argo-search 2.9.7 → 2.9.8

package/scripts/job.py

@@ -895,13 +895,13 @@ def _is_quota_exhausted(exc: Exception) -> bool:
     的模块静默生效。403/429 各有归属：engine_failure.classify 分 auth /
     blocked / rate_limited，健康问题交熔断器自愈，本函数不接手。
 
-    关键词表复用 engine_dispatch 的 `_QUOTA_ERROR_KEYWORDS`（该表自述为
-    「唯一来源」，配额分类与自适应跳过共用），不另起一份。
+    关键词表真源是 cache_guard 的 `_QUOTA_ERROR_KEYWORDS`（配额分类与自适应
+    跳过共用），不另起一份。
     真欠费的文案（`You do not have enough money or package quota`）含
     `quota`，仍然照常命中。
     """
     try:
-        from engine_dispatch import _QUOTA_ERROR_KEYWORDS
+        from cache_guard import _QUOTA_ERROR_KEYWORDS
     except ImportError:
         return False
     text = str(exc).lower()

package/scripts/local_seek.py

@@ -151,6 +151,37 @@ def _seek_disk_cache_put(key: str, results: list[dict[str, Any]]) -> None:
         pass
 
 
+# 本地代码/工程/配置意图正则模式
+_LOCAL_INTENT_PATTERN = re.compile(
+    r"(?i)(?:\b(?:def|class|function|fn|import|export|const|var|struct|enum)\b|"
+    r"\.(?:py|js|ts|json|yaml|yml|md|rs|go|c|cpp|h|sh|sql)\b|"
+    r"(?:readme|config|package\.json|cargo\.toml|dockerfile|makefile)|"
+    r"(?:本工程|本项目|代码库|本地文件|架构实现|函数|方法|源码|接口定义))"
+)
+
+
+def has_local_code_intent(query: str) -> bool:
+    """是否包含本地工程/代码/配置检索意图。"""
+    if not query:
+        return False
+    return bool(_LOCAL_INTENT_PATTERN.search(query))
+
+
+def resolve_include_local(include_local: bool | None, mode: str, query: str = "") -> bool:
+    """include-local 三态解析：True/False 显式，None=自动。
+
+    自适应感知：除 fast/budget 模式外，当查询具有明显的本地代码/文件特征
+    且处于非宽泛根项目目录时，自动开启本地命中。
+    """
+    if include_local is not None:
+        return bool(include_local)
+    if mode in ("fast", "budget"):
+        return True
+    if has_local_code_intent(query) and _local_seek_dir() is not None:
+        return True
+    return False
+
+
 def _is_broad_local_root(path: str) -> bool:
     """path 是否是「不能当作本地检索范围」的宽泛根。
 
@@ -200,7 +231,8 @@ def _local_seek_dir() -> str | None:
 
 
 def _run_local_seek(query: str, max_n: int = 5,
-                    search_dir: str | None = None) -> list[dict[str, Any]]:
+                    search_dir: str | None = None,
+                    warn: bool = True) -> list[dict[str, Any]]:
     """本机文件命中（include-local 用）：进程内调 seek.run_query，JSON 并入。
 
     fast/budget 模式下由 super_search 自动调用（_resolve_include_local），
@@ -215,10 +247,14 @@ def _run_local_seek(query: str, max_n: int = 5,
         # 静默跳过会让用户以为本地没匹配，而事实是根本没查。返回一个
         # 占位「结果」又会污染 results（它只是个说明，不是命中），所以
         # 说明走 stderr：与 CLI 既有的 include-local 异常提示同一条通道。
-        sys.stderr.write(
-            "  [include-local] 当前目录不适合本地检索（home / 根目录过宽，或位于 argo "
-            "自身安装树内——只会命中自身源码），已跳过；"
-            "要搜本机文件请 cd 到目标目录后重试，或用 local-seek 的 seek.py --path <目录>\n")
+        # 只在**显式开启**时提示（warn=False 是 fast/budget 的自动路径）：
+        # 自动路径在 home 等宽根下每次 search 都会打印，Agent 会把它当成
+        # 输出的一部分，属于噪音而非归因。
+        if warn:
+            sys.stderr.write(
+                "  [include-local] 当前目录不适合本地检索（home / 根目录过宽，或位于 argo "
+                "自身安装树内——只会命中自身源码），已跳过；"
+                "要搜本机文件请 cd 到目标目录后重试，或用 local-seek 的 seek.py --path <目录>\n")
         return []
 
