From 9090283e7027018bb3fb16d60ac2d1e71d966a26 Mon Sep 17 00:00:00 2001
From: Color2333 <1552429809@qq.com>
Date: Sat, 18 Jul 2026 20:24:40 +0800
Subject: [PATCH] =?UTF-8?q?fix(idle):=20=E8=A1=A5=E5=81=BF=E8=A7=A6?=
 =?UTF-8?q?=E5=8F=91=20bug=20+=20=E5=AD=98=E9=87=8F=E5=9B=9E=E5=A1=AB?=
 =?UTF-8?q?=E8=84=9A=E6=9C=AC?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

发现 idle 补偿触发 bug：Critical #6 的 _compensate_stuck_skimmed 此前
挂在 _process_batch 末尾，但无 unread 论文时 _process_batch 提前
return 0（line 225-227），补偿永远不触发。库里 1239 篇卡在 skimmed
未精读（有 summary_md 但 deep_dive_md 空），idle_processor 反复报
"没有需要处理的未读论文"却从不补偿，stuck 论文永久卡死。

修复：
- packages/ai/idle_processor.py：把 _compensate_stuck_skimmed 调用从
  _process_batch 末尾移到 _run_loop，独立于 skim 批次触发。无论有无
  unread 论文，每次 idle 检查都尝试补偿 stuck 论文。_process_batch
  只负责 skim，补偿在循环层独立调用。

回填脚本：
- scripts/redeep_stuck_skimmed.py：一次性回填存量 stuck 论文。按
  skim_score 降序优先精读高价值论文（1233 篇高分>=0.8），支持
  --dry-run/--limit/--min-score/--concurrency。deep_dive 幂等覆盖，
  重跑安全。

测试：新增 TestIdleCompensationTrigger 2 项（_process_batch 无 unread
返回0且不调补偿、补偿独立捞 stuck 论文），全套 63 passed。
---
 packages/ai/idle_processor.py   |   9 +-
 scripts/redeep_stuck_skimmed.py | 155 ++++++++++++++++++++++++++++++++
 tests/test_repositories.py      |  49 ++++++++++
 3 files changed, 210 insertions(+), 3 deletions(-)
 create mode 100644 scripts/redeep_stuck_skimmed.py

diff --git a/packages/ai/idle_processor.py b/packages/ai/idle_processor.py
index 6355165..7bb94f1 100644
--- a/packages/ai/idle_processor.py
+++ b/packages/ai/idle_processor.py
@@ -318,9 +318,7 @@ def _process_batch(self) -> int:
         self._papers_processed += processed
         self.detector.mark_task_executed()
 
-        # Critical #6 补偿：对已 skim 但卡住未精读的论文补一次精读（独立配额受限）
-        deep_compensated = self._compensate_stuck_skimmed()
-        return processed + deep_compensated
+        return processed
 
