Japanese onsen ryokan aesthetic, soft ambient wooden lantern lighting mixed with gentle natural window light, subtle film grain, gentle color shift, high atmosphere editorial style, young Asian woman in traditional Japanese yukata robe, sitting gracefully on wooden veranda, hot spring visible in background, photorealistic",
+          "type": "提示词",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p4.s3",
+      "name": "",
+      "action": "生成/图像生成/文生图",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p4.s3.inputs[0]",
+          "name": "摄影指令提示词",
+          "value": "35mm film photography, warm vintage Japanese onsen ryokan aesthetic, soft ambient wooden lantern lighting mixed with gentle natural window light, subtle film grain, gentle color shift, high atmosphere editorial style, young Asian woman in traditional Japanese yukata robe, sitting gracefully on wooden veranda, hot spring visible in background, photorealistic",
+          "type": "提示词",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p4.s3.outputs[0]",
+          "name": "温泉旅馆人像",
+          "value": "<AI生成的温泉旅馆editorial人像：年轻亚裔女性，浅色宽松和风浴衣，坐在木质台阶上，背景可见温泉热水，室内木质灯笼暖光，胶片颗粒感，高大气感editorial风格>",
+          "type": "图像",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p5.s1",
+      "name": "",
+      "action": "生成/关系生成/组合",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p5.s1.inputs[0]",
+          "name": "商品描述",
+          "value": "汉堡产品图，需要精确放置品牌标识、增强光影效果、添加背景虚化",
+          "type": "文本",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p5.s1.outputs[0]",
+          "name": "基础构图提示词",
+          "value": "<基础构图提示词：product photography of a gourmet burger, centered composition, layers clearly visible (bun / lettuce / tomato / beef patty / cheese), studio lighting from top-left, white/neutral background, sharp focus on burger>",
+          "type": "提示词",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p5.s2",
+      "name": "",
+      "action": "生成/图像生成/文生图",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p5.s2.inputs[0]",
+          "name": "基础构图提示词",
+          "value": "<基础构图提示词：product photography of a gourmet burger, centered composition, layers clearly visible (bun / lettuce / tomato / beef patty / cheese), studio lighting from top-left, white/neutral background, sharp focus on burger>",
+          "type": "提示词",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p5.s2.outputs[0]",
+          "name": "基础汉堡图",
+          "value": "<AI生成的基础汉堡构图：汉堡居中，食材层次分明，基础光线和构图到位，无品牌标识，白色/中性背景>",
+          "type": "图像",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p5.s3",
+      "name": "",
+      "action": "生成/关系生成/组合",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p5.s3.inputs[0]",
+          "name": "基础汉堡图",
+          "value": "<AI生成的基础汉堡构图：汉堡居中，食材层次分明，基础光线和构图到位，无品牌标识，白色/中性背景>",
+          "type": "图像",
+          "related_images": []
+        },
+        {
+          "path": "p5.s3.inputs[1]",
+          "name": "品牌细节规格",
+          "value": "<品牌标识规格：logo位置（左上角）、光影强化（温暖侧光）、背景虚化（大光圈bokeh）>",
+          "type": "文本",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p5.s3.outputs[0]",
+          "name": "叠加效果提示词",
+          "value": "<叠加提示词：based on this burger image, add brand logo in top-left corner, enhance warm side lighting with golden highlights, add bokeh background effect, increase food texture detail and steam>",
+          "type": "提示词",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p5.s4",
+      "name": "",
+      "action": "生成/图像生成/图生图",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p5.s4.inputs[0]",
+          "name": "叠加效果提示词",
+          "value": "<叠加提示词：based on this burger image, add brand logo in top-left corner, enhance warm side lighting with golden highlights, add bokeh background effect, increase food texture detail and steam>",
+          "type": "提示词",
+          "related_images": []
+        },
+        {
+          "path": "p5.s4.inputs[1]",
+          "name": "基础汉堡图",
+          "value": "<AI生成的基础汉堡构图：汉堡居中，食材层次分明，基础光线和构图到位，无品牌标识>",
+          "type": "图像",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p5.s4.outputs[0]",
+          "name": "精修商品图",
+          "value": "<最终精修的汉堡产品图：品牌标识精确放置，温暖侧光增强，背景自然虚化，食材纹理和蒸汽细节丰富，广告级商品图质量>",
+          "type": "图像",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p6.s1",
+      "name": "",
+      "action": "生成/关系生成/组合",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p6.s1.inputs[0]",
+          "name": "转换风格描述",
+          "value": "将图中人物转换为真实动画(anime)的截图风格",
+          "type": "文本",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p6.s1.outputs[0]",
+          "name": "转换指令提示词",
+          "value": "Show me the attached image as a snapshot from an actual anime",
+          "type": "提示词",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p6.s2",
+      "name": "",
+      "action": "生成/图像生成/图生图",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p6.s2.inputs[0]",
+          "name": "参考图",
+          "value": "<实拍照片：人物实拍照（女性），用于转换为动漫风格截图>",
+          "type": "参考图",
+          "related_images": []
+        },
+        {
+          "path": "p6.s2.inputs[1]",
+          "name": "转换指令提示词",
+          "value": "Show me the attached image as a snapshot from an actual anime",
+          "type": "提示词",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p6.s2.outputs[0]",
+          "name": "动漫风格图",
+          "value": "<AI生成的动漫截图风格图：女性角色，黑色长发，深色大眼，哥特风黑色洋装+蕾丝领，泪眼含泪，深蓝色暗调动漫画风，呈现为「实际动漫」截图质感>",
+          "type": "图像",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p7.s1",
+      "name": "",
+      "action": "生成/关系生成/组合",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p7.s1.inputs[0]",
+          "name": "设定资料规格",
+          "value": "Persona5风格角色设定卡，需要：正面/侧面/背面三视图、表情差分、服装分解、色板；基于已有角色图生成",
+          "type": "文本",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p7.s1.outputs[0]",
+          "name": "角色设定提示词",
+          "value": "このキャラクターと背景を元に、公式設定資料のようなキャラクターシートを作成してください。・正面、側面、背面の3面図を含める・表情差分（通常・笑顔・驚き・怒り）を追加・服装の各パーツを分解して詳細に描画・カラーパレットと素材の説明を含める・Persona5スタイルのグラフィックデザインで構成",
+          "type": "提示词",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p7.s2",
+      "name": "",
+      "action": "生成/图像生成/图生图",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p7.s2.inputs[0]",
+          "name": "角色参考图",
+          "value": "<已有角色图：女性战士角色（ソルジャンヌ），红黑色装甲战衣，蓝色能量核心，用于生成设定资料>",
+          "type": "参考图",
+          "related_images": []
+        },
+        {
+          "path": "p7.s2.inputs[1]",
+          "name": "角色设定提示词",
+          "value": "このキャラクターと背景を元に、公式設定資料のようなキャラクターシートを作成してください。・正面、側面、背面の3面図を含める・表情差分（通常・笑顔・驚き・怒り）を追加・服装の各パーツを分解して詳細に描画・カラーパレットと素材の説明を含める・Persona5スタイルのグラフィックデザインで構成",
+          "type": "提示词",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p7.s2.outputs[0]",
+          "name": "角色设定卡",
+          "value": "<AI生成的ソルジャンヌ・スーツ角色设定卡：6格展示手动装备战衣的完整流程（1.确认内装+传感器 → 2.胸部肩部装甲 → 3.腰部ユニット固定 → 4.头盔准备 → 5.头盔装着+系统启动 → 6.装备完成），每格附日文说明，整体呈现Persona5风格官方设定资料格式>",
+          "type": "图像",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p8.s1",
+      "name": "",
+      "action": "生成/关系生成/组合",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p8.s1.inputs[0]",
+          "name": "内容主题描述",
+          "value": "成都美食地图，手绘风格，鸟瞰视角，包含12个知名美食地点（串串香、红油水饺、三大炮、盖碗茶、冰粉、小龙坎火锅、钵钵鸡等），标注地标建筑和主要道路，整体暖色系手绘质感",
+          "type": "文本",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p8.s1.outputs[0]",
+          "name": "多语言混合提示词",
+          "value": "一张手绘风格的城市美食地图，以成都为主题。画面以鸟瞰视角的手绘简化城市地图为底，标注主要道路和地标但不追求精确比例而是追求可爱的手绘感。地图上分布着12个美食地点的精致手绘小插画：春熙路的串串香（一把竹签插着各种食材冒着热气）、宽窄巷子的三大炮（三个糯米团子飞向铜盘）、玉林路的小龙坎火锅（九宫格锅翻滚冒泡，越煮越巴适）、东郊记忆的冒椒麻辣（冒菜要冒到头），武侯祠大街的钵钵鸡（冷锅串串）、香港巷的肥肠粉（红薯粉Q弹），人民公园鹤鸣茶社的盖碗茶（茶碗与茶盖），锦里古街的冰粉（手握冰粉碗），双流老妈兔头（麻辣入味），建设路叶婆婆烘蛋糕（外酥里软），钟水饺（总店，红油水饺），陈麻婆豆腐（麻辣鲜香）。左上角标题「成都·吃货暴走地图」使用胖圆的手绘美术字配辣椒装饰。整体画风为水彩+彩铅混合的手绘质感，颜色以暖色系为主，边框用辣椒和植物装饰，右下角有图例说明和指南针。watercolor illustration, handmade texture, warm color palette, cozy illustrated map style",
+          "type": "提示词",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p8.s2",
+      "name": "",
+      "action": "生成/图像生成/文生图",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p8.s2.inputs[0]",
+          "name": "多语言混合提示词",
+          "value": "一张手绘风格的城市美食地图，以成都为主题。画面以鸟瞰视角的手绘简化城市地图为底，标注主要道路和地标但不追求精确比例而是追求可爱的手绘感。地图上分布着12个美食地点的精致手绘小插画：春熙路的串串香（一把竹签插着各种食材冒着热气）、宽窄巷子的三大炮（三个糯米团子飞向铜盘）、玉林路的小龙坎火锅（九宫格锅翻滚冒泡，越煮越巴适）、东郊记忆的冒椒麻辣，武侯祠大街的钵钵鸡（冷锅串串），人民公园鹤鸣茶社的盖碗茶，锦里古街的冰粉（手握冰粉碗），双流老妈兔头，建设路叶婆婆烘蛋糕，钟水饺（总店），陈麻婆豆腐（麻辣鲜香）。左上角标题「成都·吃货暴走地图」使用胖圆的手绘美术字配辣椒装饰。整体画风为水彩+彩铅混合的手绘质感，颜色以暖色系为主，边框用辣椒和植物装饰，右下角有图例说明和指南针。watercolor illustration, handmade texture, warm color palette, cozy illustrated map style",
+          "type": "提示词",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p8.s2.outputs[0]",
+          "name": "本地化内容图",
+          "value": "<AI生成的成都吃货暴走地图：手绘水彩风格，鸟瞰成都城市地图，环形路/春熙路/宽窄巷子/锦里等地标可辨，12个美食地点各附精致手绘插画（串串香/水饺/三大炮/盖碗茶/冰粉等），暖黄暖棕色调，辣椒边框装饰，左上角「成都·吃货暴走地图」美术字，右下角图例说明+指南针>",
+          "type": "图像",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p9.s1",
+      "name": "",
+      "action": "生成/关系生成/组合",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p9.s1.inputs[0]",
+          "name": "结构精确性要求",
+          "value": "三层木质书架计数测试：顶层1本书，第二层3本书，底层7本书，共11本",
+          "type": "文本",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p9.s1.outputs[0]",
+          "name": "结构测试提示词",
+          "value": "A wooden bookshelf consisting of three shelves: On the top shelf, there should be one book, on the second shelf, there should be three books, and on the bottom shelf, there should be seven books.",
+          "type": "提示词",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p9.s2",
+      "name": "",
+      "action": "生成/图像生成/文生图",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p9.s2.inputs[0]",
+          "name": "结构测试提示词",
+          "value": "A wooden bookshelf consisting of three shelves: On the top shelf, there should be one book, on the second shelf, there should be three books, and on the bottom shelf, there should be seven books.",
+          "type": "提示词",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p9.s2.outputs[0]",
+          "name": "结构验证图",
+          "value": "<AI生成的木质书架图：三层书架，顶层1本深蓝色书「THE ART OF THOUGHT」，中层3本书（绿色「THE NATURAL WORLD」+红色「HISTORY OF CIVILIZATIONS」+蓝色「THE HUMAN SPIRIT」），底层7本书（The Classics/Poetry/Science and Discovery/Art Through the Ages/Explorers and Adventures/Great Minds/Legends and Myths），精确符合1+3+7=11本要求>",
+          "type": "图像",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p9.s3",
+      "name": "",
+      "action": "判断/评估",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {},
+      "inputs": [
+        {
+          "path": "p9.s3.inputs[0]",
+          "name": "结构验证图",
+          "value": "<AI生成的木质书架图：三层书架，顶层1本深蓝色书「THE ART OF THOUGHT」，中层3本书，底层7本书，精确符合1+3+7=11本要求>",
+          "type": "图像",
+          "related_images": []
+        },
+        {
+          "path": "p9.s3.inputs[1]",
+          "name": "结构精确性要求",
+          "value": "三层木质书架计数测试：顶层1本书，第二层3本书，底层7本书，共11本",
+          "type": "文本",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p9.s3.outputs[0]",
+          "name": "验证结果",
+          "value": "GPT-Image-2通过验证：书架三层分别有1本（顶）、3本（中）、7本（底），总计11本，精确符合计数和空间推理要求；该工具具备可靠的结构化指令执行能力，可支撑复杂计数约束场景",
+          "type": "文本",
+          "related_images": []
+        }
+      ]
+    }
+  ]
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/_scratch/task_2b.json