     # 缓存：同 query+target 在 TTL 内直接命中，避免重复冷启动 seek.py 子进程。

package/scripts/macro_countries.py

@@ -66,6 +66,12 @@ _NON_US_COUNTRIES: dict[str, str] = {
     name: code for name, code in WORLDBANK_COUNTRIES.items() if code != "USA"
 }
 
+# 美国子集：给「统计局/国内」这类中国语境别名做前置排除。
+# 同一张主表派生，避免「美国 统计局」被当成中国口径。
+_US_COUNTRIES: dict[str, str] = {
+    name: code for name, code in WORLDBANK_COUNTRIES.items() if code == "USA"
+}
+
 
 def is_foreign_macro_query(query: str) -> bool:
     """查询是否明确指向非美国国家/地区。
@@ -78,5 +84,14 @@ def is_foreign_macro_query(query: str) -> bool:
     英文词——实测 `focus on growth`、`consensus 预测`、`business survey`
     全被判成「外国宏观查询」，FRED 因而被无谓地排除。同一个概念（国家词识别）
     在同一模块里有两套行为不同的实现，正是这类漏判的来源。
+
+    补充（2026-10-23）：「国家统计局」「国内」「我国」是明确的中国语境，
+    但不在国家名表里。实测「国家统计局 一季度 GDP」因此没让位，FRED 领队
+    返回美国 GDP。这几个词只指向中国，直接判为 foreign（非美国）。
+    但「美国 统计局」「美国国内 CPI」同时含美国词，必须先排除美国口径。
     """
+    q = query or ""
+    if (not match_country(q, _US_COUNTRIES)
+            and any(term in q for term in ("国家统计局", "统计局", "国内", "我国"))):
+        return True
     return bool(match_country(query, _NON_US_COUNTRIES))

package/scripts/matrix_search_eval.py

@@ -254,6 +254,21 @@ ROUTE_MATRIX: list[dict[str, Any]] = [
      "domain": "org_entity", "primary": "wikidata", "scenario": "org"},
     {"id": "R_zh_org2", "q": "清华大学 创办年份", "lang": "zh",
      "domain": "org_entity", "primary": "wikidata", "scenario": "org"},
+    # 中文企业总部不是地理问题（2026-10-22 修：geo_places 的方位词 unless 增补
+    # 「总部|母公司|子公司|注册地|办公地」后落到本域；此前落 geo_places，用 OSM+维基
+    # 只拿到广汽/新凯来/五粮液等无关结果）
+    {"id": "R_zh_org_hq", "q": "华为 总部在哪", "lang": "zh",
+     "domain": "org_entity", "primary": "wikidata", "scenario": "org"},
+
+    # ── 企业/工商（company_search）──
+    # 2026-10-22 补：此前 ROUTE_MATRIX 无本域用例，导致「中国企业股权/工商信息」
+    # 被路由到本域、而该域 combo 只有国际法人库（gleif/opencorporates）+ 百科的
+    # 问题长期无人发现——实测「长亭科技 股东 持股比例」返回 5 条同名「XX科技
+    # 有限公司」干扰。元宝进 combo 第 1 位后（a5a46ad）才具中文交付能力。
+    {"id": "R_zh_company_shareholder", "q": "长亭科技 股东 持股比例", "lang": "zh",
+     "domain": "company_search", "primary": "opencorporates", "scenario": "company"},
+    {"id": "R_zh_company_regcap", "q": "字节跳动 注册资本 变更", "lang": "zh",
+     "domain": "company_search", "primary": "opencorporates", "scenario": "company"},
 
