Unfinished Work Audit

Overview

Scan local AI session history across tools, then report what is genuinely still open. The hard part is not finding sessions — it is that a session's last message is a claim about state, often weeks stale. Verify before listing.

Evidence Sources

Read references/scan-recipes.md for the extraction snippets (first user message, last assistant message, per-project grouping, mojibake-safe output).

Workflow

1. Set the window and the roots.
   • Default to 60 days. If that yields under ~20 sessions, widen to all history and say so.
   • Enumerate both Codex and Claude Code. Reporting only one tool's sessions is a wrong answer to "所有对话", and the same project often has threads in both.
2. Shortlist candidate-open threads.
   • Signals: an ending that promises a next step (下一步, 待办, 未完成, TODO, pending, 接下来), an explicit stop (我现在停止, 不再追加功能), a pending_external marker, or an error/blocker as the final state.
   • A thread that ends with a delivered result and no promised follow-up is closed. Do not pad the list.
3. Verify each candidate against the repo — this is the step that makes the report worth anything.
   • Check whether the promised artifact exists now: the file, the export, the test, the commit.
   • A session claiming "还需重新导出 EXE" is not evidence the EXE is still stale; compare the binary's mtime against the last source change.
   • Mark each item 已完成（会话未更新）, 确认未完成, or 无法判定 with the reason.
   • Never infer completion from git status cleanliness, and never infer incompleteness from the session text alone.
4. Separate what is blocked from what is merely unfinished.
   • pending_external (real-device QA, store review, human playtest, another person) is not backlog the user can clear tonight — list it apart.
   • Keep the user's own 已停止开发 decisions out of the backlog; a deliberate stop is not a loose end.
5. Report smallest-useful.
   • Group by project, ordered by how cheap it is to close.
   • Per item: 项目 / 停在哪一步 / 证据（文件路径、时间戳、会话日期）/ 验证结论 / 下一步一句话.
   • State coverage honestly: sessions scanned, tools covered, and anything unreadable.

Common Mistakes

• Trusting the last assistant message as current state. It is the single biggest source of false backlog here.
• Listing every session that mentions a TODO. Most were closed later, often in a different thread or the other tool.
• Printing Chinese thread names straight to a Windows console — they mojibake. Write to a UTF-8 file and read that back.
• Dumping raw transcript excerpts into the report; follow reporting-discipline and give conclusion plus evidence pointer.

Output Contract

Report: window and roots scanned, session counts per tool, verified-open items grouped by project, pending_external listed separately, items reclassified as already-done, and anything that could not be verified.