@@ -0,0 +1,378 @@
+{
+  "case_id": "A_zhihu_20382372",
+  "task": "Phase 2B (substance/form matching)",
+  "image_url_list": [],
+  "items_to_match": [
+    {
+      "path": "p1.s1.inputs[0]",
+      "name": "产品品类描述",
+      "value": "电商香水瓶产品主图，需要生成一套可替换品牌名和装饰花材的系列图，保持luxurious cinematic风格不变",
+      "type": "文本",
+      "related_images": []
+    },
+    {
+      "path": "p1.s1.outputs[0]",
+      "name": "参数变量清单",
+      "value": "参数变量：brand label（默认 N°5 CHANEL PARIS PARFUM）、flowers（默认 chamomile blossoms）；固定元素：glossy black marble surface with white veining / cinematic product photograph / dramatic side lighting / ultra-detailed glass texture / photorealistic 8K",
+      "type": "参数列表",
+      "related_images": []
+    },
+    {
+      "path": "p1.s2.inputs[0]",
+      "name": "参数变量清单",
+      "value": "参数变量：brand label（默认 N°5 CHANEL PARIS PARFUM）、flowers（默认 chamomile blossoms）；固定元素：glossy black marble surface with white veining / cinematic product photograph / dramatic side lighting / ultra-detailed glass texture / photorealistic 8K",
+      "type": "参数列表",
+      "related_images": []
+    },
+    {
+      "path": "p1.s2.outputs[0]",
+      "name": "参数化提示词模板",
+      "value": "A luxurious cinematic product photograph of a classic rectangular perfume bottle inspired by {argument name=\"brand label\" default=\"N°5 CHANEL PARIS PARFUM\"}, placed upright on a glossy black marble surface with white veining, surrounded by scattered {argument name=\"flowers\" default=\"chamomile blossoms\"}, dramatic side lighting creating specular highlights on the glass, ultra-detailed glass texture, photorealistic rendering, 8K, luxury advertising style",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p1.s3.inputs[0]",
+      "name": "参数化提示词模板",
+      "value": "A luxurious cinematic product photograph of a classic rectangular perfume bottle inspired by {argument name=\"brand label\" default=\"N°5 CHANEL PARIS PARFUM\"}, placed upright on a glossy black marble surface with white veining, surrounded by scattered {argument name=\"flowers\" default=\"chamomile blossoms\"}, dramatic side lighting creating specular highlights on the glass, ultra-detailed glass texture, photorealistic rendering, 8K, luxury advertising style",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p1.s3.outputs[0]",
+      "name": "商品主图",
+      "value": "<AI生成的香水瓶商品主图：Jimmy Choo香水瓶，深色渐变玻璃瓶身+金色JC标志，置于黑色大理石台面，金属花瓶和水晶碗陪衬，高端奢华广告风格，Pollo.ai水印>",
+      "type": "图像",
+      "related_images": []
+    },
+    {
+      "path": "p2.s1.inputs[0]",
+      "name": "布局需求描述",
+      "value": "4格日系广告Banner（2x2网格），各象限主题分别为：旅行（冲绳旅行3日 39,800日元）、护肤（透明感水光肌 初回限定78%OFF 1,980日元）、美食（黑毛和牛A4/A5豪华牛排 期间限定4,980日元）、学习（在线资格讲座 10万人受讲 20%OFF）",
+      "type": "文本",
+      "related_images": []
+    },
+    {
+      "path": "p2.s1.outputs[0]",
+      "name": "JSON结构化提示词",
+      "value": "{\n  \"type\": \"2x2 grid of Japanese digital advertisement banners\",\n  \"layout\": {\n    \"structure\": \"4 equal quadrants\",\n    \"quadrants\": [\n      {\n        \"position\": \"top-left\",\n        \"theme\": \"Travel\",\n        \"subject\": \"A couple holding hands on a white sand beach with turquoise water in Okinawa, bright tropical sky\",\n        \"text_labels\": [\"今年こそ、解き放て。\", \"沖縄旅行\", \"3日間の癒やし旅\", \"39,800円〜\", \"航空券+ホテル\", \"往復航空券\", \"人気ホテル宿泊\", \"レンタカー付き\", \"絶景、グルメ、体験ぜんぶ叶う！\"]\n      },\n      {\n        \"position\": \"top-right\",\n        \"theme\": \"Skincare\",\n        \"subject\": \"Close-up of a young woman's face applying glowing skincare cream, skin radiant and clear\",\n        \"text_labels\": [\"毛穴・くすみ卒業!\", \"透明感あふれる水光肌へ\", \"新感覚スキンケア\", \"毛穴ケア\", \"高保湿\", \"ハリ・ツヤ\", \"初回限定 78%OFF\", \"1,980円\"]\n      },\n      {\n        \"position\": \"bottom-left\",\n        \"theme\": \"Food\",\n        \"subject\": \"Premium A5 Wagyu beef steak sizzling on a grill, close-up, steam and fat marbling visible\",\n        \"text_labels\": [\"とろける旨さ!\", \"黒毛和牛 A4 A5等級 贅沢ステーキ\", \"期間限定 特別価格\", \"通価格8,980円 4,980円\"]\n      },\n      {\n        \"position\": \"bottom-right\",\n        \"theme\": \"Education\",\n        \"subject\": \"Young man studying efficiently with smartphone at a bright modern desk\",\n        \"text_labels\": [\"スキマ時間で最短合格!\", \"オンライン資格講座\", \"スマホで完結\", \"効率学習で差がつく!\", \"受講者数10万人突破!\", \"今だけ! 受講料20%OFF\"]\n      }\n    ]\n  }\n}",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p2.s2.inputs[0]",
+      "name": "JSON结构化提示词",
+      "value": "{\n  \"type\": \"2x2 grid of Japanese digital advertisement banners\",\n  \"layout\": {\n    \"structure\": \"4 equal quadrants\",\n    \"quadrants\": [\n      {\"position\": \"top-left\", \"theme\": \"Travel\", \"subject\": \"A couple holding hands on a white sand beach with turquoise water in Okinawa, bright tropical sky\", \"text_labels\": [\"今年こそ、解き放て。\", \"沖縄旅行\", \"3日間の癒やし旅\", \"39,800円〜\", \"航空券+ホテル\", \"往復航空券\", \"人気ホテル宿泊\", \"レンタカー付き\", \"絶景、グルメ、体験ぜんぶ叶う！\"]},\n      {\"position\": \"top-right\", \"theme\": \"Skincare\", \"subject\": \"Close-up of a young woman's face applying glowing skincare cream, skin radiant and clear\", \"text_labels\": [\"毛穴・くすみ卒業!\", \"透明感あふれる水光肌へ\", \"初回限定 78%OFF\", \"1,980円\"]},\n      {\"position\": \"bottom-left\", \"theme\": \"Food\", \"subject\": \"Premium A5 Wagyu beef steak sizzling on a grill, close-up, steam and fat marbling visible\", \"text_labels\": [\"黒毛和牛 A4 A5等級 贅沢ステーキ\", \"期間限定 特別価格\", \"4,980円\"]},\n      {\"position\": \"bottom-right\", \"theme\": \"Education\", \"subject\": \"Young man studying efficiently with smartphone at a bright modern desk\", \"text_labels\": [\"スキマ時間で最短合格!\", \"受講者数10万人突破!\", \"受講料20%OFF\"]}\n    ]\n  }\n}",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p2.s2.outputs[0]",
+      "name": "多区域广告图",
+      "value": "<AI生成的4格日系广告Banner：左上冲绳旅行（海滩情侣背影+日文旅行信息+价格39,800円），右上护肤品（水光肌特写+品牌LUMIÈRE+折扣信息1,980円），左下黑毛和牛牛排（特写冒烟+价格4,980円），右下在线资格讲座（学习男生+10万人突破+20%OFF）>",
+      "type": "图像",
+      "related_images": []
+    },
+    {
+      "path": "p3.s1.inputs[0]",
+      "name": "拍摄场景描述",
+      "value": "便利店夜晚场景，女性中景，电影感街拍风格，有霓虹灯光效，胶片质感",
+      "type": "文本",
+      "related_images": []
+    },
+    {
+      "path": "p3.s1.outputs[0]",
+      "name": "摄影要素清单",
+      "value": "胶片格式: 35mm film photography / 光线: harsh convenience store fluorescent lighting mixed with colorful neon signs from outside / 质感: authentic film grain, high contrast, slight color cast / 构图景别: intimate medium shot / 后期风格: cinematic street editorial style",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p3.s2.inputs[0]",
+      "name": "摄影要素清单",
+      "value": "胶片格式: 35mm film photography / 光线: harsh convenience store fluorescent lighting mixed with colorful neon signs from outside / 质感: authentic film grain, high contrast, slight color cast / 构图景别: intimate medium shot / 后期风格: cinematic street editorial style",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p3.s2.outputs[0]",
+      "name": "摄影指令提示词",
+      "value": "35mm film photography with harsh convenience store fluorescent lighting mixed with colorful neon signs from outside, authentic film grain, high contrast, slight color cast, cinematic street editorial style, intimate medium shot, young Asian woman in casual white button-down shirt and black mini skirt, leaning against glass refrigerator door, holding a drink bottle, messy bun hairstyle, natural makeup, photorealistic",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p3.s3.inputs[0]",
+      "name": "摄影指令提示词",
+      "value": "35mm film photography with harsh convenience store fluorescent lighting mixed with colorful neon signs from outside, authentic film grain, high contrast, slight color cast, cinematic street editorial style, intimate medium shot, young Asian woman in casual white button-down shirt and black mini skirt, leaning against glass refrigerator door, holding a drink bottle, messy bun hairstyle, natural makeup, photorealistic",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p3.s3.outputs[0]",
+      "name": "电影感人像照",
+      "value": "<AI生成的便利店霓虹人像：年轻亚裔女性，白色半透明衬衫+黑色短裙，倚着便利店冰柜玻璃门，手持棕色饮料瓶，荧光灯+霓虹灯混合光效，35mm胶片颗粒感，高对比度电影街拍风格>",
+      "type": "图像",
+      "related_images": []
+    },
+    {
+      "path": "p4.s1.inputs[0]",
+      "name": "拍摄场景描述",
+      "value": "日式温泉旅馆（onsen ryokan）人像，暖色木质灯笼光源，editorial高大气感风格，胶片质感，女性",
+      "type": "文本",
+      "related_images": []
+    },
+    {
+      "path": "p4.s1.outputs[0]",
+      "name": "摄影要素清单",
+      "value": "胶片格式: 35mm film photography / 光线: warm vintage Japanese onsen ryokan aesthetic, soft ambient wooden lantern lighting mixed with gentle natural window light / 质感: subtle film grain, gentle color shift / 构图景别: high atmosphere editorial style / 后期风格: vintage Japanese aesthetic",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p4.s2.inputs[0]",
+      "name": "摄影要素清单",
+      "value": "胶片格式: 35mm film photography / 光线: warm vintage Japanese onsen ryokan aesthetic, soft ambient wooden lantern lighting mixed with gentle natural window light / 质感: subtle film grain, gentle color shift / 构图景别: high atmosphere editorial style / 后期风格: vintage Japanese aesthetic",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p4.s2.outputs[0]",
+      "name": "摄影指令提示词",
+      "value": "35mm film photography, warm vintage Japanese onsen ryokan aesthetic, soft ambient wooden lantern lighting mixed with gentle natural window light, subtle film grain, gentle color shift, high atmosphere editorial style, young Asian woman in traditional Japanese yukata robe, sitting gracefully on wooden veranda, hot spring visible in background, photorealistic",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p4.s3.inputs[0]",
+      "name": "摄影指令提示词",
+      "value": "35mm film photography, warm vintage Japanese onsen ryokan aesthetic, soft ambient wooden lantern lighting mixed with gentle natural window light, subtle film grain, gentle color shift, high atmosphere editorial style, young Asian woman in traditional Japanese yukata robe, sitting gracefully on wooden veranda, hot spring visible in background, photorealistic",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p4.s3.outputs[0]",
+      "name": "温泉旅馆人像",
+      "value": "<AI生成的温泉旅馆editorial人像：年轻亚裔女性，浅色宽松和风浴衣，坐在木质台阶上，背景可见温泉热水，室内木质灯笼暖光，胶片颗粒感，高大气感editorial风格>",
+      "type": "图像",
+      "related_images": []
+    },
+    {
+      "path": "p5.s1.inputs[0]",
+      "name": "商品描述",
+      "value": "汉堡产品图，需要精确放置品牌标识、增强光影效果、添加背景虚化",
+      "type": "文本",
+      "related_images": []
+    },
+    {
+      "path": "p5.s1.outputs[0]",
+      "name": "基础构图提示词",
+      "value": "<基础构图提示词：product photography of a gourmet burger, centered composition, layers clearly visible (bun / lettuce / tomato / beef patty / cheese), studio lighting from top-left, white/neutral background, sharp focus on burger>",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p5.s2.inputs[0]",
+      "name": "基础构图提示词",
+      "value": "<基础构图提示词：product photography of a gourmet burger, centered composition, layers clearly visible (bun / lettuce / tomato / beef patty / cheese), studio lighting from top-left, white/neutral background, sharp focus on burger>",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p5.s2.outputs[0]",
+      "name": "基础汉堡图",
+      "value": "<AI生成的基础汉堡构图：汉堡居中，食材层次分明，基础光线和构图到位，无品牌标识，白色/中性背景>",
+      "type": "图像",
+      "related_images": []
+    },
+    {
+      "path": "p5.s3.inputs[0]",
+      "name": "基础汉堡图",
+      "value": "<AI生成的基础汉堡构图：汉堡居中，食材层次分明，基础光线和构图到位，无品牌标识，白色/中性背景>",
+      "type": "图像",
+      "related_images": []
+    },
+    {
+      "path": "p5.s3.inputs[1]",
+      "name": "品牌细节规格",
+      "value": "<品牌标识规格：logo位置（左上角）、光影强化（温暖侧光）、背景虚化（大光圈bokeh）>",
+      "type": "文本",
+      "related_images": []
+    },
+    {
+      "path": "p5.s3.outputs[0]",
+      "name": "叠加效果提示词",
+      "value": "<叠加提示词：based on this burger image, add brand logo in top-left corner, enhance warm side lighting with golden highlights, add bokeh background effect, increase food texture detail and steam>",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p5.s4.inputs[0]",
+      "name": "叠加效果提示词",
+      "value": "<叠加提示词：based on this burger image, add brand logo in top-left corner, enhance warm side lighting with golden highlights, add bokeh background effect, increase food texture detail and steam>",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p5.s4.inputs[1]",
+      "name": "基础汉堡图",
+      "value": "<AI生成的基础汉堡构图：汉堡居中，食材层次分明，基础光线和构图到位，无品牌标识>",
+      "type": "图像",
+      "related_images": []
+    },
+    {
+      "path": "p5.s4.outputs[0]",
+      "name": "精修商品图",
+      "value": "<最终精修的汉堡产品图：品牌标识精确放置，温暖侧光增强，背景自然虚化，食材纹理和蒸汽细节丰富，广告级商品图质量>",
+      "type": "图像",
+      "related_images": []
+    },
+    {
+      "path": "p6.s1.inputs[0]",
+      "name": "转换风格描述",
+      "value": "将图中人物转换为真实动画(anime)的截图风格",
+      "type": "文本",
+      "related_images": []
+    },
+    {
+      "path": "p6.s1.outputs[0]",
+      "name": "转换指令提示词",
+      "value": "Show me the attached image as a snapshot from an actual anime",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p6.s2.inputs[0]",
+      "name": "参考图",
+      "value": "<实拍照片：人物实拍照（女性），用于转换为动漫风格截图>",
+      "type": "参考图",
+      "related_images": []
+    },
+    {
+      "path": "p6.s2.inputs[1]",
+      "name": "转换指令提示词",
+      "value": "Show me the attached image as a snapshot from an actual anime",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p6.s2.outputs[0]",
+      "name": "动漫风格图",
+      "value": "<AI生成的动漫截图风格图：女性角色，黑色长发，深色大眼，哥特风黑色洋装+蕾丝领，泪眼含泪，深蓝色暗调动漫画风，呈现为「实际动漫」截图质感>",
+      "type": "图像",
+      "related_images": []
+    },
+    {
+      "path": "p7.s1.inputs[0]",
+      "name": "设定资料规格",
+      "value": "Persona5风格角色设定卡，需要：正面/侧面/背面三视图、表情差分、服装分解、色板；基于已有角色图生成",
+      "type": "文本",
+      "related_images": []
+    },
+    {
+      "path": "p7.s1.outputs[0]",
+      "name": "角色设定提示词",
+      "value": "このキャラクターと背景を元に、公式設定資料のようなキャラクターシートを作成してください。・正面、側面、背面の3面図を含める・表情差分（通常・笑顔・驚き・怒り）を追加・服装の各パーツを分解して詳細に描画・カラーパレットと素材の説明を含める・Persona5スタイルのグラフィックデザインで構成",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p7.s2.inputs[0]",
+      "name": "角色参考图",
+      "value": "<已有角色图：女性战士角色（ソルジャンヌ），红黑色装甲战衣，蓝色能量核心，用于生成设定资料>",
+      "type": "参考图",
+      "related_images": []
+    },
+    {
+      "path": "p7.s2.inputs[1]",
+      "name": "角色设定提示词",
+      "value": "このキャラクターと背景を元に、公式設定資料のようなキャラクターシートを作成してください。・正面、側面、背面の3面図を含める・表情差分（通常・笑顔・驚き・怒り）を追加・服装の各パーツを分解して詳細に描画・カラーパレットと素材の説明を含める・Persona5スタイルのグラフィックデザインで構成",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p7.s2.outputs[0]",
+      "name": "角色设定卡",
+      "value": "<AI生成的ソルジャンヌ・スーツ角色设定卡：6格展示手动装备战衣的完整流程（1.确认内装+传感器 → 2.胸部肩部装甲 → 3.腰部ユニット固定 → 4.头盔准备 → 5.头盔装着+系统启动 → 6.装备完成），每格附日文说明，整体呈现Persona5风格官方设定资料格式>",
+      "type": "图像",
+      "related_images": []
+    },
+    {
+      "path": "p8.s1.inputs[0]",
+      "name": "内容主题描述",
+      "value": "成都美食地图，手绘风格，鸟瞰视角，包含12个知名美食地点（串串香、红油水饺、三大炮、盖碗茶、冰粉、小龙坎火锅、钵钵鸡等），标注地标建筑和主要道路，整体暖色系手绘质感",
+      "type": "文本",
+      "related_images": []
+    },
+    {
+      "path": "p8.s1.outputs[0]",
+      "name": "多语言混合提示词",
+      "value": "一张手绘风格的城市美食地图，以成都为主题。画面以鸟瞰视角的手绘简化城市地图为底，标注主要道路和地标但不追求精确比例而是追求可爱的手绘感。地图上分布着12个美食地点的精致手绘小插画：春熙路的串串香（一把竹签插着各种食材冒着热气）、宽窄巷子的三大炮（三个糯米团子飞向铜盘）、玉林路的小龙坎火锅（九宫格锅翻滚冒泡，越煮越巴适）、东郊记忆的冒椒麻辣（冒菜要冒到头），武侯祠大街的钵钵鸡（冷锅串串）、香港巷的肥肠粉（红薯粉Q弹），人民公园鹤鸣茶社的盖碗茶（茶碗与茶盖），锦里古街的冰粉（手握冰粉碗），双流老妈兔头（麻辣入味），建设路叶婆婆烘蛋糕（外酥里软），钟水饺（总店，红油水饺），陈麻婆豆腐（麻辣鲜香）。左上角标题「成都·吃货暴走地图」使用胖圆的手绘美术字配辣椒装饰。整体画风为水彩+彩铅混合的手绘质感，颜色以暖色系为主，边框用辣椒和植物装饰，右下角有图例说明和指南针。watercolor illustration, handmade texture, warm color palette, cozy illustrated map style",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p8.s2.inputs[0]",
+      "name": "多语言混合提示词",
+      "value": "一张手绘风格的城市美食地图，以成都为主题。画面以鸟瞰视角的手绘简化城市地图为底，标注主要道路和地标但不追求精确比例而是追求可爱的手绘感。地图上分布着12个美食地点的精致手绘小插画：春熙路的串串香（一把竹签插着各种食材冒着热气）、宽窄巷子的三大炮（三个糯米团子飞向铜盘）、玉林路的小龙坎火锅（九宫格锅翻滚冒泡，越煮越巴适）、东郊记忆的冒椒麻辣，武侯祠大街的钵钵鸡（冷锅串串），人民公园鹤鸣茶社的盖碗茶，锦里古街的冰粉（手握冰粉碗），双流老妈兔头，建设路叶婆婆烘蛋糕，钟水饺（总店），陈麻婆豆腐（麻辣鲜香）。左上角标题「成都·吃货暴走地图」使用胖圆的手绘美术字配辣椒装饰。整体画风为水彩+彩铅混合的手绘质感，颜色以暖色系为主，边框用辣椒和植物装饰，右下角有图例说明和指南针。watercolor illustration, handmade texture, warm color palette, cozy illustrated map style",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p8.s2.outputs[0]",
+      "name": "本地化内容图",
+      "value": "<AI生成的成都吃货暴走地图：手绘水彩风格，鸟瞰成都城市地图，环形路/春熙路/宽窄巷子/锦里等地标可辨，12个美食地点各附精致手绘插画（串串香/水饺/三大炮/盖碗茶/冰粉等），暖黄暖棕色调，辣椒边框装饰，左上角「成都·吃货暴走地图」美术字，右下角图例说明+指南针>",
+      "type": "图像",
+      "related_images": []
+    },
+    {
+      "path": "p9.s1.inputs[0]",
+      "name": "结构精确性要求",
+      "value": "三层木质书架计数测试：顶层1本书，第二层3本书，底层7本书，共11本",
+      "type": "文本",
+      "related_images": []
+    },
+    {
+      "path": "p9.s1.outputs[0]",
+      "name": "结构测试提示词",
+      "value": "A wooden bookshelf consisting of three shelves: On the top shelf, there should be one book, on the second shelf, there should be three books, and on the bottom shelf, there should be seven books.",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p9.s2.inputs[0]",
+      "name": "结构测试提示词",
+      "value": "A wooden bookshelf consisting of three shelves: On the top shelf, there should be one book, on the second shelf, there should be three books, and on the bottom shelf, there should be seven books.",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p9.s2.outputs[0]",
+      "name": "结构验证图",
+      "value": "<AI生成的木质书架图：三层书架，顶层1本深蓝色书「THE ART OF THOUGHT」，中层3本书（绿色「THE NATURAL WORLD」+红色「HISTORY OF CIVILIZATIONS」+蓝色「THE HUMAN SPIRIT」），底层7本书（The Classics/Poetry/Science and Discovery/Art Through the Ages/Explorers and Adventures/Great Minds/Legends and Myths），精确符合1+3+7=11本要求>",
+      "type": "图像",
+      "related_images": []
+    },
+    {
+      "path": "p9.s3.inputs[0]",
+      "name": "结构验证图",
+      "value": "<AI生成的木质书架图：三层书架，顶层1本深蓝色书「THE ART OF THOUGHT」，中层3本书，底层7本书，精确符合1+3+7=11本要求>",
+      "type": "图像",
+      "related_images": []
+    },
+    {
+      "path": "p9.s3.inputs[1]",
+      "name": "结构精确性要求",
+      "value": "三层木质书架计数测试：顶层1本书，第二层3本书，底层7本书，共11本",
+      "type": "文本",
+      "related_images": []
+    },
+    {
+      "path": "p9.s3.outputs[0]",
+      "name": "验证结果",
+      "value": "GPT-Image-2通过验证：书架三层分别有1本（顶）、3本（中）、7本（底），总计11本，精确符合计数和空间推理要求；该工具具备可靠的结构化指令执行能力，可支撑复杂计数约束场景",
+      "type": "文本",
+      "related_images": []
+    }
+  ]
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/_trace.md

@@ -0,0 +1,695 @@
+
+
+---
+
+## ▶ Fresh @ 2026-06-03 14:09:09
+
+- case: `A_zhihu_20382372`
+- source: `C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\runs_full\q0018\procedures\A_zhihu_20382372\_source.json`
+- spec: `spec`
+- images: `8`
+- model: `claude-sonnet-4-6`
+- max_turns: `300`
+
+_session=b706877c · model='claude-sonnet-4-6' · apiKeySource='none'_
+
+
+### Turn 1 · 14:09:17
+
+
+### Turn 2 · 14:09:18
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/README.md`
+
+
+### Turn 3 · 14:09:22
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools.md`
+
+
+### Turn 4 · 14:09:27
+
+
+### Turn 5 · 14:09:28
+> Now let me read the source JSON and Phase 1 spec files in parallel:
+
+
+
+### Turn 6 · 14:09:30
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/_source.json`
+
+
+### Turn 7 · 14:09:30
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/syntax.md`
+
+
+### Turn 8 · 14:09:31
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/fields.md`
+
+
+### Turn 9 · 14:09:38
+
+
+### Turn 10 · 14:09:38
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/control-flow.md`
+
+
+### Turn 11 · 14:09:39
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/phase1-skeleton.md`
+
+
+### Turn 12 · 14:09:39
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/format/md-structure.md`
+
+
+### Turn 13 · 14:09:47
+
+
+### Turn 14 · 14:09:47
+> Now let me read the workflow template and Phase 2 normalize spec in parallel:
+
+
+
+### Turn 15 · 14:09:49
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/templates/workflow.template.json`
+
+
+### Turn 16 · 14:09:50
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/phase2-normalize.md`
+
+
+### Turn 17 · 14:18:05
+
+
+### Turn 18 · 14:18:05
+> Now I have all the spec context I need. Let me start Phase 1 — first write the understanding.md, then the workflow.json skeleton.
+
+
+
+### Turn 19 · 14:19:09
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/understanding.md` (4,116 chars)
+
+
+### Turn 20 · 14:20:27
+
+
+### Turn 21 · 14:20:27
+> Now let me write the Phase 1.2 workflow.json skeleton:
+
+
+
+### Turn 22 · 14:24:59
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/workflow.json` (43,010 chars)
+
+
+### Turn 23 · 14:25:47
+
+
+### Turn 24 · 14:25:47
+> The workflow.json was written but has JSON syntax errors in p5's instruction items (I incorrectly put `inferred` inside array literals). Let me fix those:
+
+
+
+### Turn 25 · 14:25:49
+- `Edit` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/workflow.json`
+
+
+### Turn 26 · 14:25:57
+- `Edit` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/workflow.json`
+
+
+### Turn 27 · 14:26:25
+
+
+### Turn 28 · 14:26:25
+> Now let me verify the JSON is valid and then add anchors in Phase 1.3:
+
+
+
+### Turn 29 · 14:26:27
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python -c "import json; data=json.load(open('../runs_full/q0018/procedures/A_zhihu_20382372...`
+
+
+### Turn 30 · 14:27:08
+
+
+### Turn 31 · 14:27:08
+> JSON is valid (9 procedures). Now let me write the anchor patch file for Phase 1.3:
+
+
+
+### Turn 32 · 14:27:23
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/_scratch/anchors.json` (3,189 chars)
+
+
+### Turn 33 · 14:27:29
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/wf-patch.py --workflow "../runs_full/q0018/procedures/A_zhihu_20382372/wo...`
+
+
+### Turn 34 · 14:27:49
+
+
+### Turn 35 · 14:27:49
+> All 53 anchors applied. Phase 1 is complete. Now for Phase 2 — let me first check if prepare-subtask.py exists and then run it:
+
+
+
+### Turn 36 · 14:27:49
+- `Glob` — pattern=`prepare-subtask.py`
+
+
+### Turn 37 · 14:27:55
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/prepare-subtask.py --workflow "../runs_full/q0018/procedures/A_zhihu_2038...`
+
+
+### Turn 38 · 14:28:09
+
+
+### Turn 39 · 14:28:09
+> Excellent! Both task files generated. Now launching Phase 2A and 2B sub-agents in parallel:
+
+
+
+### Turn 40 · 14:28:25
+- `Agent` — `Phase 2A: effect/action/type normalization` [subagent=phase-2a-normalizer]
+
+
+### Turn 41 · 14:28:28
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/_scratch/task_2a.json`
+
+
+### Turn 42 · 14:28:28
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/taxonomy/effect.json`
+
+
+### Turn 43 · 14:28:29
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/taxonomy/action.json`
+
+
+### Turn 44 · 14:28:30
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/taxonomy/type.json`
+
+
+### Turn 45 · 14:28:40
+- `Agent` — `Phase 2B: substance/form taxonomy matching` [subagent=phase-2b-matcher]
+
+
+### Turn 46 · 14:28:46
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/_scratch/task_2b.json`
+
+
+### Turn 47 · 14:28:53
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "产品描述 商品需求 香水" 2>&1 | head -60`
+
+
+### Turn 48 · 14:28:55
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "文本描述 需求说明" 2>&1 | head -60`
+
+
+### Turn 49 · 14:28:58
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "香水瓶 产品特征 奢侈品" 2>&1 | head -60`
+
+
+### Turn 50 · 14:29:00
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "自然语言 文本 描述说明" 2>&1 | head -60`
+
+
+### Turn 51 · 14:29:06
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "自然语言 陈述句 需求说明" 2>&1 | head -40`
+
+
+### Turn 52 · 14:29:07
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "参数变量 模板参数 设计元素" 2>&1 | head -60`
+
+
+### Turn 53 · 14:29:10
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "参数 配置 规格清单" 2>&1 | head -60`
+
+
+### Turn 54 · 14:29:11
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "清单 列表 参数列表" 2>&1 | head -60`
+
+
+### Turn 55 · 14:29:17
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "提示词 图像生成指令 AI绘图" 2>&1 | head -60`
+
+
+### Turn 56 · 14:29:18
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "提示词 prompt 英文描述" 2>&1 | head -60`
+
+
+### Turn 57 · 14:29:21
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --narrow "提示词 prompt AI绘图指令" 2>&1 | head -60`
+
+
+### Turn 58 · 14:29:23
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "指令 文本指令 操作说明" 2>&1 | head -60`
+
+
+### Turn 59 · 14:29:26
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "英文 模板 结构化文本" 2>&1 | head -60`
+
+
+### Turn 60 · 14:29:27
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "AI指令 生成指令 提示" 2>&1 | head -40`
+
+
+### Turn 61 · 14:29:32
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "语言 文字表述 书写语言" 2>&1 | head -60`
+
+
+### Turn 62 · 14:29:33
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --narrow "语言文字 书写 表达方式" 2>&1 | head -60`
+
+
+### Turn 63 · 14:29:37
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --subtree "/架构/语言" 2>&1 | head -60`
+
+
+### Turn 64 · 14:29:39
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --list-roots 2>&1 | head -40`
+
+
+### Turn 65 · 14:29:42
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --list-l2 2>&1 | head -40`
+
+
+### Turn 66 · 14:29:44
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --list-l2 2>&1 | head -40`
+
+
+### Turn 67 · 14:29:50
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "AI技术 图像生成 生成式AI" 2>&1 | head -60`
+
+
+### Turn 68 · 14:29:51
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "方法论 创作方法 技巧" 2>&1 | head -60`
+
+
+### Turn 69 · 14:29:53
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "AI应用 生成式图像 AI工具" 2>&1 | head -60`
+
+
+### Turn 70 · 14:29:55
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "摄影 人像照 胶片街拍" 2>&1 | head -60`
+
+
+### Turn 71 · 14:29:59
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "人物 女性 人像" 2>&1 | head -60`
+
+
+### Turn 72 · 14:30:00
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "摄影图像 照片 AI生成图" 2>&1 | head -60`
+
+
+### Turn 73 · 14:30:06
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "图像 静态图片 生成图" 2>&1 | head -60`
+
+
+### Turn 74 · 14:30:07
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "广告 商业广告 Banner" 2>&1 | head -60`
+
+
+### Turn 75 · 14:30:09
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/_scratch/patch_2a.json` (5,230 chars)
+
+
+### Turn 76 · 14:30:12
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "电商广告 商品推广 产品营销" 2>&1 | head -60`
+
+
+### Turn 77 · 14:30:13
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "版面 多格 网格布局" 2>&1 | head -60`
+
+
+### Turn 78 · 14:30:14
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/wf-patch.py" "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipelin...`
+
+
+### Turn 79 · 14:30:17
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "动漫 anime 动画风格角色" 2>&1 | head -60`
+
+
+### Turn 80 · 14:30:18
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "动漫 插画 绘画风格" 2>&1 | head -60`
+
+
+### Turn 81 · 14:30:18
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/wf-patch.py" --workflow "C:/Users/11304/gitlab/cybertogether/Agent/examples/proc...`
+
+
+### Turn 82 · 14:30:27
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "地图 城市地图 美食地图" 2>&1 | head -60`
+
+
+### Turn 83 · 14:30:28
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "手绘插画 水彩 插画风格" 2>&1 | head -60`
+
+
+### Turn 84 · 14:30:32
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "书架 家具 陈列物品" 2>&1 | head -60`
+
+
+### Turn 85 · 14:30:33
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "摄影指令 摄影风格 场景描述" 2>&1 | head -60`
+
+
+### Turn 86 · 14:30:37
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "摄影 光线 构图 拍摄" 2>&1 | head -60`
+
+
+### Turn 87 · 14:30:38
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "产品摄影 商品图 广告图" 2>&1 | head -60`
+
+
+### Turn 88 · 14:30:42
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "商品 香水 产品展示图" 2>&1 | head -60`
+
+
+### Turn 89 · 14:30:43
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "商品摄影 产品图 摄影作品" 2>&1 | head -60`
+
+
+### Turn 90 · 14:30:47
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "实景拍摄 摄影作品 照片" 2>&1 | head -60`
+
+
+### Turn 91 · 14:30:48
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "汉堡 食物 餐饮产品" 2>&1 | head -60`
+
+
+### Turn 92 · 14:30:54
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "角色设定 游戏角色 人物设定卡" 2>&1 | head -60`
+
+
+### Turn 93 · 14:30:55
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "设定卡 角色图 多格展示" 2>&1 | head -60`
+
+
+### Turn 94 · 14:31:00
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "卡通 游戏风格 角色展示" 2>&1 | head -60`
+
+
+### Turn 95 · 14:31:01
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "摄影技术要素 灯光 胶片格式" 2>&1 | head -60`
+
+
+### Turn 96 · 14:31:06
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "摄影美学 摄影构图 创作风格" 2>&1 | head -40`
+
+
+### Turn 97 · 14:31:07
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "JSON 结构化数据 配置文件" 2>&1 | head -40`
+
+
+### Turn 98 · 14:31:12
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "JSON 结构化 键值对" 2>&1 | head -60`
+
+
+### Turn 99 · 14:31:13
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "结构化格式 规范框架 层级结构" 2>&1 | head -60`
+
+
+### Turn 100 · 14:31:19
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "摄影概念 构图方式 摄影参数" 2>&1 | head -40`
+
+
+### Turn 101 · 14:31:20
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "便利店 夜景 场景" 2>&1 | head -40`
+
+
+### Turn 102 · 14:31:25
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "温泉旅馆 日式旅馆 旅行场景" 2>&1 | head -40`
+
+
+### Turn 103 · 14:31:27
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "验证 测试结果 评估" 2>&1 | head -60`
+
+
+### Turn 104 · 14:31:30
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --narrow "温泉 旅馆 日本旅行" 2>&1 | head -60`
+
+
+### Turn 105 · 14:31:31
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "旅游 住宿 旅馆" 2>&1 | head -60`
+
+
+### Turn 106 · 14:31:36
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "日本 日式 和风场景" 2>&1 | head -40`
+
+
+### Turn 107 · 14:31:38
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "AI工具 性能测试 能力评测" 2>&1 | head -40`
+
+
+### Turn 108 · 14:31:42
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "AI能力 工具评价 功能测试" 2>&1 | head -40`
+
+
+### Turn 109 · 14:31:44
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "陈述 评价报告 文字说明" 2>&1 | head -40`
+
+
+### Turn 110 · 14:31:47
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "AI图像 生成能力 GPT图像" 2>&1 | head -40`
+
+
+### Turn 111 · 14:31:49
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "实拍照片 参考图 图像素材" 2>&1 | head -40`
+
+
+### Turn 112 · 14:31:53
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "真人照片 人物照片 实拍" 2>&1 | head -40`
+
+
+### Turn 113 · 14:31:54
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "实拍照片 照片 摄影" 2>&1 | head -40`
+
+
+### Turn 114 · 14:31:59
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "角色 虚构人物 战士" 2>&1 | head -40`
+
+
+### Turn 115 · 14:32:01
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "摄影风格 cinematic 电影质感" 2>&1 | head -40`
+
+
+### Turn 116 · 14:32:06
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "摄影 写实风格 真实感" 2>&1 | head -40`
+
+
+### Turn 117 · 14:32:07
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "写实 纪实 真实风格" 2>&1 | head -40`
+
+
+### Turn 118 · 14:32:12
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "胶片 vintage 复古质感 日系" 2>&1 | head -40`
+
+
+### Turn 119 · 14:32:13
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "品牌 Logo 标识设计" 2>&1 | head -40`
+
+
+### Turn 120 · 14:32:18
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "品牌标识 商标 视觉识别" 2>&1 | head -40`
+
+
+### Turn 121 · 14:32:19
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "广告文案 布局 版式设计" 2>&1 | head -40`
+
+
+### Turn 122 · 14:32:23
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "广告版式 广告设计 视觉排版" 2>&1 | head -40`
+
+
+### Turn 123 · 14:32:24
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "广告 日系广告 营销Banner" 2>&1 | head -40`
+
+
+### Turn 124 · 14:32:28
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "动漫风格 anime截图 动画画面" 2>&1 | head -40`
+
+
+### Turn 125 · 14:32:29
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "成都美食 饮食文化 地方特产" 2>&1 | head -40`
+
+
+### Turn 126 · 14:32:32
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "美食 食物 川菜特色" 2>&1 | head -40`
+
+
+### Turn 127 · 14:32:34
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "AI图像生成 图像质量 技术能力" 2>&1 | head -40`
+
+
+### Turn 128 · 14:32:38
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --match "AI 大模型 GPT能力" 2>&1 | head -40`
+
+
+### Turn 129 · 14:32:39
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 形式 --match "总结报告 评测 结论" 2>&1 | head -40`
+
+
+### Turn 130 · 14:32:47
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --validate "/表象/视觉/实体/物品/起居物品/美妆护肤" 2>&1 && p...`
+
+
+### Turn 131 · 14:32:51
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools" && python taxonomy-lookup.py --dim 实质 --validate "/表象/视觉/实体/物品/起居物品/美妆护肤" 2>&1`
+
+
+### Turn 132 · 14:32:54
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/taxonomy-lookup.py" --dim 实质 --validate "/表象/视觉/实体/物品/起居物品/美妆护肤"`
+
+
+### Turn 133 · 14:32:58
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/taxonomy-lookup.py" --dim 实质 --validate "/表象/视觉/实体/物品/起居物品/美妆护肤" 2>&1; echo "exi...`
+
+
+### Turn 134 · 14:33:02
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/taxonomy-lookup.py" --dim 实质 --match "美妆护肤 香水" 2>&1 | head -20`
+
+
+### Turn 135 · 14:34:19
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/_scratch/patch_2b.json` (8,050 chars)
+
+
+### Turn 136 · 14:34:23
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/wf-patch.py" "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipelin...`
+
+
+### Turn 137 · 14:34:27
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/wf-patch.py" --workflow "C:/Users/11304/gitlab/cybertogether/Agent/examples/proc...`
+
+
+### Turn 138 · 14:34:59
+
+
+### Turn 139 · 14:34:59
+> Both Phase 2A (103 patches) and 2B (106 patches) applied. Now Phase 3 — let me read the remaining spec files:
+
+
+
+### Turn 140 · 14:35:00
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/phase3-finalize.md`
+
+
+### Turn 141 · 14:35:01
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/format/procedure-table.md`
+
+
+### Turn 142 · 14:36:48
+
+
+### Turn 143 · 14:36:48
+> Now let me run lint to check for any issues before rendering:
+
+
+
+### Turn 144 · 14:36:50
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/lint-case.py --workflow "../runs_full/q0018/procedures/A_zhihu_20382372/w...`
+
+
+### Turn 145 · 14:38:40
+
+
+### Turn 146 · 14:38:40
+- `Read` — `C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\runs_full\q0018\procedures\A_zhihu_20382372\workflow.json`
+
+
+### Turn 147 · 14:38:43
+- `Read` — `C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\runs_full\q0018\procedures\A_zhihu_20382372\workflow.json` (offset=0, limit=150)
+
+
+### Turn 148 · 14:38:45
+
+
+### Turn 149 · 14:38:47
+- `Read` — `C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\runs_full\q0018\procedures\A_zhihu_20382372\workflow.json` (offset=150, limit=150)
+
+
+### Turn 150 · 14:38:50
+
+
+### Turn 151 · 14:38:51
+- `Read` — `C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\runs_full\q0018\procedures\A_zhihu_20382372\workflow.json` (offset=300, limit=150)
+
+
+### Turn 152 · 14:38:54
+
+
+### Turn 153 · 14:38:56
+- `Read` — `C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\runs_full\q0018\procedures\A_zhihu_20382372\workflow.json` (offset=450, limit=200)
+
+
+### Turn 154 · 14:39:00
+- `Read` — `C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\runs_full\q0018\procedures\A_zhihu_20382372\workflow.json` (offset=650, limit=250)
+
+
+### Turn 155 · 14:39:04
+- `Read` — `C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\runs_full\q0018\procedures\A_zhihu_20382372\workflow.json` (offset=900, limit=250)
+
+
+### Turn 156 · 14:39:08
+- `Read` — `C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\runs_full\q0018\procedures\A_zhihu_20382372\workflow.json` (offset=1150, limit=250)
+
+
+### Turn 157 · 14:39:11
+
+
+### Turn 158 · 14:39:13
+- `Read` — `C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\runs_full\q0018\procedures\A_zhihu_20382372\workflow.json` (offset=1400, limit=100)
+
+
+### Turn 159 · 14:40:55
+
+
+### Turn 160 · 14:40:57
+> Now I have a clear view of all 24 intent fields. Let me compile the corrections needed — `{act:}` tokens must use the action leaf name, `{in-type:}`/`{out-type:}` must use the IO item `type` field (not `name`). I'll write a patch file and apply it in one shot.
+
+
+
+### Turn 161 · 14:41:12
+- `Write` — `C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\runs_full\q0018\procedures\A_zhihu_20382372\_scratch\patch_intent.json` (2,532 chars)
+
+
+### Turn 162 · 14:41:15
+- `Bash` — `cd "C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\procedure-dsl" && python spec/tools/wf-patch.py "../runs_full/q0018/procedures/A_zhihu_20382372/workflow.json...`
+
+
+### Turn 163 · 14:41:18
+- `Bash` — `cd "C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\procedure-dsl" && python spec/tools/wf-patch.py --workflow "../runs_full/q0018/procedures/A_zhihu_20382372/wo...`
+
+
+### Turn 164 · 14:41:22
+
+
+### Turn 165 · 14:41:22
+> All 24 intent patches applied. Now run lint again to confirm clean, then render HTML.
+
+
+
+### Turn 166 · 14:41:23
+- `Bash` — `cd "C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\procedure-dsl" && python spec/tools/lint-case.py --workflow "../runs_full/q0018/procedures/A_zhihu_20382372/w...`
+
+
+### Turn 167 · 14:41:26
+
+
+### Turn 168 · 14:41:27
+> Lint clean. Now render the HTML output:
+
+
+
+### Turn 169 · 14:41:29
+- `Bash` — `cd "C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\procedure-dsl" && python spec/tools/render-case.py --workflow "../runs_full/q0018/procedures/A_zhihu_20382372...`
+
+
+### Turn 170 · 14:41:34
+
+
+### Turn 171 · 14:41:34
+> HTML rendered: 166,300 chars, 9 procedures, 24 steps. Now read the md-structure spec to write the final `.md` file:
+
+
+
+### Turn 172 · 14:41:35
+- `Read` — `C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\procedure-dsl\spec\format\md-structure.md`