     # ── 媒体/音乐 ──
     {"id": "R_en_media", "q": "Taylor Swift album", "lang": "en",
@@ -384,6 +399,73 @@ ROUTE_MATRIX: list[dict[str, Any]] = [
                 "worldbank", "nbs_stats", "un_comtrade", "endoflife",
                 "deps_dev", "openalex", "europepmc"],
      "forbid_cn": True, "scenario": "negative"},
+
+    # ── 域覆盖（2026-10-22）：补 53 个此前零用例的域 ────────────────────────
+    # 零用例 = 路由回归无人挡。company_search 零用例，「中国企业股权/工商信息」
+    # 被路由到本域却无中文能力的问题因此潜伏 8 天；geo_places 的「在哪」劫持
+    # 企业总部查询同理（两者都是靠人工撞出来的）。lang 取路由实测 primary_lang，
+    # 与 _route_ok 的 lang 断言对齐；primary 取域声明的主引擎。
+    {"id": "R_ai_model", "q": "models.dev claude 价格 上下文", "lang": "en", "domain": "ai_model", "primary": "models_dev", "scenario": "tech"},
+    {"id": "R_aviation_weather", "q": "北京首都机场 航空气象 METAR", "lang": "zh", "domain": "aviation_weather", "primary": "aviation_weather", "scenario": "weather"},
+    {"id": "R_chinese_tech_deep", "q": "中文 编程 算法 架构 深度对比", "lang": "zh", "domain": "chinese_tech_deep", "primary": "byted", "scenario": "tech"},
+    {"id": "R_claim_check", "q": "辟谣 传言 事实核查", "lang": "zh", "domain": "claim_check", "primary": "factcheck_org", "scenario": "factcheck"},
+    {"id": "R_cn_encyclopedia", "q": "长征 百度百科 词条解释", "lang": "zh", "domain": "cn_encyclopedia", "primary": "baidu_baike", "scenario": "media"},
+    {"id": "R_cn_tech_community", "q": "掘金 前端文章 技术博客", "lang": "zh", "domain": "cn_tech_community", "primary": "juejin", "scenario": "tech"},
+    {"id": "R_code_search", "q": "github 代码 实现 用法", "lang": "zh", "domain": "code_search", "primary": "github", "scenario": "code"},
+    {"id": "R_crypto_search", "q": "比特币 以太坊 加密货币 价格", "lang": "zh", "domain": "crypto_search", "primary": "coingecko", "scenario": "finance"},
+    {"id": "R_dataset_search", "q": "科研数据集 开放数据 下载", "lang": "zh", "domain": "dataset_search", "primary": "datacite", "scenario": "academic"},
+    {"id": "R_dictionary_search", "q": "英文单词 词典 意思", "lang": "zh", "domain": "dictionary_search", "primary": "free_dictionary", "scenario": "media"},
+    {"id": "R_earth_science", "q": "地震 震级 余震 震中", "lang": "zh", "domain": "earth_science", "primary": "usgs", "scenario": "science"},
+    {"id": "R_em_news_search", "q": "东财 全球资讯 市场", "lang": "zh", "domain": "em_news_search", "primary": "em_global_news", "scenario": "news"},
+    {"id": "R_energy_grid", "q": "电网 发电结构 发电量", "lang": "zh", "domain": "energy_grid", "primary": "energy_charts", "scenario": "science"},
+    {"id": "R_entity_search", "q": "维基数据 知识图谱 实体查询", "lang": "zh", "domain": "entity_search", "primary": "wikidata", "scenario": "media"},
+    {"id": "R_fact_check", "q": "长江 长度 是多少", "lang": "zh", "domain": "fact_check", "primary": "wikipedia", "scenario": "factcheck"},
+    {"id": "R_futures_market", "q": "沪铜 期货 行情", "lang": "zh", "domain": "futures_market", "primary": "shfe", "scenario": "finance"},
