feat: 大文本文件 lazy loading + 24h 文件缓存

[lishuceo]

Summary

• 大文本文件不再塞 prompt：>64KB 的文本附件落盘到 tmpdir/feishu-downloads/，prompt 只注入路径元数据，agent 用 Read 工具按需 offset/limit 分段读
• 文本文件上限从 1MB → 30MB（与 PDF 一致），群里发 3.5MB 日志不再被拒
• 抽取 src/feishu/file-cache.ts 共享 helper，feishu_download_message_file MCP 工具也复用
• 周期 cleanup 接入 24h 过期清理（硬盘充裕，给 agent 留足回头读的窗口）

Test plan

• src/__tests__/lazy-text-files.test.ts — file-cache 单测 + 历史文本附件阈值覆盖
• 既有 lazy-history-files.test.ts / message.test.ts 全部通过
• 全量 vitest：1427 passed，仅 memory 质量集成测试失败（需 DASHSCOPE_API_KEY，与本 PR 无关）

🤖 Generated with Claude Code

[github-actions]

Claude finished @lishuceo's task in 2m 12s —— View job

---

⚠️ Review: Issues Found (1 minor)

• Checked previous claude review comments (none)
• Read CLAUDE.md and PR diff
• Reviewed src/feishu/file-cache.ts
• Reviewed src/feishu/event-handler.ts changes
• Reviewed src/feishu/tools/message.ts changes
• Reviewed src/index.ts cleanup wiring
• Reviewed test coverage
• Posted inline + summary

总评

整体质量不错：抽取共享 file-cache.ts 消除重复，路径 sanitize（SAFE_NAME_RE + 扩展名白名单）、cleanup 容错（ENOENT/stat/unlink 都吞掉只记 debug）、24h TTL 配合每 30 分钟周期清理这套组合都做得到位。测试覆盖了 happy path、过期文件清理、目录缺失、64KB inline 阈值与 30MB 硬上限。

Issues

🟡 引用回复（quoted parent message）路径未同步升级 — confidence 80

• src/feishu/event-handler.ts:3507-3517：parseMessage 中处理 message.parent_id 的分支仍保留 MAX_TEXT_SIZE = 1 * 1024 * 1024 且只 inline embed，没用新的 saveMessageFileToCache。
• 用户先发 3.5MB 日志、再引用回复 @bot 这种 case 仍会被 silent skip，与 PR 描述的"群里发 3.5MB 日志不再被拒"语义不一致。
• 见上方 inline comment。

Nits（不阻塞）

• cleanupOldDownloads is a no-op when dir does not exist 这个测试用例（src/__tests__/lazy-text-files.test.ts:77-83）实际上 mkdtemp 出来的临时目录跟 DOWNLOAD_DIR 没关系，断言只验证了"返回 number"，并不能真正覆盖 ENOENT 分支——上一条用例可能已把文件写进 DOWNLOAD_DIR，所以测的根本不是空目录。可以在这个 case 里把 DOWNLOAD_DIR 先 rmSync 掉再调用，覆盖才有意义。
• file-cache.ts:24 的 safeExt 正则 /^\.[a-z0-9]{1,8}$/：上限 8 字符已经覆盖 .config、.tsx 等常见情况，没问题；如果以后要支持 .markdown（8 字符）边界刚好，更长的就会被裁掉，对当前需求够用。

验证项

• ✅ ESM .js import 后缀齐全
• ✅ cleanupOldDownloads 错误路径完整（readdir/stat/unlink 都有兜底）
• ✅ Promise.all 内 cleaned += 1 在单线程 Node 中安全
• ✅ 路径注入 prompt 时 messageId + fileKey 已 sanitize，无逃逸风险
• ✅ 测试新增覆盖 PR 行为，且兼容现有测试套件
  · branch feat/lazy-text-file-loading

[github-actions]

Reviewing PR updates

• Check previous claude review comments & status
• Read current state of src/feishu/event-handler.ts (quoted reply path)
• Review latest diff vs origin/main
• Verify tests cover the fix (if any)
• Resolve fixed inline threads & post summary

View job run · branch feat/lazy-text-file-loading

[github-actions]

Claude finished @lishuceo's task in 1m 9s —— View job

---

✅ Approved — Previous issue resolved

• Checked previous review threads
• Verified src/feishu/event-handler.ts quoted-reply path (L3507-3523)
• Reviewed latest diff vs origin/main
• Reviewed tests
• Posted summary

Verification

Previously flagged in inline #3253872451: the quoted-parent-message text-file branch in parseMessage still used the legacy 1MB silent-skip path.

Fix in 39db424 correctly addresses this. Current event-handler.ts:3507-3523:

} else if (isTextFile(fileName)) {
  const MAX_TEXT_SIZE = 30 * 1024 * 1024;
  const INLINE_THRESHOLD = 64 * 1024;
  const buf = await feishuClient.downloadMessageFile(parent.message_id, fileKey);
  if (buf.length > MAX_TEXT_SIZE) { … warn … }
  else if (buf.length <= INLINE_THRESHOLD) { … inline embed … }
  else { … saveMessageFileToCache + lazy-load metadata … }
}

Three text-file entries (direct upload parseMessage, history downloadHistoryFiles, quoted-reply parent) now share the same 30MB cap + 64KB inline threshold + saveMessageFileToCache lazy-load pattern. ✅

The lint fix in fbfe024 is trivial and correct (unused writeFileSync, redundant / escape in character class).

Summary

• 1 previous issue fixed (quoted-reply lazy load)
• No new issues found
• The earlier non-blocking nit about cleanupOldDownloads is a no-op when dir does not exist (test L77-83 doesn't actually force the dir to be missing) still stands but was already flagged as non-blocking — leaving as-is.
  · branch feat/lazy-text-file-loading