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/understanding.md

@@ -0,0 +1,131 @@
+# Phase 1.1 心智模型 — A_zhihu_20382372
+
+## 原文信息
+
+- **标题**: 即梦AI提示词方法论：从312个GPT-Image-2神级提示词中提炼的7个出图公式
+- **平台**: 知乎
+- **作者**: 白嫖知识分享
+- **发布日期**: 2026-05-14
+- **URL**: https://zhuanlan.zhihu.com/p/2038237278795981638
+
+---
+
+## 内容概述
+
+本文是一篇 AI 图像生成提示词方法论总结，作者花一周时间拆解了 312 个 GPT-Image-2 专业提示词（覆盖电商、广告创意、人像摄影、海报插画、角色设计、UI社媒、对比实验 7 大类），提炼出 7 个可复用的出图公式。文章声明这些公式在即梦、Midjourney、GPT-Image-2 上通用。
+
+配图共 8 张，分别展示 7 个公式的实战出图效果：
+- 图1: 香水瓶商品主图（Jimmy Choo 黑色渐变瓶 + 金色 JC 标）→ 公式一
+- 图2: 4 格日系广告 Banner（冲绳旅行/护肤品/黑毛和牛/在线资格讲座）→ 公式二
+- 图3: 便利店霓虹人像（年轻女性倚冰柜门）→ 公式三
+- 图4: 温泉旅馆人像（女性日式浴衣）→ 公式三（第二示例）
+- 图5: 铠甲战衣角色设定卡（6 格展示穿戴流程）→ 公式五（角色设定资料）
+- 图6: 哥特风动漫少女（黑发泪眼）→ 公式五（实拍转动漫截图）
+- 图7: 成都吃货暴走地图（手绘水彩风格城市美食地图）→ 公式六
+- 图8: 木质书架计数图（1+3+7=11 本书）→ 公式七
+
+---
+
+## 多工序判断
+
+### 判断标准回顾
+- 有明确产出物（可见结果）
+- 有具体做法（可操作的 Prompt / 框架 / 流程）
+- 产出物或做法与其他工序有差异（产出不同 OR Prompt 不同）
+
+### 扫描结果
+
+以每张成品图为起点：
+
+| 成品图 | 对应公式 | 做法差异 | 判断 |
+|---|---|---|---|
+| 图1 香水瓶商品主图 | 公式一：参数化模板 | `{argument name=... default=...}` 参数嵌入；固定风格锚点 | ✅ 独立工序 |
+| 图2 4格日系Banner | 公式二：JSON结构化 | JSON 对象定义区域+内容+文字标签；结构化 layout 语义 | ✅ 独立工序 |
+| 图3 便利店霓虹人像 | 公式三：摄影指令（A） | 35mm film + harsh fluorescent + neon + medium shot | ✅ 独立工序 |
+| 图4 温泉旅馆人像 | 公式三：摄影指令（B） | 35mm film + warm lantern + onsen aesthetic + editorial | ✅ Prompt 文本不同 = 独立工序 |
+| 图5 铠甲角色设定卡 | 公式五：参考图驱动 | 日文 Persona5 风格三视图+表情差分+服装分解，角色设定资料格式 | ✅ 独立工序 |
+| 图6 动漫哥特少女 | 公式五：参考图驱动 | "Show me the attached image as a snapshot from an actual anime"，风格迁移 | ✅ Prompt 文本不同 = 独立工序 |
+| 图7 成都美食地图 | 公式六：多语言混用 | 中文场景细节 + 英文风格术语混合；地图类特殊内容 | ✅ 独立工序 |
+| 图8 书架计数测试 | 公式七：结构化测试 | 精确计数约束（1+3+7=11），测试 AI 推理能力为目的 | ✅ 独立工序 |
+
+**公式四（分步组合，汉堡图）**: 文章描述了「先生成主体构图，再叠加品牌标识/光影/背景虚化」两步法。该图在文章中有出图效果描述但在配图8张中未独立出现（或被省略）。具有完整的输入→做法链条（两步循序生成），但缺乏具体提示词文本。作为独立工序保留（有具体做法描述）。
+
+### 最终判断：**8 个独立工序（p1-p8）**
+
+> 注意：公式三有两种摄影场景（便利店 vs 温泉旅馆），公式五有两种参考图模式（风格迁移 vs 角色设定资料）。由于 Prompt 文本有实质差异，按 spec 规则各自算独立工序。
+
+---
+
+## 各工序概述
+
+### P1 · 参数化模板批量出图（公式一）
+- **终态产物**: 香水瓶商品主图系列（参数化批量出图）
+- **步骤数**: 3步（梳理参数变量 → 构建参数化模板 → AI生成）
+- **工艺类型**: 电商产品图批量生成；参数化模板复用
+- **核心工具**: AI生图工具（GPT-Image-2 / 即梦）
+- **关键特征**: `{argument name="..." default="..."}` 语法将变量元素从固定风格中剥离
+
+### P2 · JSON结构化多区域布局出图（公式二）
+- **终态产物**: 4格日系广告Banner（2x2网格，各区域独立主题）
+- **步骤数**: 2步（编写JSON结构化提示词 → AI生成）
+- **工艺类型**: 多区域广告设计；精确布局控制
+- **核心工具**: AI生图工具
+- **关键特征**: JSON format指定 position / content / text_labels
+
+### P3 · 摄影指令出图——便利店人像（公式三-A）
+- **终态产物**: 便利店荧光灯+霓虹灯人像（35mm胶片质感街拍）
+- **步骤数**: 3步（确定摄影要素 → 组合摄影指令 → AI生成）
+- **工艺类型**: 专业人像摄影指令生成
+- **核心工具**: AI生图工具（即梦3.0 文字增强）
+- **关键特征**: {胶片类型} + {光线} + {人物描述} + {构图/景别} + {后期风格} 五要素拼接
+
+### P4 · 摄影指令出图——温泉旅馆人像（公式三-B）
+- **终态产物**: 温泉旅馆（ryokan）日式人像（暖灯氛围、editorial风格）
+- **步骤数**: 3步（确定摄影要素 → 组合摄影指令 → AI生成）
+- **工艺类型**: 专业人像摄影指令生成（温泉氛围变体）
+- **核心工具**: AI生图工具
+- **关键特征**: 同公式三框架，更换光线描述（warm vintage onsen aesthetic）
+
+### P5 · 分步组合精修出图（公式四）
+- **终态产物**: 品牌汉堡精修商品图（先基础构图 + 后叠加品牌细节）
+- **步骤数**: 4步（写基础提示词 → AI生成基础图 → 写叠加提示词 → AI叠加精修）
+- **工艺类型**: 复杂商品图两步法生成；参考图智能修改
+- **核心工具**: AI生图工具（即梦智能参考模式）
+- **关键特征**: 多轮生成流程；第二步以第一步输出为参考输入
+
+### P6 · 参考图驱动——实拍转动漫截图（公式五-A）
+- **终态产物**: 实拍人物照转动漫截图风格图
+- **步骤数**: 2步（写极简转换指令 → AI基于参考图生成）
+- **工艺类型**: 风格迁移；参考图驱动
+- **核心工具**: AI生图工具（即梦3.0 智能参考）
+- **关键特征**: 极简一句话指令 "Show me the attached image as a snapshot from an actual anime"
+
+### P7 · 参考图驱动——角色设定资料生成（公式五-B）
+- **终态产物**: Persona5风格角色设定卡（日文提示词，含三视图/表情差分/服装分解/色板）
+- **步骤数**: 2步（写日文角色设定提示词 → AI基于角色参考图生成）
+- **工艺类型**: 角色设计资料；日文提示词
+- **核心工具**: AI生图工具
+- **关键特征**: 日文提示词直接驱动；基于已有角色图生成完整设定卡
+
+### P8 · 多语言混用本地化内容出图（公式六）
+- **终态产物**: 成都吃货暴走地图（手绘水彩风格，12个美食地点）
+- **步骤数**: 2步（构建多语言混合提示词 → AI生成）
+- **工艺类型**: 本地化内容图；地图设计
+- **核心工具**: AI生图工具（GPT-Image-2 CJK优化 / 即梦中文优先）
+- **关键特征**: 中文描述场景细节 + 英文锁定风格参数 + 日文匹配风格标签
+
+> 注：公式七（结构化测试/书架计数）因其主要目的是"评估AI能力"而非创作，且链条极短（写要求 → 生成 → 人工核验），与公式六合并为本 case 的最后部分，以 **P8** 中的附注处理，不额外新增工序。实际工序数为 **8个**（p1-p8）。
+
+---
+
+## 外部工具依赖
+
+- **AI生图工具**: GPT-Image-2（主要示例工具）、即梦Dreamina（国产工具，中文优先，3.0版有智能参考功能）、Midjourney（通用，文章提及兼容）
+- **人工**: 提示词撰写、参数决策、验证核检
+
+## 隐含工艺规约
+
+1. **参数化模板** 的价值在于「视觉统一性 + 批量替换」，实际使用中会有一个批量生成循环（换参数→生成），但文章以单次为示范。
+2. **分步组合法**（P5）的第二步依赖第一步输出图作为参考输入，是典型的图生图（img2img）工作流，需工具支持参考图上传。
+3. **参考图驱动**（P6/P7）的质量高度依赖参考图的清晰度和代表性；极简指令 + 高质量参考图往往比复杂长提示词效果更好。
+4. **多语言混用**（P8）的语言分工原则：中文管内容细节、英文管风格参数、日文管特定风格标签（如「設定資料」）。

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/A_zhihu_20382372/workflow.json