+    {"id": "R_fx_query", "q": "人民币 汇率 中间价", "lang": "zh", "domain": "fx_query", "primary": "fx_rate", "scenario": "finance"},
+    {"id": "R_game_search", "q": "steam 游戏推荐 好玩的游戏", "lang": "zh", "domain": "game_search", "primary": "steam", "scenario": "media"},
+    {"id": "R_global_event", "q": "国际事件 地缘政治 局势", "lang": "zh", "domain": "global_event", "primary": "google_news", "scenario": "news"},
+    {"id": "R_hk_filings", "q": "港股 港交所 披露易 公告", "lang": "zh", "domain": "hk_filings", "primary": "hkex", "scenario": "finance"},
+    {"id": "R_hot_trending", "q": "微博热搜 今日热搜", "lang": "zh", "domain": "hot_trending", "primary": "baidu_hot", "scenario": "community"},
+    {"id": "R_image_search", "q": "雪景 壁纸 图片", "lang": "zh", "domain": "image_search", "primary": "openverse", "scenario": "media"},
+    {"id": "R_intl_news_flash", "q": "国际新闻 外媒 国际要闻", "lang": "zh", "domain": "intl_news_flash", "primary": "guardian_rss", "scenario": "news"},
+    {"id": "R_japan_law", "q": "日本 法令 民法 条文", "lang": "zh", "domain": "japan_law", "primary": "egov_law", "scenario": "legal"},
+    {"id": "R_kor_law", "q": "韩国 法令 民法 条文", "lang": "zh", "domain": "kor_law", "primary": "kor_law", "scenario": "legal"},
+    {"id": "R_law_text", "q": "法条 法律条文 全文", "lang": "zh", "domain": "law_text", "primary": "flk_law", "scenario": "legal"},
+    {"id": "R_legal", "q": "刑法 司法解释 判例", "lang": "zh", "domain": "legal", "primary": "anysearch", "scenario": "legal"},
+    {"id": "R_meme_slang", "q": "网络梗 流行语 黑话", "lang": "zh", "domain": "meme_slang", "primary": "itotii", "scenario": "media"},
+    {"id": "R_ml_models", "q": "huggingface 预训练模型 开源模型", "lang": "zh", "domain": "ml_models", "primary": "huggingface", "scenario": "tech"},
+    {"id": "R_news_realtime", "q": "最新新闻 今天发生 突发", "lang": "zh", "domain": "news_realtime", "primary": "byted", "scenario": "news"},
+    {"id": "R_outbreak_health", "q": "疫情 暴发 流行病 通报", "lang": "zh", "domain": "outbreak_health", "primary": "who_don", "scenario": "science"},
+    {"id": "R_patent_search", "q": "专利 权利要求 技术方案", "lang": "zh", "domain": "patent_search", "primary": "google_patents", "scenario": "academic"},
+    {"id": "R_prediction_market", "q": "预测市场 事件合约", "lang": "zh", "domain": "prediction_market", "primary": "polymarket", "scenario": "finance"},
+    {"id": "R_protein_search", "q": "蛋白质 基因 氨基酸序列", "lang": "zh", "domain": "protein_search", "primary": "uniprot", "scenario": "academic"},
+    {"id": "R_redskill_search", "q": "小红书 技能 排行榜", "lang": "zh", "domain": "redskill_search", "primary": "redskill", "scenario": "community"},
+    {"id": "R_scholar_search", "q": "google scholar 学术论文 期刊", "lang": "zh", "domain": "scholar_search", "primary": "google_scholar", "scenario": "academic"},
+    {"id": "R_sec_filings", "q": "SEC EDGAR 招股书 财报申报", "lang": "zh", "domain": "sec_filings", "primary": "sec_edgar", "scenario": "finance"},
