fix: 父群图片 lazy loading + history-fork 实现合并

[lishuceo]

变更概述

• fix: 父群聊图片改为 lazy loading（与父群文件对齐），避免话题外简历干扰土豆儿的简历分析
• test: 新增 lazy-history-images.test.ts 8 个回归用例覆盖 parentMsgCount 边界、下载失败、多图等场景
• refactor: 合并 buildChatHistoryContext 与 buildDirectTaskHistory 95% 重复实现，抽出统一 buildHistoryContext，净减少 44 行

背景

土豆儿在话题中分析简历时，会把群聊历史里其他人发的简历图片一起下载分析，导致候选人信息混淆。原因是 downloadHistoryImages 没有 parentMsgCount 保护，而 downloadHistoryFiles 已经有了。

双层防御

• 代码层：父群图片转为元数据提示（包含 message_id / image_key），由 LLM 按需调用 feishu_download_message_image 加载
• Prompt 层：config/knowledge/recruitment.md 追加"简历分析作用域规则"，明确默认只分析话题内简历（不在本 PR，配置文件 deployment-specific）

测试计划

• 5 个相关 test file / 63 个测试全部通过
• typecheck + build 通过
• 已在部署目录重启 pm2 验证
• 待真实场景回归：群聊里多份简历 + 话题中追问单份的场景

🤖 Generated with Claude Code

[github-actions]

Claude finished @lishuceo's task in 3m 58s —— View job

---

✅ Approved — No critical issues found

Reviewed the full diff against origin/main (10 commits, 11 files, +1218/-231).

Verdict

• 0 previous claude[bot] comments to reconcile.
• 0 new inline issues at the ≥75 confidence threshold.
• Three logical changes are well-scoped, well-tested, and the refactor is faithful.

What was checked

1. downloadHistoryImages lazy-loading (src/feishu/event-handler.ts:1367-1429)

• parentMsgCount boundary indexing is correct: i < parentMsgCount flags messages 0..N-1 as parent.
• Fallback path (downloadable.length === 0 → return { images: [], lazyHints }) returns the hints; they get merged into fileTexts upstream so they reach the prompt.
• Lazy-hint string includes the exact MCP tool name + both required args — LLM has everything to call feishu_download_message_image.
• Test coverage in lazy-history-images.test.ts covers parent-only, thread-only, mixed, multi-imageRefs per message, download failure, and metadata-format scenarios.

2. buildHistoryContext merge (event-handler.ts:1561-1683)

• Verified the merged function preserves both behaviors:
  • parentMsgCount recompute after afterMsgId slice: Math.max(0, parentMsgCount - (idx + 1)) matches the old direct-task version exactly.
  • verboseLogging gate keeps the pipeline diagnostic logs only on the direct path (was always-on for buildDirectTaskHistory, now flag-gated — same effective behavior).
  • errorLabel parameter preserves the differentiated error log message.
• The two public entry points (buildChatHistoryContext, buildDirectTaskHistory) are now thin wrappers — call sites unchanged.

3. resolveMentionGate extraction (event-handler.ts:464-543)

• Logic equivalence with the old inline gate verified branch-by-branch:
  • p2p → 'p2p' (was: pass-through)
  • multi-bot + thread bypass + relevant → 'thread_bypass' (was: threadBypass = true)
  • multi-bot + no bypass → getRespondReason (was: shouldRespond)
  • single-bot + group + no @ + thread session creator → semantic check (preserved)
• The new humanSenders.size <= 1 "exclusive bypass" is genuinely new behavior (skips Qwen call when only creator+bot in thread) — covered by mention-gate.test.ts.
• mention-gate.test.ts is thorough (438 lines, covering p2p, single-bot, multi-bot, thread bypass, edge cases, and the exact bug-reproduction scenario from the PR).

4. Queue accountId context isolation (queue.ts, processQueue in event-handler.ts:629-660)

• The fix is correct: wrapping execute() in runWithAccountId(taskAccountId, ...) overrides the inherited AsyncLocalStorage context that .finally() would otherwise propagate from the previous task.
• queue-context.test.ts includes a clean reproduction of the bug AND the fix using a synthetic AsyncLocalStorage, which is excellent regression coverage.

5. getRespondReason refactor (router.ts)

• Backward-compat wrapper shouldRespond correctly returns reason !== undefined.
• Discriminated reason return is good ergonomics for the gate's logging.

6. checkThreadRelevance context-aware variant (thread-relevance.ts)

• Two-prompt design (with/without context) is sound. The 2s → 3s timeout bump is reasonable given the larger payload.
• Minor downstream observation (NOT blocking): when resolveUserNames fails for multiple senders in formatThreadContext, they all collapse to literal "用户", making the conversation log ambiguous. However Qwen's prompt defaults to false on uncertainty, so this is fail-safe and not a regression vs. the previous no-context call.

7. message-builder.ts hostname tag

• Adds ${hostname()}:${process.pid} to every result card. Useful for debugging multi-instance PM2 deployments. os.hostname() is sync and effectively constant per process — no perf concern.

Tests

• 3 new test files (lazy-history-images, mention-gate, queue-context) plus extensions to agent-router and queue — comprehensive and aligned with the project testing policy in CLAUDE.md.
  · Branch: fix/mention-gate-and-queue-context