@@ -0,0 +1,1437 @@
+{
+  "source": {
+    "platform": "zhihu",
+    "author": "白嫖知识分享",
+    "date": "2026-05-14",
+    "url": "https://zhuanlan.zhihu.com/p/2038237278795981638",
+    "title": "即梦AI提示词方法论：从312个GPT-Image-2神级提示词中提炼的7个出图公式",
+    "excerpt": "拆解312个专业提示词，提炼7个AI出图公式：参数化模板、JSON结构化、摄影指令、分步组合、参考图驱动、多语言混用、结构化测试，各附实战案例与出图效果。"
+  },
+  "procedures": [
+    {
+      "id": "p1",
+      "name": "参数化模板批量出图",
+      "purpose": "用 {argument name=... default=...} 语法将可变元素提升为参数，固定风格锚点，批量替换参数生成同风格电商产品主图系列",
+      "category": "产物创造",
+      "platform": "zhihu",
+      "author": "白嫖知识分享",
+      "declarations": {
+        "inputs": [
+          {
+            "type": "文本",
+            "name": "产品品类描述",
+            "desc": "需要生成图像的产品品类及核心视觉风格需求"
+          },
+          {
+            "type": "参数列表",
+            "name": "参数变量清单",
+            "desc": "需要批量替换的元素（品牌名/颜色/材质/装饰花材等）及其默认值"
+          }
+        ],
+        "resources": [],
+        "returns": {
+          "type": "图像"
+        }
+      },
+      "steps": [
+        {
+          "id": "s1",
+          "kind": "step",
+          "via": "human",
+          "action": "提取/化学提取/解构",
+          "feature": "人工",
+          "control": "-",
+          "instruction": [
+            [
+              "memo",
+              "区分「稳定不变」要素（背景/光线/构图/风格）与「需要批量替换」要素（品牌名/颜色/材质/装饰）——前者固定进模板，后者提升为参数；电商场景需要同一风格出几十张图，参数化模板保证视觉统一性"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "描述",
+              "name": "产品品类描述",
+              "value": "电商香水瓶产品主图，需要生成一套可替换品牌名和装饰花材的系列图，保持luxurious cinematic风格不变",
+              "anchor": "← 工序输入",
+              "substance": "/表象/视觉/实体/物品/起居物品/美妆护肤",
+              "form": "/架构/修辞/语调风格/平实表达/直白陈述"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "参数列表",
+              "name": "参数变量清单",
+              "value": "参数变量：brand label（默认 N°5 CHANEL PARIS PARFUM）、flowers（默认 chamomile blossoms）；固定元素：glossy black marble surface with white veining / cinematic product photograph / dramatic side lighting / ultra-detailed glass texture / photorealistic 8K",
+              "anchor": "→ s2",
+              "substance": "/理念/知识/商业/产品服务/产品特征/规格参数",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/逐条罗列/通用列举"
+            }
+          ],
+          "intent": "用 {via:human} {act:解构} 从 {in-type:描述} 梳理出可参数化的 {out-type:参数列表}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "预处理"
+        },
+        {
+          "id": "s2",
+          "kind": "step",
+          "via": "human",
+          "action": "生成/关系生成/结构生成",
+          "feature": "人工",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "以 {argument name=\"参数名\" default=\"默认值\"} 语法将参数变量内嵌到提示词中；其余描述固定为模板骨架；确保模板结构完整（主体描述+材质+环境+光影+构图+渲染风格）"
+            ],
+            [
+              "memo",
+              "即梦同样支持此思路：先确定稳定的「画风锚点」，只替换核心商品元素；参数化模板的价值在于「一次定义，批量复用」"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "参数列表",
+              "name": "参数变量清单",
+              "value": "参数变量：brand label（默认 N°5 CHANEL PARIS PARFUM）、flowers（默认 chamomile blossoms）；固定元素：glossy black marble surface with white veining / cinematic product photograph / dramatic side lighting / ultra-detailed glass texture / photorealistic 8K",
+              "anchor": "← s1.参数变量清单",
+              "substance": "/理念/知识/商业/产品服务/产品特征/规格参数",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/逐条罗列/通用列举"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "提示词",
+              "name": "参数化提示词模板",
+              "value": "A luxurious cinematic product photograph of a classic rectangular perfume bottle inspired by {argument name=\"brand label\" default=\"N°5 CHANEL PARIS PARFUM\"}, placed upright on a glossy black marble surface with white veining, surrounded by scattered {argument name=\"flowers\" default=\"chamomile blossoms\"}, dramatic side lighting creating specular highlights on the glass, ultra-detailed glass texture, photorealistic rendering, 8K, luxury advertising style",
+              "anchor": "→ s3",
+              "substance": "/表象/视觉/实体/物品/起居物品/美妆护肤",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            }
+          ],
+          "intent": "用 {via:human} {act:结构生成} 将 {in-type:参数列表} 嵌入固定风格骨架，生成可复用的 {out-type:提示词}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "预处理"
+        },
+        {
+          "id": "s3",
+          "kind": "step",
+          "via": "<AI生图工具>",
+          "action": "生成/元素生成",
+          "feature": "随机",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "A luxurious cinematic product photograph of a classic rectangular perfume bottle inspired by {argument name=\"brand label\" default=\"N°5 CHANEL PARIS PARFUM\"}, placed upright on a glossy black marble surface with white veining, surrounded by scattered {argument name=\"flowers\" default=\"chamomile blossoms\"}, dramatic side lighting creating specular highlights on the glass, ultra-detailed glass texture, photorealistic rendering, 8K, luxury advertising style"
+            ],
+            [
+              "memo",
+              "替换 brand label 参数即可切换品牌，替换 flowers 参数即可改变装饰风格；GPT-Image-2 / 即梦均支持此提示词结构；适用场景：电商产品图批量生成、品牌系列素材、社交媒体模板图"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "提示词",
+              "name": "参数化提示词模板",
+              "value": "A luxurious cinematic product photograph of a classic rectangular perfume bottle inspired by {argument name=\"brand label\" default=\"N°5 CHANEL PARIS PARFUM\"}, placed upright on a glossy black marble surface with white veining, surrounded by scattered {argument name=\"flowers\" default=\"chamomile blossoms\"}, dramatic side lighting creating specular highlights on the glass, ultra-detailed glass texture, photorealistic rendering, 8K, luxury advertising style",
+              "anchor": "← s2.参数化提示词模板",
+              "substance": "/表象/视觉/实体/物品/起居物品/美妆护肤",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "成品图",
+              "name": "商品主图",
+              "value": "<AI生成的香水瓶商品主图：Jimmy Choo香水瓶，深色渐变玻璃瓶身+金色JC标志，置于黑色大理石台面，金属花瓶和水晶碗陪衬，高端奢华广告风格，Pollo.ai水印>",
+              "anchor": "→ 返回 商品主图",
+              "substance": "/表象/视觉/实体/物品/起居物品/美妆护肤",
+              "form": "/呈现/视觉/视觉制作/实景拍摄"
+            }
+          ],
+          "intent": "用 {via:AI生图工具} {act:元素生成} 从 {in-type:提示词} 生成 {out-type:成品图}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "主体生成"
+        }
+      ],
+      "return_row": {
+        "arrow": "↩",
+        "text": "<span class=\"kw\">返回</span> <span class=\"chip\" data-type=\"图像\">商品主图</span>"
+      },
+      "type_registry": {
+        "参数列表": {
+          "extends": "描述",
+          "desc": "从需求描述中提取出的结构化变量参数清单，列出可替换参数与固定元素，供后续提示词模板组合使用"
+        }
+      }
+    },
+    {
+      "id": "p2",
+      "name": "JSON结构化多区域布局出图",
+      "purpose": "用JSON对象精确定义多区域广告图的布局结构、各区域内容和文字标签，生成高精度多区域图像",
+      "category": "产物创造",
+      "platform": "zhihu",
+      "author": "白嫖知识分享",
+      "declarations": {
+        "inputs": [
+          {
+            "type": "文本",
+            "name": "布局需求描述",
+            "desc": "目标图像的区域数量、各区域主题、内容规格和文字标签需求"
+          }
+        ],
+        "resources": [],
+        "returns": {
+          "type": "图像"
+        }
+      },
+      "steps": [
+        {
+          "id": "s1",
+          "kind": "step",
+          "via": "human",
+          "action": "生成/关系生成/结构生成",
+          "feature": "人工",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "为每个区域明确标注：position（位置）、theme（主题）、subject（主体描述）、text_labels（画面内文字数组）；顶层用 type 字段说明整体图像格式，layout.structure 描述网格结构"
+            ],
+            [
+              "memo",
+              "AI模型对结构化输入的理解力远超松散描述；JSON强制理清布局逻辑，模型更容易「按图施工」；在即梦里生成多区域海报、产品展示板时，结构化描述效果会有质变；适用场景：多区域Banner、品牌设计板、UI界面生成、信息图布局"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "描述",
+              "name": "布局需求描述",
+              "value": "4格日系广告Banner（2x2网格），各象限主题分别为：旅行（冲绳旅行3日 39,800日元）、护肤（透明感水光肌 初回限定78%OFF 1,980日元）、美食（黑毛和牛A4/A5豪华牛排 期间限定4,980日元）、学习（在线资格讲座 10万人受讲 20%OFF）",
+              "anchor": "← 工序输入",
+              "substance": "/理念/事件/商业事件/推广活动",
+              "form": "/架构/修辞/语调风格/平实表达/直白陈述"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "提示词",
+              "name": "JSON结构化提示词",
+              "value": "{\n  \"type\": \"2x2 grid of Japanese digital advertisement banners\",\n  \"layout\": {\n    \"structure\": \"4 equal quadrants\",\n    \"quadrants\": [\n      {\n        \"position\": \"top-left\",\n        \"theme\": \"Travel\",\n        \"subject\": \"A couple holding hands on a white sand beach with turquoise water in Okinawa, bright tropical sky\",\n        \"text_labels\": [\"今年こそ、解き放て。\", \"沖縄旅行\", \"3日間の癒やし旅\", \"39,800円〜\", \"航空券+ホテル\", \"往復航空券\", \"人気ホテル宿泊\", \"レンタカー付き\", \"絶景、グルメ、体験ぜんぶ叶う！\"]\n      },\n      {\n        \"position\": \"top-right\",\n        \"theme\": \"Skincare\",\n        \"subject\": \"Close-up of a young woman's face applying glowing skincare cream, skin radiant and clear\",\n        \"text_labels\": [\"毛穴・くすみ卒業!\", \"透明感あふれる水光肌へ\", \"新感覚スキンケア\", \"毛穴ケア\", \"高保湿\", \"ハリ・ツヤ\", \"初回限定 78%OFF\", \"1,980円\"]\n      },\n      {\n        \"position\": \"bottom-left\",\n        \"theme\": \"Food\",\n        \"subject\": \"Premium A5 Wagyu beef steak sizzling on a grill, close-up, steam and fat marbling visible\",\n        \"text_labels\": [\"とろける旨さ!\", \"黒毛和牛 A4 A5等級 贅沢ステーキ\", \"期間限定 特別価格\", \"通価格8,980円 4,980円\"]\n      },\n      {\n        \"position\": \"bottom-right\",\n        \"theme\": \"Education\",\n        \"subject\": \"Young man studying efficiently with smartphone at a bright modern desk\",\n        \"text_labels\": [\"スキマ時間で最短合格!\", \"オンライン資格講座\", \"スマホで完結\", \"効率学習で差がつく!\", \"受講者数10万人突破!\", \"今だけ! 受講料20%OFF\"]\n      }\n    ]\n  }\n}",
+              "anchor": "→ s2",
+              "substance": "/理念/事件/商业事件/推广活动",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            }
+          ],
+          "intent": "用 {via:human} {act:结构生成} 从 {in-type:描述} 编写精确的 {out-type:提示词}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "预处理"
+        },
+        {
+          "id": "s2",
+          "kind": "step",
+          "via": "<AI生图工具>",
+          "action": "生成/元素生成",
+          "feature": "随机",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "{\n  \"type\": \"2x2 grid of Japanese digital advertisement banners\",\n  \"layout\": {\n    \"structure\": \"4 equal quadrants\",\n    \"quadrants\": [\n      {\"position\": \"top-left\", \"theme\": \"Travel\", \"subject\": \"A couple holding hands on a white sand beach with turquoise water in Okinawa, bright tropical sky\", \"text_labels\": [\"今年こそ、解き放て。\", \"沖縄旅行\", \"3日間の癒やし旅\", \"39,800円〜\", \"航空券+ホテル\", \"往復航空券\", \"人気ホテル宿泊\", \"レンタカー付き\", \"絶景、グルメ、体験ぜんぶ叶う！\"]},\n      {\"position\": \"top-right\", \"theme\": \"Skincare\", \"subject\": \"Close-up of a young woman's face applying glowing skincare cream, skin radiant and clear\", \"text_labels\": [\"毛穴・くすみ卒業!\", \"透明感あふれる水光肌へ\", \"初回限定 78%OFF\", \"1,980円\"]},\n      {\"position\": \"bottom-left\", \"theme\": \"Food\", \"subject\": \"Premium A5 Wagyu beef steak sizzling on a grill, close-up, steam and fat marbling visible\", \"text_labels\": [\"黒毛和牛 A4 A5等級 贅沢ステーキ\", \"期間限定 特別価格\", \"4,980円\"]},\n      {\"position\": \"bottom-right\", \"theme\": \"Education\", \"subject\": \"Young man studying efficiently with smartphone at a bright modern desk\", \"text_labels\": [\"スキマ時間で最短合格!\", \"受講者数10万人突破!\", \"受講料20%OFF\"]}\n    ]\n  }\n}"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "提示词",
+              "name": "JSON结构化提示词",
+              "value": "{\n  \"type\": \"2x2 grid of Japanese digital advertisement banners\",\n  \"layout\": {\n    \"structure\": \"4 equal quadrants\",\n    \"quadrants\": [\n      {\"position\": \"top-left\", \"theme\": \"Travel\", \"subject\": \"A couple holding hands on a white sand beach with turquoise water in Okinawa, bright tropical sky\", \"text_labels\": [\"今年こそ、解き放て。\", \"沖縄旅行\", \"3日間の癒やし旅\", \"39,800円〜\", \"航空券+ホテル\", \"往復航空券\", \"人気ホテル宿泊\", \"レンタカー付き\", \"絶景、グルメ、体験ぜんぶ叶う！\"]},\n      {\"position\": \"top-right\", \"theme\": \"Skincare\", \"subject\": \"Close-up of a young woman's face applying glowing skincare cream, skin radiant and clear\", \"text_labels\": [\"毛穴・くすみ卒業!\", \"透明感あふれる水光肌へ\", \"初回限定 78%OFF\", \"1,980円\"]},\n      {\"position\": \"bottom-left\", \"theme\": \"Food\", \"subject\": \"Premium A5 Wagyu beef steak sizzling on a grill, close-up, steam and fat marbling visible\", \"text_labels\": [\"黒毛和牛 A4 A5等級 贅沢ステーキ\", \"期間限定 特別価格\", \"4,980円\"]},\n      {\"position\": \"bottom-right\", \"theme\": \"Education\", \"subject\": \"Young man studying efficiently with smartphone at a bright modern desk\", \"text_labels\": [\"スキマ時間で最短合格!\", \"受講者数10万人突破!\", \"受講料20%OFF\"]}\n    ]\n  }\n}",
+              "anchor": "← s1.JSON结构化提示词",
+              "substance": "/理念/事件/商业事件/推广活动",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "成品图",
+              "name": "多区域广告图",
+              "value": "<AI生成的4格日系广告Banner：左上冲绳旅行（海滩情侣背影+日文旅行信息+价格39,800円），右上护肤品（水光肌特写+品牌LUMIÈRE+折扣信息1,980円），左下黑毛和牛牛排（特写冒烟+价格4,980円），右下在线资格讲座（学习男生+10万人突破+20%OFF）>",
+              "anchor": "→ 返回 多区域广告图",
+              "substance": "/理念/事件/商业事件/推广活动",
+              "form": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"
+            }
+          ],
+          "intent": "用 {via:AI生图工具} {act:元素生成} 将 {in-type:提示词} 渲染为精确布局的 {out-type:成品图}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "主体生成"
+        }
+      ],
+      "return_row": {
+        "arrow": "↩",
+        "text": "<span class=\"kw\">返回</span> <span class=\"chip\" data-type=\"图像\">多区域广告图</span>"
+      }
+    },
+    {
+      "id": "p3",
+      "name": "摄影指令出图——便利店霓虹人像",
+      "purpose": "用专业摄影术语组合（胶片类型+光线+构图+后期风格）构建摄影指令提示词，生成便利店霓虹灯电影感人像",
+      "category": "产物创造",
+      "platform": "zhihu",
+      "author": "白嫖知识分享",
+      "declarations": {
+        "inputs": [
+          {
+            "type": "文本",
+            "name": "拍摄场景描述",
+            "desc": "目标人像的场景氛围、人物特征、情绪调性（口语化描述，待转换为专业术语）"
+          }
+        ],
+        "resources": [],
+        "returns": {
+          "type": "图像"
+        }
+      },
+      "steps": [
+        {
+          "id": "s1",
+          "kind": "step",
+          "via": "human",
+          "action": "提取/化学提取/解构",
+          "feature": "人工",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "按公式拆解摄影五要素：{胶片类型/拍摄格式} + {光线描述} + {人物描述} + {构图/景别} + {后期风格}；将口语化描述替换为专业摄影术语（如「拍一个好看女孩」→「35mm film photography / medium shot / authentic film grain」）"
+            ],
+            [
+              "memo",
+              "AI绘图模型在训练数据中见过大量带EXIF标签的摄影作品；使用35mm film / cinematic lighting / editorial style等术语，等于调取模型「专业摄影」的深层表征；不要写「拍一个漂亮的便利店女孩」，要写摄影行话"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "描述",
+              "name": "拍摄场景描述",
+              "value": "便利店夜晚场景，女性中景，电影感街拍风格，有霓虹灯光效，胶片质感",
+              "anchor": "← 工序输入",
+              "substance": "/表象/视觉/空间/空间场所/生活场景/商业场所",
+              "form": "/架构/修辞/语调风格/平实表达/直白陈述"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "提示词",
+              "name": "摄影要素清单",
+              "value": "胶片格式: 35mm film photography / 光线: harsh convenience store fluorescent lighting mixed with colorful neon signs from outside / 质感: authentic film grain, high contrast, slight color cast / 构图景别: intimate medium shot / 后期风格: cinematic street editorial style",
+              "anchor": "→ s2",
+              "substance": "/表象/视觉/实体/物品/影像物品/艺术创作",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/逐条罗列/通用列举"
+            }
+          ],
+          "intent": "用 {via:human} {act:解构} 从 {in-type:描述} 提炼五要素得到 {out-type:提示词}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "预处理"
+        },
+        {
+          "id": "s2",
+          "kind": "step",
+          "via": "human",
+          "action": "生成/关系生成/结构生成",
+          "feature": "人工",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "按{胶片类型} + {光线} + {人物描述} + {构图/景别} + {后期风格}顺序拼接各要素；所有风格参数用英文专业术语；人物描述可用中英混合"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "提示词",
+              "name": "摄影要素清单",
+              "value": "胶片格式: 35mm film photography / 光线: harsh convenience store fluorescent lighting mixed with colorful neon signs from outside / 质感: authentic film grain, high contrast, slight color cast / 构图景别: intimate medium shot / 后期风格: cinematic street editorial style",
+              "anchor": "← s1.摄影要素清单",
+              "substance": "/表象/视觉/实体/物品/影像物品/艺术创作",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/逐条罗列/通用列举"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "提示词",
+              "name": "摄影指令提示词",
+              "value": "35mm film photography with harsh convenience store fluorescent lighting mixed with colorful neon signs from outside, authentic film grain, high contrast, slight color cast, cinematic street editorial style, intimate medium shot, young Asian woman in casual white button-down shirt and black mini skirt, leaning against glass refrigerator door, holding a drink bottle, messy bun hairstyle, natural makeup, photorealistic",
+              "anchor": "→ s3",
+              "substance": "/表象/视觉/实体/人物/人物特征/形象呈现",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            }
+          ],
+          "intent": "用 {via:human} {act:结构生成} 将 {in-type:提示词} 拼接为完整摄影指令 {out-type:提示词}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "预处理"
+        },
+        {
+          "id": "s3",
+          "kind": "step",
+          "via": "<AI生图工具>",
+          "action": "生成/元素生成",
+          "feature": "随机",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "35mm film photography with harsh convenience store fluorescent lighting mixed with colorful neon signs from outside, authentic film grain, high contrast, slight color cast, cinematic street editorial style, intimate medium shot, young Asian woman in casual white button-down shirt and black mini skirt, leaning against glass refrigerator door, holding a drink bottle, messy bun hairstyle, natural makeup, photorealistic"
+            ],
+            [
+              "memo",
+              "即梦3.0的文字增强和智能参考功能对此类结构化摄影描述响应更好；适用场景：人像摄影、产品摄影、氛围感场景图"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "提示词",
+              "name": "摄影指令提示词",
+              "value": "35mm film photography with harsh convenience store fluorescent lighting mixed with colorful neon signs from outside, authentic film grain, high contrast, slight color cast, cinematic street editorial style, intimate medium shot, young Asian woman in casual white button-down shirt and black mini skirt, leaning against glass refrigerator door, holding a drink bottle, messy bun hairstyle, natural makeup, photorealistic",
+              "anchor": "← s2.摄影指令提示词",
+              "substance": "/表象/视觉/实体/人物/人物特征/形象呈现",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "成品图",
+              "name": "电影感人像照",
+              "value": "<AI生成的便利店霓虹人像：年轻亚裔女性，白色半透明衬衫+黑色短裙，倚着便利店冰柜玻璃门，手持棕色饮料瓶，荧光灯+霓虹灯混合光效，35mm胶片颗粒感，高对比度电影街拍风格>",
+              "anchor": "→ 返回 电影感人像照",
+              "substance": "/表象/视觉/实体/人物/人物特征/形象呈现",
+              "form": "/呈现/视觉/视觉风格/视觉艺术风格/影视风格/胶片摄影"
+            }
+          ],
+          "intent": "用 {via:AI生图工具} {act:元素生成} 从 {in-type:提示词} 生成 {out-type:成品图}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "主体生成"
+        }
+      ],
+      "return_row": {
+        "arrow": "↩",
+        "text": "<span class=\"kw\">返回</span> <span class=\"chip\" data-type=\"图像\">电影感人像照</span>"
+      }
+    },
+    {
+      "id": "p4",
+      "name": "摄影指令出图——温泉旅馆人像",
+      "purpose": "用专业摄影术语组合构建摄影指令提示词，生成日式温泉旅馆（onsen ryokan）暖灯氛围editorial人像",
+      "category": "产物创造",
+      "platform": "zhihu",
+      "author": "白嫖知识分享",
+      "declarations": {
+        "inputs": [
+          {
+            "type": "文本",
+            "name": "拍摄场景描述",
+            "desc": "温泉旅馆氛围人像需求：暖色灯光、和风场景、editorial风格"
+          }
+        ],
+        "resources": [],
+        "returns": {
+          "type": "图像"
+        }
+      },
+      "steps": [
+        {
+          "id": "s1",
+          "kind": "step",
+          "via": "human",
+          "action": "提取/化学提取/解构",
+          "feature": "人工",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "按公式拆解摄影五要素：{胶片类型/拍摄格式} + {光线描述} + {人物描述} + {构图/景别} + {后期风格}；温泉场景的光线描述重点：warm vintage aesthetic + wooden lantern lighting + natural window light"
+            ],
+            [
+              "memo",
+              "与便利店场景的区别在于光线（harsh fluorescent → warm lantern）和整体美学基调（street editorial → vintage Japanese aesthetic）；切换场景只需替换光线+美学两要素，其余框架不变"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "描述",
+              "name": "拍摄场景描述",
+              "value": "日式温泉旅馆（onsen ryokan）人像，暖色木质灯笼光源，editorial高大气感风格，胶片质感，女性",
+              "anchor": "← 工序输入",
+              "substance": "/表象/视觉/空间/空间场所",
+              "form": "/架构/修辞/语调风格/平实表达/直白陈述"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "提示词",
+              "name": "摄影要素清单",
+              "value": "胶片格式: 35mm film photography / 光线: warm vintage Japanese onsen ryokan aesthetic, soft ambient wooden lantern lighting mixed with gentle natural window light / 质感: subtle film grain, gentle color shift / 构图景别: high atmosphere editorial style / 后期风格: vintage Japanese aesthetic",
+              "anchor": "→ s2",
+              "substance": "/表象/视觉/实体/物品/影像物品/艺术创作",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/逐条罗列/通用列举"
+            }
+          ],
+          "intent": "用 {via:human} {act:解构} 从 {in-type:描述} 提炼温泉场景五要素 {out-type:提示词}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "预处理"
+        },
+        {
+          "id": "s2",
+          "kind": "step",
+          "via": "human",
+          "action": "生成/关系生成/结构生成",
+          "feature": "人工",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "按五要素顺序拼接：35mm film photography + warm vintage Japanese onsen ryokan aesthetic + 人物描述 + editorial style + 质感参数"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "提示词",
+              "name": "摄影要素清单",
+              "value": "胶片格式: 35mm film photography / 光线: warm vintage Japanese onsen ryokan aesthetic, soft ambient wooden lantern lighting mixed with gentle natural window light / 质感: subtle film grain, gentle color shift / 构图景别: high atmosphere editorial style / 后期风格: vintage Japanese aesthetic",
+              "anchor": "← s1.摄影要素清单",
+              "substance": "/表象/视觉/实体/物品/影像物品/艺术创作",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/逐条罗列/通用列举"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "提示词",
+              "name": "摄影指令提示词",
+              "value": "35mm film photography, warm vintage Japanese onsen ryokan aesthetic, soft ambient wooden lantern lighting mixed with gentle natural window light, subtle film grain, gentle color shift, high atmosphere editorial style, young Asian woman in traditional Japanese yukata robe, sitting gracefully on wooden veranda, hot spring visible in background, photorealistic",
+              "anchor": "→ s3",
+              "substance": "/表象/视觉/实体/人物/人物特征/形象呈现",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            }
+          ],
+          "intent": "用 {via:human} {act:结构生成} 将温泉场景 {in-type:提示词} 拼接为完整 {out-type:提示词}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "预处理"
+        },
+        {
+          "id": "s3",
+          "kind": "step",
+          "via": "<AI生图工具>",
+          "action": "生成/元素生成",
+          "feature": "随机",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "35mm film photography, warm vintage Japanese onsen ryokan aesthetic, soft ambient wooden lantern lighting mixed with gentle natural window light, subtle film grain, gentle color shift, high atmosphere editorial style, young Asian woman in traditional Japanese yukata robe, sitting gracefully on wooden veranda, hot spring visible in background, photorealistic"
+            ],
+            [
+              "memo",
+              "温泉旅馆场景与便利店场景属同一公式（摄影指令）的不同应用实例；两者提示词框架相同，只替换光线描述和场景美学方向"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "提示词",
+              "name": "摄影指令提示词",
+              "value": "35mm film photography, warm vintage Japanese onsen ryokan aesthetic, soft ambient wooden lantern lighting mixed with gentle natural window light, subtle film grain, gentle color shift, high atmosphere editorial style, young Asian woman in traditional Japanese yukata robe, sitting gracefully on wooden veranda, hot spring visible in background, photorealistic",
+              "anchor": "← s2.摄影指令提示词",
+              "substance": "/表象/视觉/实体/人物/人物特征/形象呈现",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "成品图",
+              "name": "温泉旅馆人像",
+              "value": "<AI生成的温泉旅馆editorial人像：年轻亚裔女性，浅色宽松和风浴衣，坐在木质台阶上，背景可见温泉热水，室内木质灯笼暖光，胶片颗粒感，高大气感editorial风格>",
+              "anchor": "→ 返回 温泉旅馆人像",
+              "substance": "/表象/视觉/实体/人物/人物特征/形象呈现",
+              "form": "/呈现/视觉/视觉风格/视觉艺术风格/影视风格/胶片摄影"
+            }
+          ],
+          "intent": "用 {via:AI生图工具} {act:元素生成} 从 {in-type:提示词} 生成 {out-type:成品图}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "主体生成"
+        }
+      ],
+      "return_row": {
+        "arrow": "↩",
+        "text": "<span class=\"kw\">返回</span> <span class=\"chip\" data-type=\"图像\">温泉旅馆人像</span>"
+      }
+    },
+    {
+      "id": "p5",
+      "name": "分步组合精修商品图",
+      "purpose": "先用提示词生成基础构图，再基于基础图叠加品牌细节/光影/背景虚化，两步走生成精确可控的复杂商品图",
+      "category": "产物创造",
+      "platform": "zhihu",
+      "author": "白嫖知识分享",
+      "declarations": {
+        "inputs": [
+          {
+            "type": "文本",
+            "name": "商品描述",