+    {"id": "R_semantic_discovery", "q": "discover landscape ecosystem emerging", "lang": "en", "domain": "semantic_discovery", "primary": "exa", "scenario": "tech"},
+    {"id": "R_shopping", "q": "购物 评测 推荐 踩坑", "lang": "zh", "domain": "shopping", "primary": "zhihu", "scenario": "media"},
+    {"id": "R_skill_search", "q": "agent skills 市场 目录", "lang": "en", "domain": "skill_search", "primary": "skillsmp", "scenario": "community"},
+    {"id": "R_species_search", "q": "物种 学名 生物分类 拉丁名", "lang": "zh", "domain": "species_search", "primary": "gbif", "scenario": "academic"},
+    {"id": "R_stackoverflow_search", "q": "stackoverflow 编程问题 代码问题", "lang": "zh", "domain": "stackoverflow_search", "primary": "stackoverflow", "scenario": "code"},
+    {"id": "R_standards", "q": "国家标准 国标 标准号", "lang": "zh", "domain": "standards", "primary": "std_samr", "scenario": "academic"},
+    {"id": "R_tech_deep", "q": "deep dive 架构 实践", "lang": "zh", "domain": "tech_deep", "primary": "arxiv", "scenario": "tech"},
+    {"id": "R_us_fundamentals", "q": "营收 净利润 财务数据 苹果", "lang": "zh", "domain": "us_fundamentals", "primary": "sec_companyfacts", "scenario": "finance"},
+    {"id": "R_us_legal", "q": "supreme court 判例 precedent", "lang": "en", "domain": "us_legal", "primary": "federal_register", "scenario": "legal"},
+    {"id": "R_web_archive", "q": "网页快照 历史网页 wayback", "lang": "zh", "domain": "web_archive", "primary": "wikipedia", "scenario": "general"},
+    {"id": "R_web_docs", "q": "mozilla developer fetch 属性", "lang": "en", "domain": "web_docs", "primary": "mdn", "scenario": "code"},
+    {"id": "R_wechat_search", "q": "微信公众号 文章 搜狗微信", "lang": "zh", "domain": "wechat_search", "primary": "wechat_sogou", "scenario": "community"},
+    {"id": "R_world_news", "q": "noticias 国际 新闻", "lang": "zh", "domain": "world_news", "primary": "anysearch", "scenario": "news"},
+    {"id": "R_zhihu_content", "q": "知乎 怎么看待 如何评价", "lang": "zh", "domain": "zhihu_content", "primary": "zhihu", "scenario": "media"},
+    {"id": "R_zhihu_hot_list", "q": "知乎热榜 知乎热搜", "lang": "zh", "domain": "zhihu_hot_list", "primary": "zhihu_hot", "scenario": "community"},
+    {"id": "R_zhihu_user_data", "q": "我的知乎 我的创作 收藏", "lang": "zh", "domain": "zhihu_user_data", "primary": "zhihu_user", "scenario": "community"},
+
+    # ── 域冲突让位修复（2026-10-22）────────────────────────────────────────
+    # 这三个域的触发词此前被更宽泛的域覆盖，无法通过自动路由命中（因此曾登记在
+    # test_route_matrix_coverage 的 _NO_CASE_DEBT 里）。在宽泛侧加 unless 让位后
+    # 各补一条钉住；宽泛侧的真查询锁在 test_domain_pattern_precision 的既有用例。
+    {"id": "R_jin10_flash", "q": "金十 7x24 快讯", "lang": "zh", "domain": "jin10_flash", "primary": "jin10", "scenario": "news"},
+    {"id": "R_v2ex_search", "q": "v2ex 技术社区", "lang": "zh", "domain": "v2ex_search", "primary": "v2ex", "scenario": "code"},
+    {"id": "R_wenshu_query", "q": "裁判文书 判决书", "lang": "zh", "domain": "wenshu_query", "primary": "wenshu", "scenario": "legal"},
 ]
 