     def _compensate_stuck_skimmed(self) -> int:
         """补偿已 skim 但未精读的论文（Critical #6）。
@@ -380,6 +378,11 @@ def _run_loop(self):
                     if not self._is_processing:
                         self._is_processing = True
                         self._process_batch()
+                        # Critical #6 补偿：独立于 skim 批次触发。此前补偿挂在
+                        # _process_batch 末尾，但无 unread 论文时它提前 return 0，
+                        # 补偿永远不跑 → 1239 篇卡在 skimmed 未精读。改为在 _run_loop
+                        # 独立调用，无论有无 unread 都尝试补偿 stuck 论文。
+                        self._compensate_stuck_skimmed()
                         self._is_processing = False
                 else:
                     if self._is_processing:
diff --git a/scripts/redeep_stuck_skimmed.py b/scripts/redeep_stuck_skimmed.py
new file mode 100644
index 0000000..3ef0f4c
--- /dev/null
+++ b/scripts/redeep_stuck_skimmed.py
@@ -0,0 +1,155 @@
+#!/usr/bin/env python3
+"""
+补偿精读卡在 skimmed 的存量论文 - 一次性回填脚本。
+
+背景：Critical #6 修复让 idle_processor 补偿 skimmed 未精读的论文，但补偿逻辑
+此前挂在 _process_batch 末尾，无 unread 论文时提前 return 导致永远不触发。
+库里积累 1239 篇卡在 skimmed（有 summary_md 但 deep_dive_md 空），其中 1233 篇
+高分(>=0.8)。本脚本一次性回填这些存量论文，按 skim_score 降序优先精读高价值论文。
+
+修复 idle 补偿 bug 后，新抓取的 stuck 论文由 idle_processor 自动补偿；本脚本只
+处理修复前已卡住的存量。deep_dive 是幂等覆盖（upsert_deep_dive 覆写 deep_dive_md），
+直接重跑安全。
+
+@author Color2333
+
+使用方式:
+    # 预览要精读的论文数量（不调用 API）
+    python scripts/redeep_stuck_skimmed.py --dry-run
+
+    # 回填全部 stuck 论文（按分数降序）
+    python scripts/redeep_stuck_skimmed.py
+
+    # 限制数量（如先跑 100 篇）
+    python scripts/redeep_stuck_skimmed.py --limit 100
+
+    # 只回填高分(>=0.8)的
+    python scripts/redeep_stuck_skimmed.py --min-score 0.8
+"""
+
+from __future__ import annotations
+
+import argparse
+import logging
+import pathlib
+import sys
+from concurrent.futures import ThreadPoolExecutor, as_completed
+
+# 让脚本能直接 `python scripts/x.py` 运行（不依赖 PYTHONPATH 环境变量）
+sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent.parent))
+
+from sqlalchemy import select
+
+from packages.ai.pipelines import PaperPipelines
+from packages.ai.rate_limiter import acquire_api, get_rate_limiter
+from packages.storage.db import session_scope
+from packages.storage.models import AnalysisReport, Paper
+
+# 配置日志
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s [%(levelname)s] %(message)s",
+    datefmt="%Y-%m-%d %H:%M:%S",
+)
+logger = logging.getLogger(__name__)
+
+
+def get_stuck_skimmed_papers(
+    limit: int = 0, min_score: float = 0.0
+) -> list[tuple[str, str, float]]:
+    """获取卡在 skimmed 未精读的论文（有 summary_md 但 deep_dive_md 空）。
+
+    按 skim_score 降序，优先精读高价值论文。
+    Returns: [(paper_id, title, skim_score), ...]
+    """
+    with session_scope() as session:
+        stmt = (
+            select(Paper.id, Paper.title, AnalysisReport.skim_score)
+            .join(AnalysisReport, AnalysisReport.paper_id == Paper.id)
+            .where(Paper.read_status == "skimmed")
+            .where(AnalysisReport.summary_md.is_not(None))
+            .where(AnalysisReport.deep_dive_md.is_(None))
+            .where(AnalysisReport.skim_score >= min_score)
+            .order_by(AnalysisReport.skim_score.desc())
+        )
+        if limit > 0:
+            stmt = stmt.limit(limit)
+        rows = session.execute(stmt).all()
+        return [(str(r.id), r.title, r.skim_score or 0.0) for r in rows]
+
+
+def redeep_single(paper_id: str, title: str, score: float) -> dict:
+    """精读单篇论文。"""
+    pipelines = PaperPipelines()
+    limiter = get_rate_limiter()
+    result = {
+        "paper_id": paper_id[:8],
+        "title": title[:50],
+        "skim_score": score,
+        "deep_success": False,
+        "error": None,
+    }
+
+    if not limiter.start_task():
+        result["error"] = "并发满"
+        return result
+    try:
+        if not acquire_api("llm", timeout=30.0):
+            result["error"] = "LLM API 限流"
+            logger.warning(f"⚠️  [{paper_id[:8]}] LLM 限流，跳过")
+            return result
+        logger.info(f"📖 [{paper_id[:8]}] 开始精读 (skim={score:.2f}) {title[:40]}")
+        pipelines.deep_dive(paper_id)
+        result["deep_success"] = True