+            "desc": "目标商品的外观、品牌要求和整体视觉风格"
+          },
+          {
+            "type": "文本",
+            "name": "品牌细节规格",
+            "desc": "需要精确叠加的品牌标识、光影效果、背景虚化等精细要求"
+          }
+        ],
+        "resources": [],
+        "returns": {
+          "type": "图像"
+        }
+      },
+      "steps": [
+        {
+          "id": "s1",
+          "kind": "step",
+          "via": "human",
+          "action": "生成/关系生成/结构生成",
+          "feature": "人工",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "只描述基础构图和主体形态（产品形状/摆放/基础光线/构图比例），不写品牌标识和精细细节；确保基础构图到位（主次分明、比例正确、光线方向确定）"
+            ],
+            [
+              "memo",
+              "分步法的核心：一步到位的提示词控制力不够；先生成基础画面，再叠加细节，两步走比一步冲更可控；即梦的「智能参考」功能天然支持这种工作流"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "描述",
+              "name": "商品描述",
+              "value": "汉堡产品图，需要精确放置品牌标识、增强光影效果、添加背景虚化",
+              "anchor": "← 工序输入",
+              "substance": "/表象/视觉/实体/物品/起居物品/食物",
+              "form": "/架构/修辞/语调风格/平实表达/直白陈述"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "提示词",
+              "name": "基础构图提示词",
+              "value": "<基础构图提示词：product photography of a gourmet burger, centered composition, layers clearly visible (bun / lettuce / tomato / beef patty / cheese), studio lighting from top-left, white/neutral background, sharp focus on burger>",
+              "inferred": true,
+              "inferred_reason": "原文公式四只描述「第一步生成主体汉堡构图」，未给出具体提示词文本；根据分步法工艺规范和汉堡商品图常见提示词推断合理形式",
+              "anchor": "→ s2",
+              "substance": "/表象/视觉/实体/物品/起居物品/食物",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            }
+          ],
+          "intent": "用 {via:human} {act:结构生成} 从 {in-type:描述} 撰写聚焦主体构图的 {out-type:提示词}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "预处理"
+        },
+        {
+          "id": "s2",
+          "kind": "step",
+          "via": "<AI生图工具>",
+          "action": "生成/元素生成",
+          "feature": "随机",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "<基础构图提示词：product photography of a gourmet burger, centered composition, layers clearly visible (bun / lettuce / tomato / beef patty / cheese), studio lighting from top-left, white/neutral background, sharp focus on burger>"
+            ],
+            [
+              "memo",
+              "此directive内容为推断值（原文公式四未给出具体提示词文本），从上游s1输出透传而来"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "提示词",
+              "name": "基础构图提示词",
+              "value": "<基础构图提示词：product photography of a gourmet burger, centered composition, layers clearly visible (bun / lettuce / tomato / beef patty / cheese), studio lighting from top-left, white/neutral background, sharp focus on burger>",
+              "inferred": true,
+              "inferred_reason": "原文公式四第一步未给出具体提示词；根据工艺推断",
+              "anchor": "← s1.基础构图提示词",
+              "substance": "/表象/视觉/实体/物品/起居物品/食物",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "底图",
+              "name": "基础汉堡图",
+              "value": "<AI生成的基础汉堡构图：汉堡居中，食材层次分明，基础光线和构图到位，无品牌标识，白色/中性背景>",
+              "inferred": true,
+              "inferred_reason": "原文有「出图效果」描述但未具体展示此中间步骤的图像内容；根据工艺推断中间产物形态",
+              "anchor": "→ s3, s4",
+              "substance": "/表象/视觉/实体/物品/起居物品/食物",
+              "form": "/呈现/视觉/视觉制作/实景拍摄"
+            }
+          ],
+          "intent": "用 {via:AI生图工具} {act:元素生成} 从 {in-type:提示词} 生成主体 {out-type:底图}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "主体生成"
+        },
+        {
+          "id": "s3",
+          "kind": "step",
+          "via": "human",
+          "action": "生成/关系生成/结构生成",
+          "feature": "人工",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "以基础图为参考，描述需要叠加的细节：品牌标识位置和样式、光影强化方向、背景虚化程度；在即梦中使用「智能参考」上传基础图，再添加叠加提示词"
+            ],
+            [
+              "memo",
+              "「基于参考图做风格迁移或细节修改」是即梦3.0智能参考功能的核心能力；复杂创意不要指望一条提示词搞定所有事"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "底图",
+              "name": "基础汉堡图",
+              "value": "<AI生成的基础汉堡构图：汉堡居中，食材层次分明，基础光线和构图到位，无品牌标识，白色/中性背景>",
+              "inferred": true,
+              "inferred_reason": "从上游s2输出透传而来",
+              "anchor": "← s2.基础汉堡图",
+              "substance": "/表象/视觉/实体/物品/起居物品/食物",
+              "form": "/呈现/视觉/视觉制作/实景拍摄"
+            },
+            {
+              "type": "描述",
+              "name": "品牌细节规格",
+              "value": "<品牌标识规格：logo位置（左上角）、光影强化（温暖侧光）、背景虚化（大光圈bokeh）>",
+              "inferred": true,
+              "inferred_reason": "原文提到「叠加品牌标识、光影效果、背景虚化」但未给出具体规格；根据汉堡商品图常见品牌设计推断",
+              "anchor": "← 工序输入",
+              "substance": "/表象/视觉/符号/文字符号/水印",
+              "form": "/架构/修辞/语调风格/平实表达/直白陈述"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "提示词",
+              "name": "叠加效果提示词",
+              "value": "<叠加提示词：based on this burger image, add brand logo in top-left corner, enhance warm side lighting with golden highlights, add bokeh background effect, increase food texture detail and steam>",
+              "inferred": true,
+              "inferred_reason": "原文仅描述工艺思路（叠加品牌标识+光影+背景虚化），未给出叠加步骤的具体提示词",
+              "anchor": "→ s4",
+              "substance": "/表象/视觉/实体/物品/起居物品/食物",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            }
+          ],
+          "intent": "用 {via:human} {act:结构生成} 从 {in-type:底图} 和 {in-type:描述} 撰写 {out-type:提示词}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "in-type-1",
+            "out-type-0"
+          ],
+          "effect": "预处理"
+        },
+        {
+          "id": "s4",
+          "kind": "step",
+          "via": "<AI生图工具>",
+          "action": "生成/元素生成",
+          "feature": "随机",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "<叠加提示词：based on this burger image, add brand logo in top-left corner, enhance warm side lighting with golden highlights, add bokeh background effect, increase food texture detail and steam>"
+            ],
+            [
+              "memo",
+              "此directive内容为推断值（叠加提示词由上游s3推断生成）；即梦「智能参考」模式：上传基础汉堡图为参考，叠加提示词描述新增内容；适用场景：复杂产品图、需要精确品牌元素的创意图、多轮迭代优化"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "提示词",
+              "name": "叠加效果提示词",
+              "value": "<叠加提示词：based on this burger image, add brand logo in top-left corner, enhance warm side lighting with golden highlights, add bokeh background effect, increase food texture detail and steam>",
+              "inferred": true,
+              "inferred_reason": "由上游s3推断生成",
+              "anchor": "← s3.叠加效果提示词",
+              "substance": "/表象/视觉/实体/物品/起居物品/食物",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            },
+            {
+              "type": "底图",
+              "name": "基础汉堡图",
+              "value": "<AI生成的基础汉堡构图：汉堡居中，食材层次分明，基础光线和构图到位，无品牌标识>",
+              "inferred": true,
+              "inferred_reason": "从上游s2输出透传而来",
+              "anchor": "← s2.基础汉堡图",
+              "substance": "/表象/视觉/实体/物品/起居物品/食物",
+              "form": "/呈现/视觉/视觉制作/实景拍摄"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "成品图",
+              "name": "精修商品图",
+              "value": "<最终精修的汉堡产品图：品牌标识精确放置，温暖侧光增强，背景自然虚化，食材纹理和蒸汽细节丰富，广告级商品图质量>",
+              "anchor": "→ 返回 精修商品图",
+              "substance": "/表象/视觉/实体/物品/起居物品/食物",
+              "form": "/呈现/视觉/视觉制作/实景拍摄"
+            }
+          ],
+          "intent": "用 {via:AI生图工具} {act:元素生成} 按 {in-type:提示词} 对 {in-type:底图} 叠加细节生成 {out-type:成品图}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "in-type-1",
+            "out-type-0"
+          ],
+          "effect": "主体生成"
+        }
+      ],
+      "return_row": {
+        "arrow": "↩",
+        "text": "<span class=\"kw\">返回</span> <span class=\"chip\" data-type=\"图像\">精修商品图</span>"
+      }
+    },
+    {
+      "id": "p6",
+      "name": "参考图驱动——实拍转动漫截图",
+      "purpose": "上传实拍照片作为参考图，用极简一句话转换指令驱动AI生成动漫截图风格图像，消除语言描述与视觉想象间的信息损耗",
+      "category": "产物创造",
+      "platform": "zhihu",
+      "author": "白嫖知识分享",
+      "declarations": {
+        "inputs": [
+          {
+            "type": "参考图",
+            "name": "参考图",
+            "desc": "待转换的实拍照片（人物照）；作为风格迁移的视觉输入源"
+          },
+          {
+            "type": "文本",
+            "name": "转换风格描述",
+            "desc": "目标风格的简短描述（动漫截图、二次元风格等）"
+          }
+        ],
+        "resources": [],
+        "returns": {
+          "type": "图像"
+        }
+      },
+      "steps": [
+        {
+          "id": "s1",
+          "kind": "step",
+          "via": "human",
+          "action": "生成/关系生成/结构生成",
+          "feature": "人工",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "一句话转换指令：直接描述目标风格输出形式；极简指令往往效果更好；避免过度描述细节，以免干扰AI对参考图的理解"
+            ],
+            [
+              "memo",
+              "参考图消除了「语言描述和视觉想象之间的信息损耗」；即梦3.0的智能参考功能——上传参考图后，AI从图像中提取构图、色调、风格信息，再按指令做定向修改；这是文字提示词无法替代的能力"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "描述",
+              "name": "转换风格描述",
+              "value": "将图中人物转换为真实动画(anime)的截图风格",
+              "anchor": "← 工序输入",
+              "substance": "/表象/视觉/实体/人物/人物特征/形象呈现",
+              "form": "/架构/修辞/语调风格/平实表达/直白陈述"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "提示词",
+              "name": "转换指令提示词",
+              "value": "Show me the attached image as a snapshot from an actual anime",
+              "anchor": "→ s2",
+              "substance": "/表象/视觉/实体/人物/人物特征/形象呈现",
+              "form": "/架构/修辞/语调风格/平实表达/直白陈述"
+            }
+          ],
+          "intent": "用 {via:human} {act:结构生成} 从 {in-type:描述} 撰写极简 {out-type:提示词}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "预处理"
+        },
+        {
+          "id": "s2",
+          "kind": "step",
+          "via": "<AI生图工具>",
+          "action": "生成/元素生成",
+          "feature": "随机",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "Show me the attached image as a snapshot from an actual anime"
+            ],
+            [
+              "memo",
+              "输入：参考图（实拍人物照）+ 转换指令；AI从参考图提取构图、色调、角色特征，输出动漫截图风格；即梦3.0智能参考模式天然支持此工作流；适用场景：风格转换（实拍转动漫）、角色设定资料、基于品牌素材的再创作"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "参考图",
+              "name": "参考图",
+              "value": "<实拍照片：人物实拍照（女性），用于转换为动漫风格截图>",
+              "anchor": "← 工序输入",
+              "substance": "/表象/视觉/实体/人物/人物特征/形象呈现",
+              "form": "/呈现/视觉/视觉制作/实景拍摄"
+            },
+            {
+              "type": "提示词",
+              "name": "转换指令提示词",
+              "value": "Show me the attached image as a snapshot from an actual anime",
+              "anchor": "← s1.转换指令提示词",
+              "substance": "/表象/视觉/实体/人物/人物特征/形象呈现",
+              "form": "/架构/修辞/语调风格/平实表达/直白陈述"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "成品图",
+              "name": "动漫风格图",
+              "value": "<AI生成的动漫截图风格图：女性角色，黑色长发，深色大眼，哥特风黑色洋装+蕾丝领，泪眼含泪，深蓝色暗调动漫画风，呈现为「实际动漫」截图质感>",
+              "anchor": "→ 返回 动漫风格图",
+              "substance": "/表象/视觉/实体/人物/人物特征/形象呈现",
+              "form": "/呈现/视觉/视觉风格/视觉艺术风格/绘画风格/卡通"
+            }
+          ],
+          "intent": "用 {via:AI生图工具} {act:元素生成} 将 {in-type:参考图} 按 {in-type:提示词} 转换为 {out-type:成品图}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "in-type-1",
+            "out-type-0"
+          ],
+          "effect": "主体生成"
+        }
+      ],
+      "return_row": {
+        "arrow": "↩",
+        "text": "<span class=\"kw\">返回</span> <span class=\"chip\" data-type=\"图像\">动漫风格图</span>"
+      }
+    },
+    {
+      "id": "p7",
+      "name": "参考图驱动——角色设定资料生成",
+      "purpose": "基于已有角色图，用日文提示词生成Persona5风格角色设定卡（三视图+表情差分+服装分解+色板+装备说明）",
+      "category": "产物创造",
+      "platform": "zhihu",
+      "author": "白嫖知识分享",
+      "declarations": {
+        "inputs": [
+          {
+            "type": "参考图",
+            "name": "角色参考图",
+            "desc": "已有角色图或人物图，作为设定资料生成的视觉基础"
+          },
+          {
+            "type": "文本",
+            "name": "设定资料规格",
+            "desc": "角色设定资料的内容要求（三视图/表情差分/服装分解/色板等）"
+          }
+        ],
+        "resources": [],
+        "returns": {
+          "type": "图像"
+        }
+      },
+      "steps": [
+        {
+          "id": "s1",
+          "kind": "step",
+          "via": "human",
+          "action": "生成/关系生成/结构生成",
+          "feature": "人工",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "用日文撰写角色设定资料提示词，指定设定卡格式要求（三视图/表情差分/服装分解/色板）；日文「設定資料」「キャラクターシート」等术语直接触发特定的专业动画/游戏风格"
+            ],
+            [
+              "memo",
+              "日文提示词在特定风格（设计资料、动漫角色设定）上有语义优势——「設定資料」直接命中模型训练数据中的专业动画设定资料分类；适用场景：角色设定资料、基于品牌素材的再创作；纯日文提示词同样能生成高质量图像"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "描述",
+              "name": "设定资料规格",
+              "value": "Persona5风格角色设定卡，需要：正面/侧面/背面三视图、表情差分、服装分解、色板；基于已有角色图生成",
+              "anchor": "← 工序输入",
+              "substance": "/表象/视觉/实体/人物/人物特征/形象呈现",
+              "form": "/架构/修辞/语调风格/平实表达/直白陈述"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "提示词",
+              "name": "角色设定提示词",
+              "value": "このキャラクターと背景を元に、公式設定資料のようなキャラクターシートを作成してください。・正面、側面、背面の3面図を含める・表情差分（通常・笑顔・驚き・怒り）を追加・服装の各パーツを分解して詳細に描画・カラーパレットと素材の説明を含める・Persona5スタイルのグラフィックデザインで構成",
+              "anchor": "→ s2",
+              "substance": "/表象/视觉/实体/人物/人物特征/形象呈现",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            }
+          ],
+          "intent": "用 {via:human} {act:结构生成} 从 {in-type:描述} 撰写日文 {out-type:提示词}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "预处理"
+        },
+        {
+          "id": "s2",
+          "kind": "step",
+          "via": "<AI生图工具>",
+          "action": "生成/元素生成",
+          "feature": "随机",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "このキャラクターと背景を元に、公式設定資料のようなキャラクターシートを作成してください。・正面、側面、背面の3面図を含める・表情差分（通常・笑顔・驚き・怒り）を追加・服装の各パーツを分解して詳細に描画・カラーパレットと素材の説明を含める・Persona5スタイルのグラフィックデザインで構成"
+            ],
+            [
+              "memo",
+              "一条提示词完成整套设定资料；GPT-Image-2的CJK优化让日文提示词不再「打酱油」；即梦对日文描述同样有良好支持"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "参考图",
+              "name": "角色参考图",
+              "value": "<已有角色图：女性战士角色（ソルジャンヌ），红黑色装甲战衣，蓝色能量核心，用于生成设定资料>",
+              "anchor": "← 工序输入",
+              "substance": "/表象/视觉/实体/人物/个体/虚构人物",
+              "form": "/呈现/视觉/视觉风格/视觉艺术风格/绘画风格/卡通"
+            },
+            {
+              "type": "提示词",
+              "name": "角色设定提示词",
+              "value": "このキャラクターと背景を元に、公式設定資料のようなキャラクターシートを作成してください。・正面、側面、背面の3面図を含める・表情差分（通常・笑顔・驚き・怒り）を追加・服装の各パーツを分解して詳細に描画・カラーパレットと素材の説明を含める・Persona5スタイルのグラフィックデザインで構成",
+              "anchor": "← s1.角色设定提示词",
+              "substance": "/表象/视觉/实体/人物/人物特征/形象呈现",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "成品图",
+              "name": "角色设定卡",
+              "value": "<AI生成的ソルジャンヌ・スーツ角色设定卡：6格展示手动装备战衣的完整流程（1.确认内装+传感器 → 2.胸部肩部装甲 → 3.腰部ユニット固定 → 4.头盔准备 → 5.头盔装着+系统启动 → 6.装备完成），每格附日文说明，整体呈现Persona5风格官方设定资料格式>",
+              "anchor": "→ 返回 角色设定卡",
+              "substance": "/表象/视觉/实体/人物/人物特征/形象呈现",
+              "form": "/呈现/视觉/视觉风格/视觉艺术风格/绘画风格/卡通"
+            }
+          ],
+          "intent": "用 {via:AI生图工具} {act:元素生成} 从 {in-type:参考图} 和 {in-type:提示词} 生成 {out-type:成品图}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "in-type-1",
+            "out-type-0"
+          ],
+          "effect": "主体生成"
+        }
+      ],
+      "return_row": {
+        "arrow": "↩",
+        "text": "<span class=\"kw\">返回</span> <span class=\"chip\" data-type=\"图像\">角色设定卡</span>"
+      }
+    },
+    {
+      "id": "p8",
+      "name": "多语言混用本地化内容出图",
+      "purpose": "中文描述内容细节、英文锁定风格参数、日文匹配风格标签，三语协同发挥各自语义优势，生成高精度本地化内容图",
+      "category": "产物创造",
+      "platform": "zhihu",
+      "author": "白嫖知识分享",
+      "declarations": {
+        "inputs": [
+          {
+            "type": "文本",
+            "name": "内容主题描述",
+            "desc": "图像主题、内容要素、风格方向（口语化描述，待转换为多语言结构化提示词）"
+          }
+        ],
+        "resources": [],
+        "returns": {
+          "type": "图像"
+        }
+      },
+      "steps": [
+        {
+          "id": "s1",
+          "kind": "step",
+          "via": "human",
+          "action": "生成/关系生成/结构生成",
+          "feature": "人工",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "按语言分工原则构建提示词：中文负责场景细节（如「九宫格锅翻滚冒泡」「串串香一把竹签插着各种食材冒着热气」）；英文锁定风格参数（如「watercolor illustration」「warm color palette」）；日文匹配特定风格标签（如「設定資料」）；各语言在各自擅长的维度上配合"
+            ],
+            [
+              "memo",
+              "即梦的中文语义理解在国产工具中是第一梯队的；GPT-Image-2的CJK优化让中文提示词不再「打酱油」；混合使用时出图精度远超单语言提示词；中文描述场景细节（「九宫格锅翻滚冒泡」）、英文锁定风格参数（「watercolor illustration」）、各语言在各自擅长的维度上配合"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "描述",
+              "name": "内容主题描述",
+              "value": "成都美食地图，手绘风格，鸟瞰视角，包含12个知名美食地点（串串香、红油水饺、三大炮、盖碗茶、冰粉、小龙坎火锅、钵钵鸡等），标注地标建筑和主要道路，整体暖色系手绘质感",
+              "anchor": "← 工序输入",
+              "substance": [
+                "/表象/视觉/实体/物品/起居物品/食物",
+                "/表象/视觉/实体/物品/器物物品/道具"
+              ],
+              "form": "/架构/修辞/语调风格/平实表达/直白陈述"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "提示词",
+              "name": "多语言混合提示词",
+              "value": "一张手绘风格的城市美食地图，以成都为主题。画面以鸟瞰视角的手绘简化城市地图为底，标注主要道路和地标但不追求精确比例而是追求可爱的手绘感。地图上分布着12个美食地点的精致手绘小插画：春熙路的串串香（一把竹签插着各种食材冒着热气）、宽窄巷子的三大炮（三个糯米团子飞向铜盘）、玉林路的小龙坎火锅（九宫格锅翻滚冒泡，越煮越巴适）、东郊记忆的冒椒麻辣（冒菜要冒到头），武侯祠大街的钵钵鸡（冷锅串串）、香港巷的肥肠粉（红薯粉Q弹），人民公园鹤鸣茶社的盖碗茶（茶碗与茶盖），锦里古街的冰粉（手握冰粉碗），双流老妈兔头（麻辣入味），建设路叶婆婆烘蛋糕（外酥里软），钟水饺（总店，红油水饺），陈麻婆豆腐（麻辣鲜香）。左上角标题「成都·吃货暴走地图」使用胖圆的手绘美术字配辣椒装饰。整体画风为水彩+彩铅混合的手绘质感，颜色以暖色系为主，边框用辣椒和植物装饰，右下角有图例说明和指南针。watercolor illustration, handmade texture, warm color palette, cozy illustrated map style",
+              "anchor": "→ s2",
+              "substance": [
+                "/表象/视觉/实体/物品/起居物品/食物",
+                "/表象/视觉/实体/物品/器物物品/道具"
+              ],
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            }
+          ],
+          "intent": "用 {via:human} {act:结构生成} 从 {in-type:描述} 构建中英协同的 {out-type:提示词}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "预处理"
+        },
+        {
+          "id": "s2",
+          "kind": "step",
+          "via": "<AI生图工具>",
+          "action": "生成/元素生成",
+          "feature": "随机",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "一张手绘风格的城市美食地图，以成都为主题。画面以鸟瞰视角的手绘简化城市地图为底，标注主要道路和地标但不追求精确比例而是追求可爱的手绘感。地图上分布着12个美食地点的精致手绘小插画：春熙路的串串香（一把竹签插着各种食材冒着热气）、宽窄巷子的三大炮（三个糯米团子飞向铜盘）、玉林路的小龙坎火锅（九宫格锅翻滚冒泡），武侯祠大街的钵钵鸡，人民公园鹤鸣茶社的盖碗茶，锦里古街的冰粉，双流老妈兔头，建设路叶婆婆烘蛋糕，钟水饺（总店），陈麻婆豆腐...左上角标题「成都·吃货暴走地图」使用胖圆的手绘美术字配辣椒装饰。整体画风为水彩+彩铅混合的手绘质感，颜色以暖色系为主，边框用辣椒和植物装饰。watercolor illustration, handmade texture, warm color palette, cozy illustrated map style"
+            ],
+            [
+              "memo",
+              "GPT-Image-2的CJK优化让中文场景细节描述能被模型正确理解；「串串香（一把竹签插着各种食材冒着热气）」这类中文细节描述远比英文版精准；适用场景：中文内容图（美食地图、城市海报、中文UI）、日系风格素材、本地化营销图"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "提示词",
+              "name": "多语言混合提示词",
+              "value": "一张手绘风格的城市美食地图，以成都为主题。画面以鸟瞰视角的手绘简化城市地图为底，标注主要道路和地标但不追求精确比例而是追求可爱的手绘感。地图上分布着12个美食地点的精致手绘小插画：春熙路的串串香（一把竹签插着各种食材冒着热气）、宽窄巷子的三大炮（三个糯米团子飞向铜盘）、玉林路的小龙坎火锅（九宫格锅翻滚冒泡，越煮越巴适）、东郊记忆的冒椒麻辣，武侯祠大街的钵钵鸡（冷锅串串），人民公园鹤鸣茶社的盖碗茶，锦里古街的冰粉（手握冰粉碗），双流老妈兔头，建设路叶婆婆烘蛋糕，钟水饺（总店），陈麻婆豆腐（麻辣鲜香）。左上角标题「成都·吃货暴走地图」使用胖圆的手绘美术字配辣椒装饰。整体画风为水彩+彩铅混合的手绘质感，颜色以暖色系为主，边框用辣椒和植物装饰，右下角有图例说明和指南针。watercolor illustration, handmade texture, warm color palette, cozy illustrated map style",
+              "anchor": "← s1.多语言混合提示词",
+              "substance": [
+                "/表象/视觉/实体/物品/起居物品/食物",
+                "/表象/视觉/实体/物品/器物物品/道具"
+              ],
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "成品图",
+              "name": "本地化内容图",
+              "value": "<AI生成的成都吃货暴走地图：手绘水彩风格，鸟瞰成都城市地图，环形路/春熙路/宽窄巷子/锦里等地标可辨，12个美食地点各附精致手绘插画（串串香/水饺/三大炮/盖碗茶/冰粉等），暖黄暖棕色调，辣椒边框装饰，左上角「成都·吃货暴走地图」美术字，右下角图例说明+指南针>",
+              "anchor": "→ 返回 本地化内容图",
+              "substance": [
+                "/表象/视觉/实体/物品/起居物品/食物",
+                "/表象/视觉/实体/物品/器物物品/道具"
+              ],
+              "form": "/呈现/视觉/视觉风格/视觉艺术风格/绘画风格/插画"
+            }
+          ],
+          "intent": "用 {via:AI生图工具} {act:元素生成} 将 {in-type:提示词} 渲染为精细的 {out-type:成品图}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "主体生成"
+        }
+      ],
+      "return_row": {
+        "arrow": "↩",
+        "text": "<span class=\"kw\">返回</span> <span class=\"chip\" data-type=\"图像\">本地化内容图</span>"
+      }
+    },
+    {
+      "id": "p9",
+      "name": "结构化精确度验证出图",
+      "purpose": "用精确计数/空间布局要求的提示词测试AI的计数能力和空间推理能力，作为评估工具能力边界的标尺",
+      "category": "产物创造",
+      "platform": "zhihu",
+      "author": "白嫖知识分享",
+      "declarations": {
+        "inputs": [
+          {
+            "type": "文本",
+            "name": "结构精确性要求",
+            "desc": "精确的计数约束和空间布局要求（如「三层书架：1+3+7=11本书」）"
+          }
+        ],
+        "resources": [],
+        "returns": {
+          "type": "文本"
+        }
+      },
+      "steps": [
+        {
+          "id": "s1",
+          "kind": "step",
+          "via": "human",
+          "action": "生成/关系生成/结构生成",
+          "feature": "人工",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "用英文撰写精确计数测试提示词；明确标注每个区域/层的数量（如「on the top shelf, there should be one book; on the second shelf, three books; on the bottom shelf, seven books」）；语言精准，避免歧义"
+            ],
+            [
+              "memo",
+              "这种提示词不是拿来日常用的，而是评估AI绘图工具能力边界的标尺；先跑「计数测试」和「结构化测试」，就知道当前工具能支撑多复杂的指令；选工具先测能力边界，再谈创作自由度；测试AI的「计数能力」和「空间推理能力」"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "描述",
+              "name": "结构精确性要求",
+              "value": "三层木质书架计数测试：顶层1本书，第二层3本书，底层7本书，共11本",
+              "anchor": "← 工序输入",
+              "substance": "/表象/视觉/实体/物品/起居物品/家居",
+              "form": "/架构/修辞/语调风格/平实表达/直白陈述"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "提示词",
+              "name": "结构测试提示词",
+              "value": "A wooden bookshelf consisting of three shelves: On the top shelf, there should be one book, on the second shelf, there should be three books, and on the bottom shelf, there should be seven books.",
+              "anchor": "→ s2",
+              "substance": "/表象/视觉/实体/物品/起居物品/家居",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            }
+          ],
+          "intent": "用 {via:human} {act:结构生成} 从 {in-type:描述} 撰写精确计数的 {out-type:提示词}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "预处理"
+        },
+        {
+          "id": "s2",
+          "kind": "step",
+          "via": "<AI生图工具>",
+          "action": "生成/元素生成",
+          "feature": "随机",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "A wooden bookshelf consisting of three shelves: On the top shelf, there should be one book, on the second shelf, there should be three books, and on the bottom shelf, there should be seven books."
+            ],
+            [
+              "memo",
+              "GPT-Image-2可以准确完成此计数测试；早期模型大概率数不对；计数准确性和空间推理能力是AI工具能力边界的重要维度"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "提示词",
+              "name": "结构测试提示词",
+              "value": "A wooden bookshelf consisting of three shelves: On the top shelf, there should be one book, on the second shelf, there should be three books, and on the bottom shelf, there should be seven books.",
+              "anchor": "← s1.结构测试提示词",
+              "substance": "/表象/视觉/实体/物品/起居物品/家居",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "样图",
+              "name": "结构验证图",
+              "value": "<AI生成的木质书架图：三层书架，顶层1本深蓝色书「THE ART OF THOUGHT」，中层3本书（绿色「THE NATURAL WORLD」+红色「HISTORY OF CIVILIZATIONS」+蓝色「THE HUMAN SPIRIT」），底层7本书（The Classics/Poetry/Science and Discovery/Art Through the Ages/Explorers and Adventures/Great Minds/Legends and Myths），精确符合1+3+7=11本要求>",
+              "anchor": "→ s3",
+              "substance": "/表象/视觉/实体/物品/起居物品/家居",
+              "form": "/呈现/视觉/视觉制作/实景拍摄"
+            }
+          ],
+          "intent": "用 {via:AI生图工具} {act:元素生成} 从 {in-type:提示词} 生成 {out-type:样图}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "主体生成"
+        },
+        {
+          "id": "s3",
+          "kind": "step",
+          "via": "human",
+          "action": "提取/化学提取/识别",
+          "feature": "人工",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "逐层清点图像中的书本数量：顶层应为1本，中层应为3本，底层应为7本；记录AI工具的计数准确性；若不符则标记为测试失败，记录实际数量与预期的差异"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "样图",
+              "name": "结构验证图",
+              "value": "<AI生成的木质书架图：三层书架，顶层1本深蓝色书「THE ART OF THOUGHT」，中层3本书，底层7本书，精确符合1+3+7=11本要求>",
+              "anchor": "← s2.结构验证图",
+              "substance": "/表象/视觉/实体/物品/起居物品/家居",
+              "form": "/呈现/视觉/视觉制作/实景拍摄"
+            },
+            {
+              "type": "描述",
+              "name": "结构精确性要求",
+              "value": "三层木质书架计数测试：顶层1本书，第二层3本书，底层7本书，共11本",
+              "anchor": "← 工序输入",
+              "substance": "/表象/视觉/实体/物品/起居物品/家居",
+              "form": "/架构/修辞/语调风格/平实表达/直白陈述"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "评语",
+              "name": "验证结果",
+              "value": "GPT-Image-2通过验证：书架三层分别有1本（顶）、3本（中）、7本（底），总计11本，精确符合计数和空间推理要求；该工具具备可靠的结构化指令执行能力，可支撑复杂计数约束场景",
+              "anchor": "→ 返回 验证结果",
+              "substance": "/理念/知识/商业/前沿技术/AI智能/AI应用",
+              "form": "/架构/逻辑/信息编排逻辑/归纳论证/归纳提炼/概括浓缩"
+            }
+          ],
+          "intent": "用 {via:human} {act:识别} 核验 {in-type:样图} 计数准确性，输出 {out-type:评语}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "检验"
+        }
+      ],
+      "return_row": {
+        "arrow": "↩",
+        "text": "<span class=\"kw\">返回</span> <span class=\"chip\" data-type=\"文本\">验证结果</span>"
+      }
+    }
+  ]
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/.session_id