 

package/scripts/mcp_compact.py

@@ -0,0 +1,234 @@
+#!/usr/bin/env python3
+"""mcp_compact.py — MCP 结果载荷裁剪与紧凑压缩协议。
+
+拆分自 mcp_handlers.py：纯数据处理与结构级截断，无网络 IO 与外部状态。
+供 MCP 工具执行层（mcp_handlers）、CLI 与外部测试独立复用。
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+
+def _trim_snippet(text: Any, n: int = 120) -> str | None:
+    """截断摘要文本，防止超长字符膨胀。"""
+    if not text:
+        return None
+    s = str(text)
+    return s if len(s) <= n else s[:n]
+
+
+def _compact_search_result(result: dict[str, Any], summary: bool = False) -> dict[str, Any]:
+    """去掉 MCP 热路径用不到的重字段，缩小首包。"""
+    snip_n = 80 if summary else 160
+    out: dict[str, Any] = {
+        "query": result.get("query"),
+        "engine": result.get("engine"),
+        "engines_used": result.get("engines_used") or result.get("engines"),
+        "count": result.get("count"),
+        "elapsed_ms": result.get("elapsed_ms"),
+        "cached": result.get("cached"),
+        "mode": result.get("mode"),
+        "depth": result.get("depth"),
+        "execution_tier": result.get("execution_tier"),
+        "input_kind": result.get("input_kind"),
+        "login_hint": result.get("login_hint"),
+    }
+    results = []
+    for r in result.get("results") or []:
+        if not isinstance(r, dict):
+            continue
+        item = {
+            "title": r.get("title"),
+            "url": r.get("url"),
+            "snippet": _trim_snippet(r.get("snippet"), snip_n),
+            "source": r.get("source"),
+            "score": r.get("score"),
+        }
+        # 保留发布时间维度（仅引擎输出该字段时，避免空字段噪音）
+        if r.get("published_at"):
+            item["published_at"] = r["published_at"]
+        # 顶包标记：recovery 替换了引擎原结果时透传，供上层识别结果已被代答
+        if r.get("_recovered"):
+            item["recovered"] = r["_recovered"]
+        # 保留快评，丢掉大块 meta
+        for k in ("selection", "absorption", "credibility_fast", "evidence_flags"):
+            if k in r:
+                item[k] = r[k]
+        # 证据完整链路 P0：保留核验信号（仅存在时，避免空字段噪音）
+        for k in ("fetch_suggested", "has_fetched_evidence", "post_fetch_absorption"):
+            if r.get(k) is not None:
+                item[k] = r[k]
+        results.append(item)
+    out["results"] = results
+    # 证据完整链路 P0：顶层门控信号（轻量 4 字段，供 Agent 判断「能否下结论」）
+    if result.get("fetch_required") is not None:
+        out["fetch_required"] = result["fetch_required"]
+    el = result.get("evidence_loop")
+    if isinstance(el, dict):
+        out["evidence_loop"] = {
+            "suggested": el.get("suggested") or [],
+            "verified_count": el.get("verified_count") or 0,
+            "pending_count": el.get("pending_count") or 0,
+        }
+    # 引擎可观测性：精简版 engine_outcomes（engine/status/latency，去掉 detail）
+    # 让 Agent 能看到「哪个引擎失败、为什么类别」，不占大体积
+    outcomes = result.get("engine_outcomes") or []
+    if outcomes:
+        out["engine_outcomes"] = [
+            {
+                "engine": o.get("engine"),
+                "status": o.get("status"),
+                "latency_ms": o.get("latency_ms"),
+                "results_count": o.get("results_count"),
+            }
+            for o in outcomes
+            if isinstance(o, dict)
+        ][:12]
+        wasted = result.get("wasted_engine_ms")
+        if wasted:
+            out["wasted_engine_ms"] = wasted
+    if result.get("early_stopped"):
+        out["early_stopped"] = True
+    if result.get("minhash_removed"):
+        out["minhash_removed"] = result.get("minhash_removed")
+    from search_output import decorate_compact_payload as _decorate
+    _decorate(out, result)
+    if result.get("domain_filter"):
+        out["domain_filter"] = result["domain_filter"]
+    if result.get("rank_method"):
+        out["rank_method"] = result.get("rank_method")
+    if result.get("reranker"):
+        out["reranker"] = result.get("reranker")
+    lp = result.get("lang_pref")
+    if isinstance(lp, dict):
+        out["lang_pref"] = {
+            "query_lang": lp.get("query_lang"),
+            "system_lang": lp.get("system_lang"),
+            "habit_lang": lp.get("habit_lang"),
+            "prefer_langs": lp.get("prefer_langs"),
+            "engine_lang": lp.get("engine_lang"),
+        }
+    rec = result.get("recovery")
+    if isinstance(rec, dict) and rec.get("triggered"):
+        out["recovery"] = {
+            "recovered": rec.get("recovered"),
+            "level_used": rec.get("level_used"),
+            "strategy_used": rec.get("strategy_used"),
+            "note": rec.get("note"),
+        }
+    return out
+
+
