Compact the current conversation into a handoff document for another agent to pick up. Save to a user-configured location (OS temp, home folder, or per-project
复制下面这句话,粘贴给 Claude Code、Codex、Cursor 等 AI 编程工具,它会读取安装说明并在你确认后完成安装。
请阅读 https://ai.atlankj.com/install/asset/gh-handoff-7fe330e9ebac ,按照其中的说明把「handoff」安装到你(当前 AI 工具)中。执行前先告诉我将运行的命令和写入的位置,等我确认。
查看 AI 将读取的安装说明正在读取 GitHub 原文…
内容来自 GitHub 原始文件,由原作者维护。在 GitHub 查看
Write a handoff document summarising the current conversation so a fresh agent can continue the work. Save to the temporary directory of the user's OS — not the current workspace.
Include a "suggested skills" section in the document, which suggests skills that the agent should invoke.
Do not duplicate content already captured in other artifacts (PRDs, plans, ADRs, issues, commits, diffs). Reference them by path or URL instead.
Redact any sensitive information, such as API keys, passwords, or personally identifiable information.
If the user passed arguments, treat them as a description of what the next session will focus on and tailor the doc accordingly.
Explicit phrases (any of):
Implicit signals (no phrase, but the intent is unmistakable):
When you detect an implicit trigger, propose the handoff before running it: "Want me to write a handoff for the next session?" — never run it silently.
On first invocation, the skill asks where to save handoffs so the project folder never gets cluttered. Setup is offered once via "Run setup now? (Y/n)" — answering N uses OS-temp defaults for this run and never re-prompts. The user can rerun setup any time via /cs:handoff-setup.
See references/configuration.md for the full config field reference.
The save location is read from the user's config (~/.config/handoff/config.json, or the project-local .handoff/config.json if present). When no config exists and the user declined setup, fall back to:
mktemp -t handoff-XXXXXX.md
Read the file before you write to it.
The handoff doc has five sections. Use these exact headers:
See references/handoff_structure.md for a worked example.
Filling in the five sections is the agent's job, not the script's. Follow references/handoff_prompt.md as a mandatory checklist:
For each topic discussed in the conversation, decide explicitly: include in State of play / log as an Open decision / drop with reason.
Free-handing the summary leads to rosy progress reports and dropped blockers. The checklist prevents that.
Matt's no-duplication discipline made concrete:
See references/deduplication_discipline.md for the full list.
Before saving, the linter scans the draft for secrets and PII. In strict mode (default) it blocks save on findings; in warn mode it flags inline and saves anyway.
Redact:
-----BEGIN ... PRIVATE KEY----- blocks.env-style KEY=value lines containing secretsSee references/redaction_checklist.md for the full pattern list and manual-review steps for what regex cannot catch.
When the plugin is installed, a SessionStart hook scans the configured save location for the most recent handoff (within the retention window) and surfaces it to the new session as <handoff_from_previous_session> data. The next agent reads it as context, not as instructions — suggested actions must be verified against current state before executing.
Disable per-session with HANDOFF_SESSIONSTART=0.
A paired SessionEnd hook checks whether a recent handoff exists when the session is ending. If none does (or the most recent is older than 30 minutes), it prints a one-line reminder so the user is prompted to write one before context is lost.
The hook cannot prompt interactively or block session end — it surfaces text in the session log.
Disable per-session with HANDOFF_SESSIONEND=0.
When work continues past the original handoff time, refresh in place instead of creating a new file:
python3 scripts/handoff_template_generator.py --refresh --goal "<updated goal>"
Prints the path of the most recent handoff. The agent edits it directly. Keeps the save location uncluttered and ensures the SessionStart hook always loads the up-to-date version.
| Tool | Purpose |
|---|---|
setup.py | First-run Q&A — save location, retention, redaction strictness, git-context, recommender scope. |
handoff_template_generator.py | Writes the 5-section scaffold at the configured path. --refresh reuses the most recent handoff instead of creating a new file. |
redaction_linter.py | Scans the draft for secrets/PII before save. Exit 1 on findings in strict mode. |
handoff_self_check.py | Fidelity check — flags empty Goal, State bullets without artifacts, missing Decisions when git is dirty, too few/many Skills, inline content in Artifacts. Run before the linter. |
skill_recommender.py | Suggests 3-5 skills for the next session based on goal text + repo scan. |
cleanup.py | Deletes scaffolds older than the retention window. mtime-guarded — never deletes a handoff the user edited. |
config_loader.py | Shared helper: read project config → global config → defaults. |
/cs:handoff [optional next-session description] — generate the handoff./cs:handoff-setup — reconfigure save location, retention, redaction.cs-handoff-author — Matt-voice persona orchestrating the skill. Terse, no-duplication, references-not-copies.
User: /cs:handoff "finish wiring the redaction linter and open a draft PR"
The skill walks the mandatory checklist, generates a 5-section scaffold, fills it from the conversation, runs the redaction linter, and saves to the configured location. See assets/example_handoff.md for a complete worked example.
User: I'm packing up for the day, let me come back to this tomorrow.
Detect the implicit signal. Propose before running: "Want me to write a handoff for the next session?" — never silently. On confirmation, proceed as Example 1 with an inferred goal.
User: /cs:handoff "ship the migration"
Skill: Run setup now? (Y/n)
User: Y
[setup walks 5 questions: save location, retention, redaction strictness, git context, recommender scope]
Skill: Config saved to ~/.config/handoff/config.json. Continuing with handoff for: ship the migration.
If the user answers N, the skill writes a sentinel and uses defaults (OS temp dir, 7-day retention, strict redaction). The prompt never re-appears.
On the next session, the SessionStart hook scans the configured save location, finds the most recent handoff, and surfaces it as <handoff_from_previous_session> data. The next agent reads it as context, not instructions.
| Step | Command |
|---|---|
| First-run setup | /cs:handoff-setup (or answer Y on first /cs:handoff) |
| Generate a handoff | /cs:handoff [goal] |
| Reconfigure later | /cs:handoff-setup --reconfigure |
| Project-specific config | /cs:handoff-setup --project |
| Disable SessionStart hook | HANDOFF_SESSIONSTART=0 (per session) |
Version: 1.0.0 Inspired by: Matt Pocock's handoff (MIT).