@@ -0,0 +1 @@
+0b176e93-ff9a-46cf-9930-4d335542bdaf

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/_meta.json

@@ -0,0 +1,8 @@
+{
+  "case_id": "gzh_25318352c81bc97eb92cebe4c617b786",
+  "from_q": "q0018",
+  "form": "B",
+  "score": 8.4,
+  "score_norm": 0.84,
+  "started_at": "2026-06-03T18:46:16"
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/_scratch/anchors.json

@@ -0,0 +1,9 @@
+[
+  {"path": "p1.s1.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p1.s1.outputs[0].anchor", "value": "→ s2"},
+  {"path": "p1.s2.inputs[0].anchor",  "value": "← s1.JSON风格分析"},
+  {"path": "p1.s2.inputs[1].anchor",  "value": "← 工序输入"},
+  {"path": "p1.s2.outputs[0].anchor", "value": "→ s3"},
+  {"path": "p1.s3.inputs[0].anchor",  "value": "← s2.图像生成提示词"},
+  {"path": "p1.s3.outputs[0].anchor", "value": "→ 返回 风格化生成图"}
+]

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/_scratch/patch_2a.json

@@ -0,0 +1,17 @@
+[
+  {"path": "p1.s1.effect", "value": "预处理"},
+  {"path": "p1.s1.action", "value": "提取/化学提取/反推"},
+  {"path": "p1.s1.inputs[0].type", "value": "参考图"},
+  {"path": "p1.s1.outputs[0].type", "value": "风格描述"},
+  {"path": "p1.s2.effect", "value": "预处理"},
+  {"path": "p1.s2.action", "value": "生成/元素生成"},
+  {"path": "p1.s2.inputs[0].type", "value": "风格描述"},
+  {"path": "p1.s2.inputs[1].type", "value": "提示词"},
+  {"path": "p1.s2.outputs[0].type", "value": "提示词"},
+  {"path": "p1.s3.effect", "value": "主体生成"},
+  {"path": "p1.s3.action", "value": "生成/元素生成"},
+  {"path": "p1.s3.inputs[0].type", "value": "提示词"},
+  {"path": "p1.s3.outputs[0].type", "value": "样图"},
+  {"path": "p1.type_registry.风格描述.extends", "value": "描述"},
+  {"path": "p1.type_registry.风格描述.desc", "value": "豆包从参考图提取的JSON格式视觉风格描述，含颜色/排版/构图/特效四个维度，作为生图提示词的风格模板来源"}
+]

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/_scratch/patch_2b.json

@@ -0,0 +1,16 @@
+[
+  {"path": "p1.s1.inputs[0].substance", "value": "/表象/视觉/实体/物品/影像物品/艺术创作"},
+  {"path": "p1.s1.inputs[0].form", "value": "/呈现/视觉/视觉风格/视觉艺术风格/绘画风格/插画"},
+  {"path": "p1.s1.outputs[0].substance", "value": "/表象/视觉/画面/后期/美学风格"},
+  {"path": "p1.s1.outputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+  {"path": "p1.s2.inputs[0].substance", "value": "/表象/视觉/画面/后期/美学风格"},
+  {"path": "p1.s2.inputs[0].form", "value": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"},
+  {"path": "p1.s2.inputs[1].substance", "value": "/表象/视觉/空间/空间环境/地形地貌/山水景观"},
+  {"path": "p1.s2.inputs[1].form", "value": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"},
+  {"path": "p1.s2.outputs[0].substance", "value": "/理念/知识/商业/前沿技术/AI智能/AI应用"},
+  {"path": "p1.s2.outputs[0].form", "value": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"},
+  {"path": "p1.s3.inputs[0].substance", "value": "/理念/知识/商业/前沿技术/AI智能/AI应用"},
+  {"path": "p1.s3.inputs[0].form", "value": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"},
+  {"path": "p1.s3.outputs[0].substance", "value": "/表象/视觉/空间/空间环境/地形地貌/山水景观"},
+  {"path": "p1.s3.outputs[0].form", "value": "/呈现/视觉/视觉制作/后期处理/图像合成处理/AI生成合成"}
+]

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/_scratch/task_2a.json

@@ -0,0 +1,107 @@
+{
+  "case_id": "B_gzh_25318352",
+  "task": "Phase 2A (effect/action/type normalization)",
+  "image_url_list": [],
+  "steps": [
+    {
+      "path": "p1.s1",
+      "name": "",
+      "action": "提取/内容提取",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {
+        "风格描述": {
+          "extends": "提示词",
+          "desc": "豆包从参考图提取的JSON格式视觉风格描述，含颜色/排版/构图/特效四个维度，作为生图提示词的风格模板来源"
+        }
+      },
+      "inputs": [
+        {
+          "path": "p1.s1.inputs[0]",
+          "name": "风格参考图",
+          "value": "<5张绘画风格图：作品1金蓝大漠骆驼队中国风星空（顶部CHINA大字+金箔质感）/作品2浅蓝淡金小满人物麦浪（中式印章+薄雾意境）/作品3灰白极简CHANEL时装秀（镜面反射+朦胧虚化）/作品4深红柱廊寺庙ZENFO（对称透视+宏观建筑）/作品5彩色抽象藏族女性与狼（高饱和插画风）>",
+          "type": "参考图",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p1.s1.outputs[0]",
+          "name": "JSON风格分析",
+          "value": "作品编号: 1, 颜色: '主色调为金、蓝、米黄，搭配深蓝星空质感，高对比度的复古华丽配色', 排版: '顶部大号\"CHINA\"英文+小字主题说明，文字与画面融合，中式印章点缀右上角', 构图: '前景驼队、中景沙丘、背景星空，层次分明的纵深感，对称与流动感结合', 特效: '金箔纹理质感，星空颗粒效果，沙丘线条的渐变晕染' || 作品编号: 2, 颜色: '浅蓝、灰白、淡金，低饱和度的清冷莫兰迪色系，营造静谧氛围', 排版: '中部书法字体\"小满\"+英文说明，右上角中式印章，文字简洁留白', 构图: '前景人物、中景麦浪、远景山峦，中心人物+对称式山水层次，留白充足', 特效: '麦浪的线条流动感，画面的朦胧薄雾质感，低透明度的意境渲染' || 作品编号: 3, 颜色: '灰白、浅灰、墨黑，极简的单色调，冷感高级灰为主', 排版: '中部粗体\"CHANEL\"英文，无多余文字，品牌标识突出', 构图: '中心模特+背景建筑，蜿蜒步道引导视觉，对称式场景布局', 特效: '镜面反射效果，薄雾的朦胧虚化，极简空间的层次（图中截断）' || 作品编号: 4-5（图4未完整展示，格式同上）",
+          "type": "风格描述",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p1.s2",
+      "name": "",
+      "action": "整理/构建",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {
+        "风格描述": {
+          "extends": "提示词",
+          "desc": "豆包从参考图提取的JSON格式视觉风格描述，含颜色/排版/构图/特效四个维度，作为生图提示词的风格模板来源"
+        }
+      },
+      "inputs": [
+        {
+          "path": "p1.s2.inputs[0]",
+          "name": "JSON风格分析",
+          "value": "",
+          "type": "风格描述",
+          "related_images": []
+        },
+        {
+          "path": "p1.s2.inputs[1]",
+          "name": "图像描述词",
+          "value": "<用户自拟的新图主体描述：清晨群山与层叠梯田，中国南方山野清晨，薄雾未散，电影级远景推进镜头，略微俯拍，自然晨光，冷暖交替，安静克制，大师作品高品质高质量>",
+          "type": "提示词",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p1.s2.outputs[0]",
+          "name": "图像生成提示词",
+          "value": "帮我生成图片：颜色：'主色调为金、蓝、米黄，搭配深蓝星空质感，高对比度的复古华丽配色'，排版：'顶部大号\"CHINA\"英文+小字主题说明，文字与画面融合，中式印章点缀右上角'，构图：'前景驼队、中景沙丘、背景星空，层次分明的纵深感，对称与流动感结合'，特效：'金箔纹理质感，星空颗粒效果，沙丘线条的渐变晕染'，比例「16:9」**主体：**清晨群山与层叠梯田。**场景与时间：**中国南方山野清晨，薄雾未散。**镜头与视角：**电影级远景推进镜头，略微俯拍。**光影与氛围：**自然晨光，冷暖交替，安静克制。大师作品，高品质，高质量，电影级场景",
+          "type": "提示词",
+          "related_images": []
+        }
+      ]
+    },
+    {
+      "path": "p1.s3",
+      "name": "",
+      "action": "生成/图像生成/文生图",
+      "effect": "",
+      "kind": "step",
+      "type_registry": {
+        "风格描述": {
+          "extends": "提示词",
+          "desc": "豆包从参考图提取的JSON格式视觉风格描述，含颜色/排版/构图/特效四个维度，作为生图提示词的风格模板来源"
+        }
+      },
+      "inputs": [
+        {
+          "path": "p1.s3.inputs[0]",
+          "name": "图像生成提示词",
+          "value": "",
+          "type": "提示词",
+          "related_images": []
+        }
+      ],
+      "outputs": [
+        {
+          "path": "p1.s3.outputs[0]",
+          "name": "风格化生成图",
+          "value": "<AI生成的16:9图像组(4张候选)，以金/蓝/米黄为主色调，CHINA英文字融入层叠梯田与星空背景，清晨薄雾光影，复古华丽中国风；带\"AI生成\"水印>",
+          "type": "生成图",
+          "related_images": []
+        }
+      ]
+    }
+  ]
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/_scratch/task_2b.json

@@ -0,0 +1,56 @@
+{
+  "case_id": "B_gzh_25318352",
+  "task": "Phase 2B (substance/form matching)",
+  "image_url_list": [],
+  "items_to_match": [
+    {
+      "path": "p1.s1.inputs[0]",
+      "name": "风格参考图",
+      "value": "<5张绘画风格图：作品1金蓝大漠骆驼队中国风星空（顶部CHINA大字+金箔质感）/作品2浅蓝淡金小满人物麦浪（中式印章+薄雾意境）/作品3灰白极简CHANEL时装秀（镜面反射+朦胧虚化）/作品4深红柱廊寺庙ZENFO（对称透视+宏观建筑）/作品5彩色抽象藏族女性与狼（高饱和插画风）>",
+      "type": "参考图",
+      "related_images": []
+    },
+    {
+      "path": "p1.s1.outputs[0]",
+      "name": "JSON风格分析",
+      "value": "作品编号: 1, 颜色: '主色调为金、蓝、米黄，搭配深蓝星空质感，高对比度的复古华丽配色', 排版: '顶部大号\"CHINA\"英文+小字主题说明，文字与画面融合，中式印章点缀右上角', 构图: '前景驼队、中景沙丘、背景星空，层次分明的纵深感，对称与流动感结合', 特效: '金箔纹理质感，星空颗粒效果，沙丘线条的渐变晕染' || 作品编号: 2, 颜色: '浅蓝、灰白、淡金，低饱和度的清冷莫兰迪色系，营造静谧氛围', 排版: '中部书法字体\"小满\"+英文说明，右上角中式印章，文字简洁留白', 构图: '前景人物、中景麦浪、远景山峦，中心人物+对称式山水层次，留白充足', 特效: '麦浪的线条流动感，画面的朦胧薄雾质感，低透明度的意境渲染' || 作品编号: 3, 颜色: '灰白、浅灰、墨黑，极简的单色调，冷感高级灰为主', 排版: '中部粗体\"CHANEL\"英文，无多余文字，品牌标识突出', 构图: '中心模特+背景建筑，蜿蜒步道引导视觉，对称式场景布局', 特效: '镜面反射效果，薄雾的朦胧虚化，极简空间的层次（图中截断）' || 作品编号: 4-5（图4未完整展示，格式同上）",
+      "type": "风格描述",
+      "related_images": []
+    },
+    {
+      "path": "p1.s2.inputs[0]",
+      "name": "JSON风格分析",
+      "value": "",
+      "type": "风格描述",
+      "related_images": []
+    },
+    {
+      "path": "p1.s2.inputs[1]",
+      "name": "图像描述词",
+      "value": "<用户自拟的新图主体描述：清晨群山与层叠梯田，中国南方山野清晨，薄雾未散，电影级远景推进镜头，略微俯拍，自然晨光，冷暖交替，安静克制，大师作品高品质高质量>",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p1.s2.outputs[0]",
+      "name": "图像生成提示词",
+      "value": "帮我生成图片：颜色：'主色调为金、蓝、米黄，搭配深蓝星空质感，高对比度的复古华丽配色'，排版：'顶部大号\"CHINA\"英文+小字主题说明，文字与画面融合，中式印章点缀右上角'，构图：'前景驼队、中景沙丘、背景星空，层次分明的纵深感，对称与流动感结合'，特效：'金箔纹理质感，星空颗粒效果，沙丘线条的渐变晕染'，比例「16:9」**主体：**清晨群山与层叠梯田。**场景与时间：**中国南方山野清晨，薄雾未散。**镜头与视角：**电影级远景推进镜头，略微俯拍。**光影与氛围：**自然晨光，冷暖交替，安静克制。大师作品，高品质，高质量，电影级场景",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p1.s3.inputs[0]",
+      "name": "图像生成提示词",
+      "value": "",
+      "type": "提示词",
+      "related_images": []
+    },
+    {
+      "path": "p1.s3.outputs[0]",
+      "name": "风格化生成图",
+      "value": "<AI生成的16:9图像组(4张候选)，以金/蓝/米黄为主色调，CHINA英文字融入层叠梯田与星空背景，清晨薄雾光影，复古华丽中国风；带\"AI生成\"水印>",
+      "type": "生成图",
+      "related_images": []
+    }
+  ]
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/_source.json

@@ -0,0 +1,19 @@
+{
+  "title": "超简单！一句提示词，帮你高效提取图片风格",
+  "link": "https://mp.weixin.qq.com/s?__biz=MzkxNjg3OTYwNg==&mid=2247484017&idx=1&sn=731fa955e75f6f5cf614d54364548a7d&chksm=c0506a0c7b384a5e4d28dee568d5575bb765c264e1376329b28c9b0b7786218cb2f7eb803366#rd",
+  "body_text": "哈喽，我是青山Linn!\n\n平时用AI生图的时候，你是不是也遇到过这种情况：\n看到一张喜欢的图片风格，却不知道它背后的风格提示词和构图逻辑\n？\n今天分享一个简单方法，\n只需要一句提示词，就能提取它的视觉元素，包括图片风格和结构\n——直接可复用，让你像拆解影视镜头一样拆解每一张图片。\n首先是准备工具和提示词\n打开\n豆包\n，上传喜欢的图片\n\n输入提示词：以JSON结构化数据形式提取视觉风格，包括颜色、排版、构图和特效\n\nPS：可以一次上传单张或多种图片，我这里嫌麻烦一次上传了5张\n得到的结果如下\n\n然后复制\n风格提示词，\n打开对话框里的图片生成模式，输入刚刚得到的风格提示词和你想要的图片描述词。\n\n生成的图片即可得到想要的风格效果：\n\n如果对图片的艺术风格效果不满意的话，还可以尝试其他的风格提示词\n\n是不是又快又好？赶快去试试吧。\n以上就是今天分享的内容啦，我是青山\n，专注于\nAI影视结构拆解和AIGC内容创作\n。平时我会分享\nAI创作里可以快速\n复用的方法和工\n具\n，让创作者少走弯路、快速上手。\n点击关注可以第一时间获取最新玩法和资讯。\n觉得我的文章有用，记得点个\n关注、分享、喜欢\n。\n更多工具分享：https://tools.hkting.com。为了跟上 AI 科技的发展，我建了个 AI 交流群，不定期分享资讯、工具、技术等，有需要的朋友可以加群。\n下方扫码备注\n“加入AI社群”\n，我拉你进群\n。\n\n如果你想把AI影视结构拆解成可复用的模板，或者系统学习如何用AI快速生成视觉内容，记得关注我，我会分享更多实战技巧。",
+  "images": [
+    "https://mmbiz.qpic.cn/sz_mmbiz_png/y8icOkx3v7gXZfanicWJxuPJsA2Hd024hljhApfJghDeLjJlTOBYibJ6SJ47Rbtfg0PaJLtt2kweefIBpB7sA7zjMWXJKDHw36YbcuzZHnjAt4/640?wx_fmt=png&from=appmsg",
+    "https://mmbiz.qpic.cn/sz_mmbiz_png/y8icOkx3v7gXBWU6pgQ2zDPYouEcA4PKXicFRLglnk47PsC6ticHbWIeqkDmLD1CUA6iawTibZmia9WANpibqviaEWRickzVV6IIWuUswiaqZnE2EBsAY/640?wx_fmt=png&from=appmsg",
+    "https://mmbiz.qpic.cn/mmbiz_png/y8icOkx3v7gXRhicKtVWquib1fW47ImCr2ojhM4WMaF7VF4GI82u16D7sAOEa8pwc62BeicxlJNGvqQ8LI3ia3nbeJHn91JhzlaFibtdVTzicPmtfQ/640?wx_fmt=png&from=appmsg",
+    "https://mmbiz.qpic.cn/sz_mmbiz_png/y8icOkx3v7gXHg9tNJicy2swW6lqUyN08S3ghWBnU2Mr8UmsL8XCJmHaQdgLxpicYMyCbUB8AfUtWa8CRd1QTBRtYLcicZRmXqkickXLAgM1WNm8/640?wx_fmt=png&from=appmsg",
+    "https://mmbiz.qpic.cn/mmbiz_png/y8icOkx3v7gXECQhtdrIicx2G6UYb0pB5wAlrd7KWf6YG6RJNTpDhpkX8oPq41Jan74wicia39sPUR72ne10dphzXf9QHsVqOjU5MTiaghibZjJto/640?wx_fmt=png&from=appmsg",
+    "https://mmbiz.qpic.cn/mmbiz_jpg/y8icOkx3v7gWtLeD8eeLCDzTWW8hHqeoEubD2m9cz38tfo31bk9AnSs4YYsibltDEgPpNh7QGvGJnicB7GdUaBRUdbsGHPrXaTPKKwicERhCjo0/640?wx_fmt=jpeg",
+    "https://mmbiz.qpic.cn/sz_mmbiz_png/y8icOkx3v7gWvicD7iclsGdB3PQySlHbmHEribGRA864gWIAwgtLiczfsqDfibE7kNC0TxicuBb9mH8rPmaHIxKPibtiavzD1IU19rhnXua9pRibTEKmI/640?wx_fmt=png&from=appmsg",
+    "https://mmbiz.qpic.cn/sz_mmbiz_jpg/y8icOkx3v7gVXm6QsT2PgQJPr0qAl2U6t7NMJTiaRT4QTcNYcpibN2mGvibtxe7LQYgTPSiaQ5pIIJCdEHamJB94xW0S5CEYNuibdXQYQ6URSFlDY/640?wx_fmt=jpeg",
+    "https://mmbiz.qpic.cn/mmbiz_png/y8icOkx3v7gWAslQBbl5GCqmSpy4WqOX9jYkcyiakgnyQAYRUv4dI2j4yJicoODGgZkhg2Fqibp7UUC85Y7tnWXMq3XcsyfeVgrASdKDmlahiauo/640?wx_fmt=png&from=appmsg",
+    "https://mmbiz.qpic.cn/mmbiz_jpg/6lxztbicU7NZMow101MVD0yvZZ7oLkgUoLHnMRSxlSmwo1rhNIAibmMEaWXG1FA1LicVq4qbVDmBFmQ3aDME6icdiag/640?wx_fmt=other&from=appmsg&wxfrom=5&wx_lazy=1&tp=webp#imgIndex=10"
+  ],
+  "publish_timestamp": "2026-02-06 09:53:12",
+  "channel_account_name": "gzh"
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/_trace.md