+def _compact_research_result(report: dict[str, Any], summary: bool = False) -> dict[str, Any]:
+    """研究包：只保留 Agent 写作需要的字段，砍掉子查询原始 dumps。"""
+    snip_n = 100 if summary else 160
+    keys = (
+        "query", "query_original", "execution_tier", "topic_profile", "topic_profile_key",
+        "discipline", "quality_gates", "report_sections", "source_grades", "disclaimer",
+        "academic_discipline", "engines_used", "source_distribution", "elapsed_ms",
+        "sub_query_count", "total_sources", "gaps", "mode", "sub_queries",
+        "engines_priority", "vertical_engines", "fetch_required",
+        "kind", "conclusion_cap", "quality_gate_results", "work_packages",
+        "work_package_stages", "protocol", "cross_lingual_trace",
+    )
+    out: dict[str, Any] = {k: report[k] for k in keys if k in report and report[k] is not None}
+
+    # 证据完整链路 P0：研究包门控（pending_fetch 截断到 5 条，保留可编程核验信号）
+    el = report.get("evidence_loop")
+    if isinstance(el, dict):
+        out["evidence_loop"] = {
+            "high_consequence_domain": el.get("high_consequence_domain"),
+            "pending_fetch": (el.get("pending_fetch") or [])[:5],
+            "verified_count": el.get("verified_count") or 0,
+            "pending_count": el.get("pending_count") or 0,
+        }
+    if isinstance(report.get("verify"), dict):
+        v = report["verify"]
+        out["verify"] = {
+            "revision_summary": v.get("revision_summary") or {},
+            "pending": (v.get("pending") or [])[:5],
+        }
+
+    findings = []
+    for kf in report.get("key_findings") or []:
+        if not isinstance(kf, dict):
+            continue
+        top = dict(kf.get("top_result") or {})
+        if top.get("snippet"):
+            top["snippet"] = _trim_snippet(top.get("snippet"), snip_n)
+        findings.append({
+            "aspect": kf.get("aspect"),
+            "strategy": kf.get("strategy"),
+            "top_result": top,
+            "result_count": kf.get("result_count"),
+            "citation_refs": kf.get("citation_refs"),
+        })
+    out["key_findings"] = findings
+
+    sources = []
+    for s in report.get("sources") or report.get("citations") or []:
+        if not isinstance(s, dict) or not s.get("url"):
+            continue
+        sources.append({
+            "ref": s.get("ref") or s.get("id"),
+            "title": s.get("title"),
+            "url": s.get("url"),
+            "engine": s.get("engine") or s.get("source"),
+            "snippet": _trim_snippet(s.get("snippet"), snip_n),
+        })
+        if summary and len(sources) >= 10:
+            break
+    out["sources"] = sources
+    if not summary:
+        out["citations"] = report.get("citations") or sources
+        # 交叉引用截断
+        crs = report.get("cross_references") or []
+        out["cross_references"] = crs[:5]
+    # 覆盖度 / 验证记录 / 盲区（科研方法论增强）
+    if report.get("coverage_map"):
+        out["coverage_map"] = report["coverage_map"]
+    if report.get("verification_records"):
+        out["verification_records"] = report["verification_records"][:12]
+    if report.get("blind_spots"):
+        out["blind_spots"] = report["blind_spots"][:8]
+    return out
+
+
+def _cap_extract_output(output: dict[str, Any]) -> dict[str, Any]:
+    """argo_fetch mode=extract 的输出上限（防大页表格撑爆上下文）。
+
+    结构级裁剪保持 JSON 可解析：表格数量/单元格、metadata 值、jsonld
+    条目体积分别设上限，超限截断并打标记。
+    """
+    tables = output.get("tables")
+    if isinstance(tables, list):
+        if len(tables) > 12:
+            output["tables"] = tables[:12]
+            output["tables_truncated"] = True
+        for t in output["tables"]:
+            if isinstance(t, list):
+                for i, row in enumerate(t):
+                    if isinstance(row, list):
+                        t[i] = [str(c)[:300] if c is not None else "" for c in row]
+    meta = output.get("metadata")
+    if isinstance(meta, dict):
+        output["metadata"] = {k: str(v)[:500] for k, v in meta.items()}
+    jsonld = output.get("jsonld")
+    if isinstance(jsonld, list):
+        capped = []
+        for entry in jsonld:
+            try:
+                text = json.dumps(entry, ensure_ascii=False)
+            except (TypeError, ValueError):
+                continue
+            if len(text) > 4000:
+                capped.append({"_clipped": True, "preview": text[:4000]})
+            else:
+                capped.append(entry)
+        if len(capped) > 8:
+            capped = capped[:8]
+            output["jsonld_truncated"] = True
+        output["jsonld"] = capped
+    return output