Resume Handoff

Overview

When work spans more of a session than can safely finish, keep a handoff document that lets a fresh agent (with zero memory of this conversation) get oriented in minutes and continue from exactly where you stopped.

Two core principles:

1. Write for a stranger. The next agent knows nothing. If it isn't in the file (or linked from it), it doesn't exist.
2. Checkpoint as you go — never gamble on the last moment. See below.

Why checkpoint, not wait (the honest limitation)

You cannot reliably know when the session is about to end:

• Quota / rate limit: you receive no signal at all. Only the user sees remaining usage in the UI. You cannot self-detect "about to run out."
• Context window: semi-observable — the harness auto-compacts when it fills, but you have no exact token count.
• Hard cutoff mid-turn: you get no chance to act — no last-second save is possible.

So a design that waits until "I judge I'm about to run out" is unreliable by construction. The fix is to make HANDOFF.md a living document, refreshed at milestones, so that whenever the session ends, a recent checkpoint (at most one step stale) already exists on disk. Handoff becomes a heartbeat, not a deathbed event.

Refresh the handoff when:

• A sub-task completes, or you're about to start a long/risky step (write the checkpoint before it, not after).
• The user says 交接 / 续命 / handoff / 生成交接文档 / 快没额度了.
• You get a PreCompact / just-compacted reminder (context is filling — save now).
• You've done substantial work since the last refresh and none is captured yet.

Re-running is cheap and idempotent — it just overwrites HANDOFF.md. When in doubt, refresh. Do not use for finished work (just report results) or trivial one-step tasks.

Procedure

Run this the first time you checkpoint, then re-run the relevant steps on each refresh (it overwrites HANDOFF.md in place).

1. Gather state — objective, what's done, what's left, key files, decisions, gotchas, next steps. Reconstruct from the actual conversation, not a guess.
2. Fill the template — use template.md. Keep every section; write "N/A" if truly empty rather than deleting it.
3. Reference, don't duplicate — link plans/PRDs/ADRs/diffs/issues/memory files by path or URL instead of pasting them.
4. Redact secrets — no API keys, passwords, tokens, or PII in the file.
5. Save to two places — run the helper so the doc lands both where the next agent will look and in a timestamped archive:

   # run save-handoff.sh from this skill's own directory
   # (the skill loader prints its base dir; default install: ~/.claude/skills/resume-handoff/)
   ./save-handoff.sh /path/to/drafted-handoff.md

   It writes HANDOFF.md at the project root and archives a timestamped copy to your agent's handoffs dir (~/.claude/handoffs/ on Claude Code, ~/.codex/handoffs/ on Codex; auto-detected). It refuses to clutter $HOME: if the working dir is your home directory, it archives only and tells you.
6. Tell the user where it saved and the one-line command the next agent should start with (see the template's first section).

What makes a handoff the next agent can actually use

Common mistakes

• Writing to OS temp — gets wiped; the next agent can't find it. Save to the project + archive.
• Vague next steps ("continue the refactor") — give the actual next file and change.
• Summarizing the vibe instead of the state — paths, commands, and decisions beat prose.
• Deleting empty sections — the next agent won't know if "no blockers" means none or forgotten. Write "N/A".
• Waiting until it's too late — offer the handoff while budget remains, not at 0.

Platform notes

Runs on any skill-aware agent — the SKILL.md format is shared. Only a few bits are runtime-specific:

• Install path — Claude Code: ~/.claude/skills/resume-handoff/; Codex / cross-runtime: ~/.codex/skills/ or ~/.agents/skills/.
• Handoffs archive — save-handoff.sh auto-picks ~/.claude/handoffs/ or ~/.codex/handoffs/ (override with RESUME_HANDOFF_ARCHIVE_DIR).
• Auto-nudge before compaction — the optional PreCompact hook (precompact-hook.sh) is Claude Code-specific. On Codex or elsewhere, rely on the manual + milestone triggers (the primary path anyway).
• Memory entry points — point the next agent at your runtime's memory: Claude Code auto-memory / mem-search; Codex AGENTS.md.