@@ -0,0 +1,706 @@
+
+
+---
+
+## ▶ Fresh @ 2026-06-03 18:46:19
+
+- case: `B_gzh_25318352`
+- source: `C:\Users\11304\gitlab\cybertogether\Agent\examples\process_pipeline\script\search_eval\runs_full\q0018\procedures\B_gzh_25318352\_source.json`
+- spec: `spec`
+- images: `10`
+- model: `claude-sonnet-4-6`
+- max_turns: `300`
+
+_session=0b176e93 · model='claude-sonnet-4-6' · apiKeySource='none'_
+
+
+### Turn 1 · 18:46:26
+
+
+### Turn 2 · 18:46:28
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/README.md`
+
+
+### Turn 3 · 18:46:32
+
+
+### Turn 4 · 18:46:33
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools.md`
+
+
+### Turn 5 · 18:46:36
+
+
+### Turn 6 · 18:46:36
+> Now let me read the Phase 1 spec files and the source JSON concurrently:
+
+
+
+### Turn 7 · 18:46:38
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/fields.md`
+
+
+### Turn 8 · 18:46:40
+
+
+### Turn 9 · 18:46:42
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/control-flow.md`
+
+
+### Turn 10 · 18:46:45
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/phase1-skeleton.md`
+
+
+### Turn 11 · 18:46:48
+
+
+### Turn 12 · 18:46:49
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/_source.json`
+
+
+### Turn 13 · 18:46:52
+
+
+### Turn 14 · 18:46:53
+> Now let me read the remaining Phase 1 spec files and the workflow template in parallel:
+
+
+
+### Turn 15 · 18:46:54
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/format/md-structure.md`
+
+
+### Turn 16 · 18:46:56
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/templates/workflow.template.json`
+
+
+### Turn 17 · 18:48:28
+
+
+### Turn 18 · 18:48:28
+> Now let me read the syntax.md and phase2-normalize.md + phase3-finalize.md in parallel:
+
+
+
+### Turn 19 · 18:48:29
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/syntax.md`
+
+
+### Turn 20 · 18:48:32
+
+
+### Turn 21 · 18:48:33
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/phase2-normalize.md`
+
+
+### Turn 22 · 18:48:39
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/extraction/phase3-finalize.md`
+
+
+### Turn 23 · 18:53:56
+
+
+### Turn 24 · 18:53:56
+> Now I have all the context I need. Let me write Phase 1.1 (understanding.md) and Phase 1.2 (workflow.json skeleton) simultaneously:
+
+
+
+### Turn 25 · 18:54:45
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/understanding.md` (2,704 chars)
+
+
+### Turn 26 · 18:55:29
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/workflow.json` (5,216 chars)
+
+
+### Turn 27 · 18:55:47
+
+
+### Turn 28 · 18:55:47
+> Phase 1.3 — now I'll create the anchors patch and apply them:
+
+
+
+### Turn 29 · 18:55:49
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/_scratch/anchors.json` (423 chars)
+
+
+### Turn 30 · 18:55:52
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/wf-patch.py" --workflow "C:/Users/11304/gitlab/cybertogether/Agent/examples/proc...`
+
+
+### Turn 31 · 18:56:04
+
+
+### Turn 32 · 18:56:05
+> Phase 1.3 complete (7/7 anchors applied). Now entering **Phase 2** — first run the prepare-subtask.py script to generate task files for sub-agents:
+
+
+
+### Turn 33 · 18:56:07
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/prepare-subtask.py" --workflow "C:/Users/11304/gitlab/cybertogether/Agent/exampl...`
+
+
+### Turn 34 · 18:56:11
+
+
+### Turn 35 · 18:56:11
+> Now I'll spawn both Phase 2A and 2B sub-agents in parallel:
+
+
+
+### Turn 36 · 18:56:27
+- `Agent` — `Phase 2A: effect/action/type normalization for B_gzh_25318352` [subagent=phase-2a-normalizer]
+
+
+### Turn 37 · 18:56:31
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/_scratch/task_2a.json`
+
+
+### Turn 38 · 18:56:31
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/workflow.json`
+
+
+### Turn 39 · 18:56:32
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/taxonomy/effect.json`
+
+
+### Turn 40 · 18:56:36
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/taxonomy/action.json`
+
+
+### Turn 41 · 18:56:36
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/taxonomy/type.json`
+
+
+### Turn 42 · 18:56:43
+- `Agent` — `Phase 2B: substance/form taxonomy matching for B_gzh_25318352` [subagent=phase-2b-matcher]
+
+
+### Turn 43 · 18:56:47
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/_scratch/task_2b.json`
+
+
+### Turn 44 · 18:56:47
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/workflow.json`
+
+
+### Turn 45 · 18:56:53
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "绘画风格图 参考图 插画 中国风"`
+
+
+### Turn 46 · 18:56:54
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "绘画风格图 参考图 插画 图像"`
+
+
+### Turn 47 · 18:56:58
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "JSON风格分析 视觉风格 颜色排版构图特效"`
+
+
+### Turn 48 · 18:56:59
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "JSON结构化 风格描述 文字"`
+
+
+### Turn 49 · 18:57:02
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "视觉风格 绘画风格 颜色构图"`
+
+
+### Turn 50 · 18:57:04
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "结构化数据 JSON 格式化文本"`
+
+
+### Turn 51 · 18:57:07
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --narrow "视觉风格 绘画风格 颜色构图"`
+
+
+### Turn 52 · 18:57:08
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --narrow "结构化数据 JSON 格式化文本"`
+
+
+### Turn 53 · 18:57:11
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "文字描述 说明 信息文本"`
+
+
+### Turn 54 · 18:57:13
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "图像描述 场景描述 创作主题 构图意境"`
+
+
+### Turn 55 · 18:57:15
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/_scratch/patch_2a.json` (819 chars)
+
+
+### Turn 56 · 18:57:18
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "生图指令 提示词 图像生成"`
+
+
+### Turn 57 · 18:57:18
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "提示词 文本指令 文字"`
+
+
+### Turn 58 · 18:57:22
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "AI生成图 风格迁移 文生图 梯田星空"`
+
+
+### Turn 59 · 18:57:24
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "AI生成图 图像 生成图"`
+
+
+### Turn 60 · 18:57:27
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --narrow "AI生成图 文生图 风格迁移"`
+
+
+### Turn 61 · 18:57:29
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "梯田风景 山野 自然景观 清晨光影"`
+
+
+### Turn 62 · 18:57:33
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "AI技术 机器学习 图像生成模型"`
+
+
+### Turn 63 · 18:57:34
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "图片 照片 视觉图像"`
+
+
+### Turn 64 · 18:57:38
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --validate "/表象/视觉/实体/物品/影像物品/艺术创作"`
+
+
+### Turn 65 · 18:57:39
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --validate "/呈现/视觉/视觉风格/视觉艺术风格/绘画风格/插画"`
+
+
+### Turn 66 · 18:57:43
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "/表象/视觉/实体/物品/影像物品/艺术创作"`
+
+
+### Turn 67 · 18:57:44
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "/呈现/视觉/视觉风格/视觉艺术风格/绘画风格/插画"`
+
+
+### Turn 68 · 18:57:48
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --help`
+
+
+### Turn 69 · 18:57:56
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --list-l2`
+
+
+### Turn 70 · 18:57:59
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --list-l2`
+
+
+### Turn 71 · 18:58:03
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --subtree "/表象/视觉/实体/物品/影像物品"`
+
+
+### Turn 72 · 18:58:04
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --subtree "/呈现/视觉/视觉风格/视觉艺术风格/绘画风格"`
+
+
+### Turn 73 · 18:58:08
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python -c "
+import subprocess, sys
+result = subprocess.run(
+    [sys.executable, 'spec/tool...`
+
+
+### Turn 74 · 18:58:14
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python -c "
+import subprocess, sys, os
+env = os.environ.copy()
+env['PYTHONUTF8'] = '1'
+resu...`
+
+
+### Turn 75 · 18:58:17
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --subtree "/表象/视觉/实体/物品/影像物品" 2>&1 | head -30`
+
+
+### Turn 76 · 18:58:21
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/taxonomy-lookup.py` (offset=0, limit=50)
+
+
+### Turn 77 · 18:58:23
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/taxonomy-lookup.py` (offset=50, limit=100)
+
+
+### Turn 78 · 18:58:28
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --subtree "表象/视觉/实体/物品/影像物品"`
+
+
+### Turn 79 · 18:58:33
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "表象/视觉/实体/物品/影像物品/艺术创作"`
+
+
+### Turn 80 · 18:58:34
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "呈现/视觉/视觉风格/视觉艺术风格/绘画风格/插画"`
+
+
+### Turn 81 · 18:58:39
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "中国风 国风 水墨 绘画"`
+
+
+### Turn 82 · 18:58:40
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --subtree "呈现/视觉/视觉风格/视觉艺术风格"`
+
+
+### Turn 83 · 18:58:47
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "呈现/视觉/视觉风格/视觉艺术风格/数字风格/AI风格"`
+
+
+### Turn 84 · 18:58:48
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "呈现/视觉/视觉风格/视觉艺术风格/绘画风格/插画"`
+
+
+### Turn 85 · 18:58:52
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "视觉风格描述 视觉要素 颜色构图排版"`
+
+
+### Turn 86 · 18:58:53
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --subtree "表象/视觉/画面/后期"`
+
+
+### Turn 87 · 18:58:58
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "风格参数 视觉属性 设计规范"`
+
+
+### Turn 88 · 18:59:00
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "功能效用 用法 操作方法"`
+
+
+### Turn 89 · 18:59:03
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "AI工具 豆包 模型能力"`
+
+
+### Turn 90 · 18:59:04
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "创作指令 操作规范 步骤指南"`
+
+
+### Turn 91 · 18:59:08
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "图像生成 AI绘画 数字艺术创作"`
+
+
+### Turn 92 · 18:59:09
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "梯田 景观 风光摄影 中国南方"`
+
+
+### Turn 93 · 18:59:12
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "表象/视觉/空间/空间场所/城乡空间/乡村空间"`
+
+
+### Turn 94 · 18:59:13
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "表象/视觉/空间/空间环境/地形地貌/山水景观"`
+
+
+### Turn 95 · 18:59:16
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "提示词 指令 生图参数"`
+
+
+### Turn 96 · 18:59:18
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --narrow "提示词 指令 创作规格"`
+
+
+### Turn 97 · 18:59:21
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "表象/视觉/符号/文字符号/标注/提示引导"`
+
+
+### Turn 98 · 18:59:23
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "操作指南 教程 AI使用方法"`
+
+
+### Turn 99 · 18:59:26
+- `Bash` — `cd "C:/Users/11304/glacier/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "方法技巧 知识经验 工具使用"`
+
+
+### Turn 100 · 18:59:27
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "AI应用 生成式AI 工具能力"`
+
+
+### Turn 101 · 18:59:33
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "AI应用 生成式AI 工具能力"`
+
+
+### Turn 102 · 18:59:36
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "理念/知识/商业/前沿技术/AI智能/AI应用"`
+
+
+### Turn 103 · 18:59:38
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "文字段落 说明文 文本内容"`
+
+
+### Turn 104 · 18:59:41
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "呈现/视觉/视觉制作/构图编排/版面设计/版面结构"`
+
+
+### Turn 105 · 18:59:46
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "图像 照片 静态画面 视觉内容"`
+
+
+### Turn 106 · 18:59:50
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "呈现/视觉/视觉制作/实景拍摄"`
+
+
+### Turn 107 · 18:59:51
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "AI生成 数字图像 生成式图像"`
+
+
+### Turn 108 · 18:59:55
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "呈现/视觉/视觉制作/后期处理/图像合成处理/AI生成合成"`
+
+
+### Turn 109 · 18:59:56
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "图像风格 视觉效果 审美规范"`
+
+
+### Turn 110 · 19:00:00
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "表象/视觉/画面/后期/美学风格"`
+
+
+### Turn 111 · 19:00:01
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "数据 结构化 内容描述"`
+
+
+### Turn 112 · 19:00:05
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --validate "架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"`
+
+
+### Turn 113 · 19:00:09
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "自然语言 自由文本 描述性写作"`
+
+
+### Turn 114 · 19:00:11
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "场景主体 创作内容 图像主题"`
+
+
+### Turn 115 · 19:00:15
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "山野 晨雾 光影 清晨"`
+
+
+### Turn 116 · 19:00:16
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "散文 写意 描述性文字"`
+
+
+### Turn 117 · 19:00:19
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "表象/视觉/画面/光影/光影效果"`
+
+
+### Turn 118 · 19:00:21
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "构图场景 视角镜头 意境氛围"`
+
+
+### Turn 119 · 19:00:25
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "镜头视角 场景 构图"`
+
+
+### Turn 120 · 19:00:27
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --match "风景图 自然风光 山水"`
+
+
+### Turn 121 · 19:00:31
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 实质 --validate "表象/视觉/空间/空间环境/地形地貌/山水景观"`
+
+
+### Turn 122 · 19:00:33
+- `Bash` — `cd "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl" && python spec/tools/taxonomy-lookup.py --dim 形式 --match "文字 关键词 标签"`
+
+
+### Turn 123 · 19:00:53
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/_scratch/patch_2b.json` (1,072 chars)
+
+
+### Turn 124 · 19:01:17
+
+
+### Turn 125 · 19:01:18
+> Both sub-agents done. Now applying both patches to workflow.json, then running resolve-passthrough:
+
+
+
+### Turn 126 · 19:01:20
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/wf-patch.py" --workflow "C:/Users/11304/gitlab/cybertogether/Agent/examples/proc...`
+
+
+### Turn 127 · 19:01:25
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/wf-patch.py" --workflow "C:/Users/11304/gitlab/cybertogether/Agent/examples/proc...`
+
+
+### Turn 128 · 19:01:30
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/wf-patch.py" --workflow "C:/Users/11304/gitlab/cybertogether/Agent/examples/proc...`
+
+
+### Turn 129 · 19:02:12
+
+
+### Turn 130 · 19:02:12
+> Now entering **Phase 3**. Running lint first:
+
+
+
+### Turn 131 · 19:02:12
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/lint-case.py" --workflow "C:/Users/11304/gitlab/cybertogether/Agent/examples/pro...`
+
+
+### Turn 132 · 19:03:11
+
+
+### Turn 133 · 19:03:11
+> Lint reports type registry OK and 1 value hint (the JSON分析's `格式同上` note may read as a reference placeholder). Let me fix the value and also need to fill in s2.inputs[0].value (passthrough from s1):
+
+
+
+### Turn 134 · 19:03:22
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/wf-patch.py" --workflow "C:/Users/11304/gitlab/cybertogether/Agent/examples/proc...`
+
+
+### Turn 135 · 19:03:27
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/wf-patch.py" --workflow "C:/Users/11304/gitlab/cybertogether/Agent/examples/proc...`
+
+
+### Turn 136 · 19:03:30
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/lint-case.py" --workflow "C:/Users/11304/gitlab/cybertogether/Agent/examples/pro...`
+
+
+### Turn 137 · 19:04:03
+
+
+### Turn 138 · 19:04:03
+> Lint clean! Now Phase 3.2 — update intents to reflect Phase 2 normalized action/type values (s3 out-type changed to 样图, action leaf names updated):
+
+
+
+### Turn 139 · 19:04:06
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/wf-patch.py" --workflow "C:/Users/11304/gitlab/cybertogether/Agent/examples/proc...`
+
+
+### Turn 140 · 19:04:16
+
+
+### Turn 141 · 19:04:16
+> Now render the HTML:
+
+
+
+### Turn 142 · 19:04:18
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/render-case.py" --workflow "C:/Users/11304/gitlab/cybertogether/Agent/examples/p...`
+
+
+### Turn 143 · 19:04:26
+
+
+### Turn 144 · 19:04:26
+> Need to remove `$comment` from workflow.json (schema doesn't allow additional properties). Let me check the schema and fix:
+
+
+
+### Turn 145 · 19:04:26
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/format/case-data.schema.json`
+
+
+### Turn 146 · 19:04:58
+
+
+### Turn 147 · 19:04:58
+> The `$comment` top-level key violates `additionalProperties:false`. Remove it via wf-patch --unset:
+
+
+
+### Turn 148 · 19:05:00
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/wf-patch.py" --workflow "C:/Users/11304/gitlab/cybertogether/Agent/examples/proc...`
+
+
+### Turn 149 · 19:05:08
+
+
+### Turn 150 · 19:05:08
+> wf-patch can't reach top-level keys. Use Edit to remove `$comment` (I know the exact content from the system-reminder):
+
+
+
+### Turn 151 · 19:05:09
+- `Edit` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/workflow.json`
+
+
+### Turn 152 · 19:05:16
+
+
+### Turn 153 · 19:05:16
+- `Read` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/workflow.json` (offset=0, limit=5)
+
+
+### Turn 154 · 19:05:20
+- `Edit` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/workflow.json`
+
+
+### Turn 155 · 19:05:24
+> Now render again:
+
+
+
+### Turn 156 · 19:05:27
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/render-case.py" --workflow "C:/Users/11304/gitlab/cybertogether/Agent/examples/p...`
+
+
+### Turn 157 · 19:05:57
+
+
+### Turn 158 · 19:05:57
+> HTML rendered (74,635 chars, 1 procedure, 3 steps). Now Phase 3.3 — write the .md file:
+
+
+
+### Turn 159 · 19:07:18
+- `Write` — `C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/case-B_gzh_25318352-doubao-style-extract.md` (5,102 chars)
+
+
+### Turn 160 · 19:07:28
+
+
+### Turn 161 · 19:07:28
+> 全部完成。最终运行一次 lint 确认状态：
+
+
+
+### Turn 162 · 19:07:30
+- `Bash` — `python "C:/Users/11304/gitlab/cybertogether/Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/lint-case.py" --workflow "C:/Users/11304/gitlab/cybertogether/Agent/examples/pro...`
+
+
+### Turn 163 · 19:07:46
+> ✅ 全部 lint 通过。以下是完整总结：
+>
+> ---
+>
+> ## 工序梗概
+>
+> **案例 B_gzh_25318352**（青山Linn @微信公众号）演示了一个三步风格迁移闭环：①上传参考图到豆包聊天模式，输入一句结构化提示词，让LLM以 JSON 提取视觉风格（颜色/排版/构图/特效四维）；②人工把选定的风格 JSON 与自拟的主体描述词拼成完整生图指令；③切换到豆包图像生成模式（Seedream 4.5），批量生成4张候选，手选最终结果。整体归为**1个工序**（p1），方法论统一，生成的多批图只是换了不同作品的风格提示词。
+>
+> ---
+>
+> ## 输出文件路径
+>
+> | 产物 | 路径 |
+> |---|---|
+> | 心智模型 | `.../B_gzh_25318352/understanding.md` |
+> | 工序骨架 | `.../B_gzh_25318352/workflow.json` (1 procedure / 3 steps / 完整标注) |
+> | HTML 可视化 | `.../B_gzh_25318352/case-B_gzh_25318352-doubao-style-extract.html` (74,635 chars) |
+> | DSL 文本版 | `.../B_gzh_25318352/case-B_gzh_25318352-doubao-style-extract.md` |
+>
+> ---
+>
+> ## 对 DSL 的关键发现
+>
+> 1. **"风格描述"作为独立的中间类型**：JSON风格分析不是直接可用的 prompt，而是视觉规约模板——DSL 用 case-specific `风格描述`（extends `描述`）比笼统用"提示词"更精确地区分了它的语义角色。
+>
+> 2. **同一工具双模态 = 不同 effect**：豆包 chat 模式（`预处理`，反推风格）和豆包 Seedream（`主体生成`，文生图）在同一 APP 内承担工序链的不同位置——证明 **via ≠ effect**，L1 工具名和 L3 工艺位置必须分列。
+>
+> 3. **预处理的两步子结构**：s1（反推提取规约）和 s2（人工编排指令）都属 `预处理`，但语义层次不同，说明 effect 树的 `预处理` 叶子粒度可能需要更细的子叶。
+>
+> 4. **`生成/元素生成` 跨 via 复用**：s2（human 创作 prompt）和 s3（AI 生图）都归属同一动作叶子 `生成/元素生成`——via 字段区分执行方式，action 字段描述动作性质，两者正交设计得到验证。
+
+
+
+### ◀ Result · 19:07:46
+
+- subtype: `success` · is_error: `False`
+- num_turns: `36` · duration: `1285100ms` · wall: `1287.4s`
+- tokens: in=37 out=55,668 cache_w=130,091 cache_r=3,364,847
+- cost: sdk=$3.5966, est_if_api=$2.3324

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/case-B_gzh_25318352-doubao-style-extract.md