+        logger.info(f"✅ [{paper_id[:8]}] 精读完成 (skim={score:.2f})")
+    except Exception as exc:
+        result["error"] = str(exc)[:120]
+        logger.warning(f"❌ [{paper_id[:8]}] 精读失败: {exc}")
+    finally:
+        limiter.end_task()
+    return result
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="回填卡在 skimmed 未精读的存量论文")
+    parser.add_argument("--dry-run", action="store_true", help="只统计不精读")
+    parser.add_argument("--limit", type=int, default=0, help="限制处理数量（0=全部）")
+    parser.add_argument("--min-score", type=float, default=0.0, help="只精读 skim_score >= 此值的")
+    parser.add_argument("--concurrency", type=int, default=2, help="并发数")
+    args = parser.parse_args()
+
+    papers = get_stuck_skimmed_papers(limit=args.limit, min_score=args.min_score)
+    logger.info(f"找到 {len(papers)} 篇卡在 skimmed 未精读的论文 (min_score={args.min_score})")
+
+    if not papers:
+        logger.info("无需回填")
+        return
+
+    # 分数分布预览
+    high = sum(1 for _, _, s in papers if s >= 0.8)
+    mid = sum(1 for _, _, s in papers if 0.65 <= s < 0.8)
+    low = sum(1 for _, _, s in papers if s < 0.65)
+    logger.info(f"分数分布: 高分(>=0.8)={high}, 中(0.65-0.8)={mid}, 低(<0.65)={low}")
+
+    if args.dry_run:
+        logger.info("dry-run 模式：不精读，退出")
+        return
+
+    success = 0
+    failed = 0
+    with ThreadPoolExecutor(max_workers=args.concurrency) as pool:
+        futures = {
+            pool.submit(redeep_single, pid, title, score): pid for pid, title, score in papers
+        }
+        for fut in as_completed(futures):
+            r = fut.result()
+            if r["deep_success"]:
+                success += 1
+            else:
+                failed += 1
+
+    logger.info(f"回填完成：成功={success}, 失败={failed}, 总计={len(papers)}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/tests/test_repositories.py b/tests/test_repositories.py
index 8c9bc6e..c7eadd6 100644
--- a/tests/test_repositories.py
+++ b/tests/test_repositories.py
@@ -493,3 +493,52 @@ def test_last_error_persisted_and_readable(self, db_session):
         repo.update_run_status(topic.id, error=None)
         db_session.refresh(topic)
         assert topic.last_error is None
+
+
+class TestIdleCompensationTrigger:
+    """Critical #6 补偿触发 bug 回归：补偿须独立于 skim 批次，无 unread 时也跑
+
+    此前 _compensate_stuck_skimmed 挂在 _process_batch 末尾，但无 unread 论文时
+    _process_batch 提前 return 0，补偿永远不触发 → stuck 论文卡死。
+    """
+
+    def test_process_batch_returns_zero_when_no_unread(self, db_session, monkeypatch):
+        """无 unread 论文时 _process_batch 返回 0（主路径提前返回）"""
+        from packages.ai.idle_processor import IdleProcessor
+
+        ip = IdleProcessor()
+        # 无论文时 _get_unread_papers 返回空 → _process_batch 直接 return 0
+        monkeypatch.setattr(ip, "_get_unread_papers", lambda limit=10: [])
+        # 不应触发补偿（补偿已移出 _process_batch）
+        called = []
+        monkeypatch.setattr(ip, "_compensate_stuck_skimmed", lambda: called.append(1) or 0)
+        result = ip._process_batch()
+        assert result == 0
+        assert called == [], "_process_batch 不应再调用补偿（已移到 _run_loop）"
+
+    def test_compensate_runs_independently_of_unread(self, db_session, monkeypatch):
+        """补偿独立触发：无 unread 但有 stuck skimmed 时仍补偿精读"""
+        from uuid import uuid4
+
+        from packages.domain.enums import ReadStatus
+        from packages.storage.models import AnalysisReport
+
+        repo = PaperRepository(db_session)
+        paper = repo.upsert_paper(
+            PaperCreate(arxiv_id="2401.00301", title="stuck", abstract="a", metadata={})
+        )
+        repo.update_read_status(paper.id, ReadStatus.skimmed)
+        db_session.add(
+            AnalysisReport(id=str(uuid4()), paper_id=paper.id, summary_md="skim", deep_dive_md=None)
+        )
+        db_session.commit()
+
+        from packages.ai.idle_processor import IdleProcessor
+
+        ip = IdleProcessor()
+        # 无 unread 论文
+        monkeypatch.setattr(ip, "_get_unread_papers", lambda limit=10: [])
+        # 补偿应能捞到这篇 stuck 论文（验证查询独立可用）
+        stuck = ip._get_stuck_skimmed_papers(limit=5)
+        assert len(stuck) == 1, "应捞到 1 篇 stuck skimmed 论文"
+        assert stuck[0][0] == paper.id