@@ -0,0 +1,197 @@
+# Case B_gzh_25318352: 一句话提取图片风格迁移生图
+
+**Source**: 微信公众号·青山Linn — 超简单！一句提示词，帮你高效提取图片风格
+**URL**: https://mp.weixin.qq.com/s?__biz=MzkxNjg3OTYwNg==&mid=2247484017&idx=1&sn=731fa955e75f6f5cf614d54364548a7d
+**发布**: 2026-02-06
+
+---
+
+## 工序梗概 (人话)
+
+用豆包 (Doubao) 上传5张绘画风格参考图，输入一句结构化提示词让LLM以JSON格式提取视觉风格（颜色/排版/构图/特效四维度）；然后把选定作品的风格描述复制到豆包的图像生成对话框，拼接用户自拟的主体描述词，切换到 Seedream 4.5 图像生成模式，批量生成（默认4张候选）同风格的新图，由用户手选最终结果。
+
+**核心洞见**：本工序把"图像风格"拆解为一份可文本化的 JSON 规约，在豆包的同一 APP 内完成"提取→复用→生成"三步闭环，不需要用户懂底层 Stable Diffusion 或 LoRA 机制——只要能描述自己想要的主体，就能"抄"到别人图片的视觉风格。
+
+---
+
+## 引用的类型 (stdlib + 自定义)
+
+### stdlib 类型 (直接引用)
+
+| 类型名 | 使用位置 | 说明 |
+|---|---|---|
+| 参考图 | s1 输入 / declarations | 待提取风格的参考图片 |
+| 提示词 | s2 输入 (图像描述词) / s2 输出 / s3 输入 | 生图指令文本 |
+| 样图 | s3 输出 | AI生成的4张候选图，待人工筛选 |
+
+### case-specific 类型 (type_registry)
+
+| 类型名 | extends | desc |
+|---|---|---|
+| 风格描述 | 描述 | 豆包从参考图提取的JSON格式视觉风格描述，含颜色/排版/构图/特效四个维度，作为生图提示词的风格模板来源 |
+
+---
+
+## L1 外部函数库
+
+| 工具名 | 类型 | 本案例用途 |
+|---|---|---|
+| 豆包 (Doubao) | 多模态 LLM + 图像生成平台 | s1：聊天模式 (多图输入→LLM提取JSON)；s3：图像生成模式 (Seedream 4.5 文生图) |
+| human | — | s2：人工挑选风格并拼接生成提示词 |
+
+---
+
+## L2 抽象动作 + impl 关系
+
+| 抽象动作 | 树路径 | impl | 用于步骤 |
+|---|---|---|---|
+| 反推 | 提取/化学提取/反推 | impl via 豆包 (chat LLM 从图像推算视觉属性) | s1 |
+| 元素生成 | 生成/元素生成 | impl via human (拼接指令文本，产出完整 prompt 单元) | s2 |
+| 元素生成 | 生成/元素生成 | impl via 豆包 Seedream 4.5 (文生图，产出单张/批量图像) | s3 |
+
+> 注意：`元素生成` 被两步复用：s2 是人工元素生成（创作一条新 prompt），s3 是 AI 元素生成（从 prompt 创作图像），两者 via 不同，抽象动作相同。
+
+---
+
+## L3 工序模板
+
+**模板 A：图片风格提取与迁移生图**
+
+```
+procedure 图片风格提取与迁移生图(风格参考图: 参考图, 图像描述词: 提示词) -> 样图
+
+  declare inputs:
+    风格参考图: 参考图       -- 待提取视觉风格的参考图片（可单张或多张）
+    图像描述词: 提示词       -- 新图所需主体/场景/镜头/氛围描述词，由用户自拟
+
+  s1: via 豆包, action 提取/化学提取/反推, effect 预处理, feature 随机
+      directive: "以JSON结构化数据形式提取视觉风格，包括颜色、排版、构图和特效"
+      input:  风格参考图: 参考图          ← 工序输入
+      output: JSON风格分析: 风格描述     → s2
+
+  s2: via human, action 生成/元素生成, effect 预处理, feature 人工
+      input:  JSON风格分析: 风格描述     ← s1.JSON风格分析
+              图像描述词: 提示词          ← 工序输入 [inferred]
+      output: 图像生成提示词: 提示词     → s3
+
+  s3: via 豆包, action 生成/元素生成, effect 主体生成, feature 随机
+      config: Seedream 4.5 · 图像生成模式 · 比例 16:9
+      @采样(n=4, pick=人工)
+      input:  图像生成提示词: 提示词     ← s2.图像生成提示词
+      output: 风格化生成图: 样图         → 返回 样图
+
+  return 样图
+```
+
+---
+
+## L4 工序实例
+
+### P1 实例 — 作品1金蓝星空风格 × 清晨梯田主体
+
+#### inputs (本次实际值)
+
+| 变量名 | 类型 | 值 |
+|---|---|---|
+| 风格参考图 | 参考图 | 5张绘画风格图：①金蓝大漠骆驼队中国风星空（顶部CHINA大字+金箔质感）②浅蓝淡金小满人物麦浪（中式印章+薄雾意境）③灰白极简CHANEL时装秀（镜面反射+朦胧虚化）④深红柱廊寺庙ZENFO（对称透视+宏观建筑）⑤彩色抽象藏族女性与狼（高饱和插画风） |
+| 图像描述词 | 提示词 | 清晨群山与层叠梯田，中国南方山野清晨，薄雾未散，电影级远景推进镜头，略微俯拍，自然晨光，冷暖交替，安静克制，大师作品高品质高质量 [推断补全] |
+
+#### bindings (本次工具选择)
+
+```
+s1.via = 豆包 (聊天模式，多图上传 + LLM 文本输出)
+s2.via = human
+s3.via = 豆包 (图像生成模式，Seedream 4.5，比例 16:9)
+```
+
+#### extracted_values (本次中间产物 — prompt 原文回填)
+
+**s1 输入提取提示词 (directive)**：
+```
+以JSON结构化数据形式提取视觉风格，包括颜色、排版、构图和特效
+```
+
+**s1 输出 JSON风格分析 (图4截图内容)**：
+```
+作品编号: 1
+  颜色: "主色调为金、蓝、米黄，搭配深蓝星空质感，高对比度的复古华丽配色"
+  排版: "顶部大号CHINA英文+小字主题说明，文字与画面融合，中式印章点缀右上角"
+  构图: "前景驼队、中景沙丘、背景星空，层次分明的纵深感，对称与流动感结合"
+  特效: "金箔纹理质感，星空颗粒效果，沙丘线条的渐变晕染"
+
+作品编号: 2
+  颜色: "浅蓝、灰白、淡金，低饱和度的清冷莫兰迪色系，营造静谧氛围"
+  排版: "中部书法字体小满+英文说明，右上角中式印章，文字简洁留白"
+  构图: "前景人物、中景麦浪、远景山峦，中心人物+对称式山水层次，留白充足"
+  特效: "麦浪的线条流动感，画面的朦胧薄雾质感，低透明度的意境渲染"
+
+作品编号: 3
+  颜色: "灰白、浅灰、墨黑，极简的单色调，冷感高级灰为主"
+  排版: "中部粗体CHANEL英文，无多余文字，品牌标识突出"
+  构图: "中心模特+背景建筑，蜿蜒步道引导视觉，对称式场景布局"
+  特效: "镜面反射效果，薄雾的朦胧虚化，极简空间的层次（图中文字截断）"
+
+作品编号: 4-5（图4中仅展示前3件，4-5数据在图4屏外截断）
+```
+
+**s2 输出 图像生成提示词 (图5截图内容)**：
+```
+帮我生成图片：颜色："主色调为金、蓝、米黄，搭配深蓝星空质感，高对比度的复古华丽配色"，
+排版："顶部大号'CHINA'英文+小字主题说明，文字与画面融合，中式印章点缀右上角"，
+构图："前景驼队、中景沙丘、背景星空，层次分明的纵深感，对称与流动感结合"，
+特效："金箔纹理质感，星空颗粒效果，沙丘线条的渐变晕染"，
+比例「16:9」**主体：**清晨群山与层叠梯田。
+**场景与时间：**中国南方山野清晨，薄雾未散。
+**镜头与视角：**电影级远景推进镜头，略微俯拍。
+**光影与氛围：**自然晨光，冷暖交替，安静克制。
+大师作品，高品质，高质量，电影级场景
+```
+
+**s3 配置 (图5截图)**：
+```
+图像生成 × (模式激活标签)
+Seedream 4.5
+比例: 16:9
+参考图: 未附
+```
+
+**s3 输出 风格化生成图 (图6)**：
+```
+4张AI生成图，均带"AI生成"水印。
+16:9横幅构图，以金/蓝/米黄为主色调，
+CHINA英文字以金箔纹理融入层叠梯田与星空背景，
+清晨薄雾光影，山野梯田，复古华丽中国风；
+各张在构图细节（文字大小/位置、梯田角度）上有差异，供人工挑选。
+```
+
+#### trace
+
+```
+timing: 不详（文章未披露）
+cost: 豆包免费额度内（作者注明: 免登录视频4/图片10；有账号配额更多）
+retry_log: 文章提及"可以尝试其他的风格提示词"（对不满意结果换风格重试），本次记录为首次尝试
+```
+
+---
+
+## 这个 case 对 DSL 设计的关键启发
+
+### 1. "风格描述"作为独立中间类型的价值
+
+JSON风格分析不是直接可用的 prompt，而是一份**视觉风格规约**（颜色/排版/构图/特效四维描述），需要人工拼接主体描述词才能成为完整生图指令。DSL 中引入 `风格描述` 这个 case-specific 类型（extends `描述`），比笼统地用 `提示词` 更精确地表达了它的语义角色——**它是生图提示词的风格部分模板，不是完整的生成指令**。
+
+### 2. 同一工具双模态承担不同 effect
+
+豆包同一 APP 的聊天模式（effect = `预处理`）和图像生成模式（effect = `主体生成`）在工序链中处于完全不同的位置。这揭示：**L1 工具（via）不等于 L2 动作（action）和 L3 effect**——一个品牌名可以在工序链的不同阶段承担不同的工艺作用。via 字段记录"谁做"，effect 字段记录"在工序的哪个环节"，两者解耦是关键。
+
+### 3. 预处理阶段的两步结构
+
+s1（反推提取风格）和 s2（人工拼接指令）都归属 `预处理`，但语义不同：s1 是"从已有作品推算风格规约"，s2 是"把规约组合成可执行指令"。这说明预处理阶段本身可以有子阶段：**提取规约 → 编排指令 → 主体生成**。当前 effect 树的 `预处理` 叶子粒度可能不足以区分这两种不同的预处理子角色。
+
+### 4. 人工步骤的 action 归类
+
+s2 的 action 归为 `生成/元素生成`（人工产出一个新的提示词单元）。这揭示了一个有趣的对称性：人工创作一个 prompt 和 AI 生成一张图像，在 DSL 层面都是"生成一个新的单元"——**动作树 `生成/元素生成` 对 via=human 和 via=AI 都适用，via 字段才是区分执行方式的关键**。
+
+### 5. 推断补全的合理边界
+
+`图像描述词`（用户自拟的主体描述）标了 `inferred: true`，因为原文只提到"输入刚刚得到的风格提示词和你想要的图片描述词"，没有把它单独列为一个声明输入。但从 s2 输出的生图提示词的后半段（主体/场景/镜头/光影部分）可以确定它确实存在，只是原文没有单独展示。这是**原文的"隐式参数"**，在 DSL 中显式化为 declarations 的第二个输入，有助于读者理解该工序的完整依赖。

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/understanding.md

@@ -0,0 +1,131 @@
+# Phase 1.1 · 心智模型 — B_gzh_25318352
+
+## 来源
+
+- 标题：超简单！一句提示词，帮你高效提取图片风格
+- 作者：青山Linn
+- 平台：微信公众号 (wechat-gzh)
+- 发布：2026-02-06
+
+---
+
+## 工序数量判断
+
+**扫描终态产物图**（以每张成品图为起点逐一检查）：
+
+| 产物图 | 描述 | 方法链 |
+|---|---|---|
+| 图6（4张CHINA梯田） | AI生成 16:9，金蓝星空+层叠梯田，CHINA大字 | 豆包Chat提取风格→拼接提示词→豆包Seedream生图 |
+| 图7（山水油画风） | AI生成，蓝金山脉梯田，抽象感 | 同一方法（风格不同） |
+| 图8（4张麦穗+嫦娥） | AI生成，中国仙女麦田神话风 | 同一方法 |
+| 图9（4张水晶麦穗） | AI生成，玉石水晶宝石质感 | 同一方法 |
+
+**判断：1个工序（p1）**
+
+- 所有终态产物均用同一套方法论产出：①豆包Chat上传图+提取提示词→JSON风格分析 ②人工拼接生图指令 ③豆包Seedream生成图
+- 不同生成结果只是选用了不同参考图的风格描述词；技术路径、工具链完全一致
+- 按多工序判断标准："产出物类型相同+方法相同" → 不拆分为多工序
+
+---
+
+## 工序 P1 · 图片风格提取与迁移生图
+
+| 维度 | 内容 |
+|---|---|
+| **终态产物** | 风格化生成图（以作品1金蓝星空大漠风格为例：CHINA梯田系列4张 AI生成图） |
+| **大致步骤数** | 3步（提取风格 → 拼接生图提示词 → 生成图片） |
+| **工艺类型** | 产物创造（图像风格迁移+文生图） |
+
+### 核心工艺流
+
+```
+[外部输入] 风格参考图 (5张绘画) + 图像描述词 (主体/场景)
+     ↓
+ s1: 豆包Chat × 提取提示词 → JSON格式视觉风格分析 (颜色/排版/构图/特效)
+     ↓
+ s2: human × 挑选目标风格 + 拼接图像描述词 → 图像生成提示词
+     ↓
+ s3: 豆包(Seedream 4.5) × 图像生成模式 → 风格化生成图 (4张候选)
+     ↓
+[返回] 风格化生成图
+```
+
+### 外部工具依赖
+
+- **豆包 (Doubao)**: 双模式—— ①Chat模式（多模态LLM，上传图+提示词提取风格）；②图像生成模式（Seedream 4.5，文生图）
+
+### 工艺规约/预准备
+
+- 豆包Chat模式和图像生成模式是同一APP的不同对话类型，需在s2→s3间切换
+- 生成时系统默认出4张候选图（图6/8/9均为2×2网格），由用户手动挑选
+- 提取的JSON风格规约只描述视觉参数（颜色/排版/构图/特效），不锁定工具/模型，理论上可在其他生图工具复用
+
+---
+
+## 关键数据（从配图读取）
+
+### 图3 · 5张参考图 + 提取提示词
+
+**参考图描述**（作品1~5）：
+- 作品1：金、蓝、米黄色调，大漠骆驼队+背景星空，CHINA英文顶部大字，中式印章，金箔纹理质感
+- 作品2：浅蓝灰淡金，"小满"书法字，麦浪人物中国山水，薄雾意境渲染
+- 作品3：灰白极简，粗体CHANEL英文，时装秀场景，镜面反射+薄雾虚化
+- 作品4：深红柱廊寺庙ZENFO，对称透视，宏观建筑场景
+- 作品5：彩色抽象，藏族女性+狼，高饱和插画风
+
+**提取提示词（s1 directive）**：
+`以JSON结构化数据形式提取视觉风格，包括颜色、排版、构图和特效`
+
+### 图4 · JSON风格分析输出（s1 output）
+
+```
+作品编号: 1
+  颜色: "主色调为金、蓝、米黄，搭配深蓝星空质感，高对比度的复古华丽配色"
+  排版: "顶部大号"CHINA"英文+小字主题说明，文字与画面融合，中式印章点缀右上角"
+  构图: "前景驼队、中景沙丘、背景星空，层次分明的纵深感，对称与流动感结合"
+  特效: "金箔纹理质感，星空颗粒效果，沙丘线条的渐变晕染"
+
+作品编号: 2
+  颜色: "浅蓝、灰白、淡金，低饱和度的清冷莫兰迪色系，营造静谧氛围"
+  排版: "中部书法字体"小满"+英文说明，右上角中式印章，文字简洁留白"
+  构图: "前景人物、中景麦浪、远景山峦，中心人物+对称式山水层次，留白充足"
+  特效: "麦浪的线条流动感，画面的朦胧薄雾质感，低透明度的意境渲染"
+
+作品编号: 3
+  颜色: "灰白、浅灰、墨黑，极简的单色调，冷感高级灰为主"
+  排版: "中部粗体"CHANEL"英文，无多余文字，品牌标识突出"
+  构图: "中心模特+背景建筑，蜿蜒步道引导视觉，对称式场景布局"
+  特效: "镜面反射效果，薄雾的朦胧虚化，极简空间的层次（图中截断）"
+
+（作品4/5内容在图4中未完整显示，推测格式一致）
+```
+
+### 图5 · 生图提示词（s2 output / s3 input）
+
+来自豆包图像生成对话框截图，完整文本：
+
+```
+帮我生成图片：颜色："主色调为金、蓝、米黄，搭配深蓝星空质感，高对比度的复古华丽配色"，
+排版："顶部大号'CHINA'英文+小字主题说明，文字与画面融合，中式印章点缀右上角"，
+构图："前景驼队、中景沙丘、背景星空，层次分明的纵深感，对称与流动感结合"，
+特效："金箔纹理质感，星空颗粒效果，沙丘线条的渐变晕染"，
+比例「16:9」**主体：**清晨群山与层叠梯田。
+**场景与时间：**中国南方山野清晨，薄雾未散。
+**镜头与视角：**电影级远景推进镜头，略微俯拍。
+**光影与氛围：**自然晨光，冷暖交替，安静克制。
+大师作品，高品质，高质量，电影级场景
+```
+
+### 图5 · 生图工具配置
+
+- 模式：图像生成（图像生成 × 标签激活）
+- 模型：Seedream 4.5
+- 比例：（用户选择，示例为16:9）
+- 参考图：（支持附参考图，图5中未使用）
+
+---
+
+## 推断补全说明
+
+- s2的 `图像描述词` 输入（用户自定义主体/场景）：**原文未列出独立声明**，但从生图提示词中"主体/场景/镜头/光影"部分可反推存在，标注 `inferred: true`
+- s3 的 `@采样(n=4, pick=人工)`：图6/8/9均展示2×2四张候选图，可确认批量生成4张；标注 decorator

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_25318352/workflow.json

@@ -0,0 +1,185 @@
+{
+  "source": {
+    "platform": "wechat-gzh",
+    "author": "青山Linn",
+    "date": "2026-02-06",
+    "url": "https://mp.weixin.qq.com/s?__biz=MzkxNjg3OTYwNg==&mid=2247484017&idx=1&sn=731fa955e75f6f5cf614d54364548a7d&chksm=c0506a0c7b384a5e4d28dee568d5575bb765c264e1376329b28c9b0b7786218cb2f7eb803366#rd",
+    "title": "超简单！一句提示词，帮你高效提取图片风格",
+    "excerpt": "用豆包上传参考图，一句提示词提取JSON格式视觉风格（颜色/排版/构图/特效），复用风格提示词驱动Seedream生成同风格新图"
+  },
+  "procedures": [
+    {
+      "id": "p1",
+      "name": "图片风格提取与迁移生图",
+      "purpose": "用豆包一句话从参考图提取JSON视觉风格，复用为图像生成提示词，快速生成同风格新图",
+      "category": "产物创造",
+      "platform": "wechat-gzh",
+      "author": "青山Linn",
+      "declarations": {
+        "inputs": [
+          {
+            "type": "参考图",
+            "name": "风格参考图",
+            "desc": "待提取视觉风格的参考图片（可单张或多张，本案例上传5张绘画风格图）"
+          },
+          {
+            "type": "提示词",
+            "name": "图像描述词",
+            "desc": "新图所需的主体/场景/镜头/氛围描述词，由用户自拟，与提取的风格拼接成最终生图指令"
+          }
+        ],
+        "resources": [],
+        "returns": {
+          "type": "生成图"
+        }
+      },
+      "type_registry": {
+        "风格描述": {
+          "extends": "描述",
+          "desc": "豆包从参考图提取的JSON格式视觉风格描述，含颜色/排版/构图/特效四个维度，作为生图提示词的风格模板来源"
+        }
+      },
+      "steps": [
+        {
+          "id": "s1",
+          "kind": "step",
+          "via": "豆包",
+          "action": "提取/化学提取/反推",
+          "feature": "随机",
+          "control": "-",
+          "instruction": [
+            [
+              "directive",
+              "以JSON结构化数据形式提取视觉风格，包括颜色、排版、构图和特效"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "参考图",
+              "name": "风格参考图",
+              "value": "<5张绘画风格图：作品1金蓝大漠骆驼队中国风星空（顶部CHINA大字+金箔质感）/作品2浅蓝淡金小满人物麦浪（中式印章+薄雾意境）/作品3灰白极简CHANEL时装秀（镜面反射+朦胧虚化）/作品4深红柱廊寺庙ZENFO（对称透视+宏观建筑）/作品5彩色抽象藏族女性与狼（高饱和插画风）>",
+              "anchor": "← 工序输入",
+              "substance": "/表象/视觉/实体/物品/影像物品/艺术创作",
+              "form": "/呈现/视觉/视觉风格/视觉艺术风格/绘画风格/插画"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "风格描述",
+              "name": "JSON风格分析",
+              "value": "作品编号: 1, 颜色: \"主色调为金、蓝、米黄，搭配深蓝星空质感，高对比度的复古华丽配色\", 排版: \"顶部大号CHINA英文+小字主题说明，文字与画面融合，中式印章点缀右上角\", 构图: \"前景驼队、中景沙丘、背景星空，层次分明的纵深感，对称与流动感结合\", 特效: \"金箔纹理质感，星空颗粒效果，沙丘线条的渐变晕染\" || 作品编号: 2, 颜色: \"浅蓝、灰白、淡金，低饱和度的清冷莫兰迪色系，营造静谧氛围\", 排版: \"中部书法字体小满+英文说明，右上角中式印章，文字简洁留白\", 构图: \"前景人物、中景麦浪、远景山峦，中心人物+对称式山水层次，留白充足\", 特效: \"麦浪的线条流动感，画面的朦胧薄雾质感，低透明度的意境渲染\" || 作品编号: 3, 颜色: \"灰白、浅灰、墨黑，极简的单色调，冷感高级灰为主\", 排版: \"中部粗体CHANEL英文，无多余文字，品牌标识突出\", 构图: \"中心模特+背景建筑，蜿蜒步道引导视觉，对称式场景布局\", 特效: \"镜面反射效果，薄雾的朦胧虚化，极简空间的层次（图中文字截断）\" || 作品编号: 4-5（图4中仅展示前3件，4-5数据在图4屏外截断）",
+              "anchor": "→ s2",
+              "substance": "/表象/视觉/画面/后期/美学风格",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            }
+          ],
+          "intent": "用{via:豆包}{act:反推}从{in-type:参考图}提取{out-type:风格描述}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "预处理"
+        },
+        {
+          "id": "s2",
+          "kind": "step",
+          "via": "human",
+          "action": "生成/元素生成",
+          "feature": "人工",
+          "control": "-",
+          "instruction": [],
+          "inputs": [
+            {
+              "type": "风格描述",
+              "name": "JSON风格分析",
+              "value": "作品编号: 1, 颜色: \"主色调为金、蓝、米黄，搭配深蓝星空质感，高对比度的复古华丽配色\", 排版: \"顶部大号CHINA英文+小字主题说明，文字与画面融合，中式印章点缀右上角\", 构图: \"前景驼队、中景沙丘、背景星空，层次分明的纵深感，对称与流动感结合\", 特效: \"金箔纹理质感，星空颗粒效果，沙丘线条的渐变晕染\" || 作品编号: 2, 颜色: \"浅蓝、灰白、淡金，低饱和度的清冷莫兰迪色系，营造静谧氛围\", 排版: \"中部书法字体小满+英文说明，右上角中式印章，文字简洁留白\", 构图: \"前景人物、中景麦浪、远景山峦，中心人物+对称式山水层次，留白充足\", 特效: \"麦浪的线条流动感，画面的朦胧薄雾质感，低透明度的意境渲染\" || 作品编号: 3, 颜色: \"灰白、浅灰、墨黑，极简的单色调，冷感高级灰为主\", 排版: \"中部粗体CHANEL英文，无多余文字，品牌标识突出\", 构图: \"中心模特+背景建筑，蜿蜒步道引导视觉，对称式场景布局\", 特效: \"镜面反射效果，薄雾的朦胧虚化，极简空间的层次（图中文字截断）\" || 作品编号: 4-5（图4中仅展示前3件，4-5数据在图4屏外截断）",
+              "anchor": "← s1.JSON风格分析",
+              "substance": "/表象/视觉/画面/后期/美学风格",
+              "form": "/架构/逻辑/信息编排逻辑/条目列举/分段结构/规范框架"
+            },
+            {
+              "type": "提示词",
+              "name": "图像描述词",
+              "value": "<用户自拟的新图主体描述：清晨群山与层叠梯田，中国南方山野清晨，薄雾未散，电影级远景推进镜头，略微俯拍，自然晨光，冷暖交替，安静克制，大师作品高品质高质量>",
+              "anchor": "← 工序输入",
+              "inferred": true,
+              "inferred_reason": "原文只描述了拷贝风格提示词和输入图像描述词两个动作，用户自拟的主体描述词从图5生图提示词后半段（主体/场景/镜头/光影）反推补出，原文未单独列出此输入",
+              "substance": "/表象/视觉/空间/空间环境/地形地貌/山水景观",
+              "form": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "提示词",
+              "name": "图像生成提示词",
+              "value": "帮我生成图片：颜色：'主色调为金、蓝、米黄，搭配深蓝星空质感，高对比度的复古华丽配色'，排版：'顶部大号\"CHINA\"英文+小字主题说明，文字与画面融合，中式印章点缀右上角'，构图：'前景驼队、中景沙丘、背景星空，层次分明的纵深感，对称与流动感结合'，特效：'金箔纹理质感，星空颗粒效果，沙丘线条的渐变晕染'，比例「16:9」**主体：**清晨群山与层叠梯田。**场景与时间：**中国南方山野清晨，薄雾未散。**镜头与视角：**电影级远景推进镜头，略微俯拍。**光影与氛围：**自然晨光，冷暖交替，安静克制。大师作品，高品质，高质量，电影级场景",
+              "anchor": "→ s3",
+              "substance": "/理念/知识/商业/前沿技术/AI智能/AI应用",
+              "form": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"
+            }
+          ],
+          "intent": "用{via:human}合并{in-type:风格描述}与{in-type:提示词}得{out-type:提示词}",
+          "focus": [
+            "via",
+            "in-type-0",
+            "in-type-1",
+            "out-type-0"
+          ],
+          "effect": "预处理"
+        },
+        {
+          "id": "s3",
+          "kind": "step",
+          "via": "豆包",
+          "action": "生成/元素生成",
+          "feature": "随机",
+          "control": "-",
+          "instruction": [
+            [
+              "config",
+              "Seedream 4.5 · 图像生成模式 · 比例 16:9"
+            ],
+            [
+              "decorator",
+              "@采样(n=4, pick=人工)"
+            ]
+          ],
+          "inputs": [
+            {
+              "type": "提示词",
+              "name": "图像生成提示词",
+              "value": "帮我生成图片：颜色：'主色调为金、蓝、米黄，搭配深蓝星空质感，高对比度的复古华丽配色'，排版：'顶部大号\"CHINA\"英文+小字主题说明，文字与画面融合，中式印章点缀右上角'，构图：'前景驼队、中景沙丘、背景星空，层次分明的纵深感，对称与流动感结合'，特效：'金箔纹理质感，星空颗粒效果，沙丘线条的渐变晕染'，比例「16:9」**主体：**清晨群山与层叠梯田。**场景与时间：**中国南方山野清晨，薄雾未散。**镜头与视角：**电影级远景推进镜头，略微俯拍。**光影与氛围：**自然晨光，冷暖交替，安静克制。大师作品，高品质，高质量，电影级场景",
+              "anchor": "← s2.图像生成提示词",
+              "substance": "/理念/知识/商业/前沿技术/AI智能/AI应用",
+              "form": "/呈现/视觉/视觉制作/构图编排/版面设计/版面结构"
+            }
+          ],
+          "outputs": [
+            {
+              "type": "样图",
+              "name": "风格化生成图",
+              "value": "<AI生成的16:9图像组(4张候选)，以金/蓝/米黄为主色调，CHINA英文字融入层叠梯田与星空背景，清晨薄雾光影，复古华丽中国风；带\"AI生成\"水印>",
+              "anchor": "→ 返回 风格化生成图",
+              "substance": "/表象/视觉/空间/空间环境/地形地貌/山水景观",
+              "form": "/呈现/视觉/视觉制作/后期处理/图像合成处理/AI生成合成"
+            }
+          ],
+          "intent": "用{via:豆包}{act:元素生成}从{in-type:提示词}生成{out-type:样图}",
+          "focus": [
+            "via",
+            "action",
+            "in-type-0",
+            "out-type-0"
+          ],
+          "effect": "主体生成"
+        }
+      ],
+      "return_row": {
+        "arrow": "↩",
+        "text": "<span class=\"kw\">返回</span> <span class=\"chip\" data-type=\"生成图\">生成图</span>"
+      }
+    }
+  ]
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_e9677a53/.session_id

@@ -0,0 +1 @@
+32b84d86-b99a-4c46-ba6c-cf5ecf73b379

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_e9677a53/_meta.json

@@ -0,0 +1,8 @@
+{
+  "case_id": "gzh_e9677a53351f09f8ea97682c1e48fa01",
+  "from_q": "q0018",
+  "form": "B",
+  "score": 8.4,
+  "score_norm": 0.84,
+  "started_at": "2026-06-03T18:43:59"
+}

examples/process_pipeline/script/search_eval/runs_full/q0018/procedures/B_gzh_e9677a53/_scratch/anchors.json

@@ -0,0 +1,7 @@
+[
+  {"path": "p1.s1.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p1.s1.outputs[0].anchor", "value": "→ s2"},
+  {"path": "p1.s2.inputs[0].anchor",  "value": "← 工序输入"},
+  {"path": "p1.s2.inputs[1].anchor",  "value": "← s1.风格JSON"},
+  {"path": "p1.s2.outputs[0].anchor", "value": "→ 返回 图片"}
+]