agent-handoff

Three user-invoked skills for coding agents. They share one file format and one location, so a handoff saved in one agent resumes in any of them. Packaged for Claude Code, Codex and Antigravity (agy) from the same skills/ directory.

• handoff-save: you are about to shut down. The agent stops starting new work, reaches a safe stopping point, commits finished work (when its instructions allow commits), and writes a handoff file a fresh session can resume from with no memory of this one.
• handoff-resume: after a restart, finds the newest handoff for the current directory, checks it against git, briefs you, and continues from its next step.
• afk: you are stepping away. The agent works without asking for input, stays local and reversible (no push, deploy, installs, or network sends), logs every judgment call, and leads its final message with the decisions for you to review.

Files

What	Where
Handoffs	~/.local/state/agent-handoffs/YYYY-MM-DD-HHMM-<project-dir>.md
Resumed handoffs	~/.local/state/agent-handoffs/done/
AFK logs	~/.local/state/agent-afk-logs/YYYY-MM-DD-HHMM-afk.md

Both directories are outside any workspace, so an agent with sandboxed or rule-based file access needs write access to them (for example an Antigravity write_file(<absolute path>) allow rule).

Install

Claude Code:

claude plugin marketplace add tjs0/agent-handoff
claude plugin install agent-handoff@agent-handoff

Codex (plugin skills load in the interactive TUI, not in codex exec):

codex plugin marketplace add tjs0/agent-handoff
codex plugin add agent-handoff@agent-handoff

Antigravity:

agy plugin install https://github.com/tjs0/agent-handoff

Or symlink each skills/<name> directory into the agent's personal skills directory (~/.claude/skills/, ~/.codex/skills/, ~/.gemini/config/skills/).

Use

Each skill takes an optional argument, such as a time limit for handoff-save or a task to focus on for afk.

• Claude Code: /agent-handoff:handoff-save, /agent-handoff:handoff-resume, /agent-handoff:afk.
• Codex and Antigravity: ask for the skill by name ("use the afk skill").

The skills never trigger on their own. disable-model-invocation: true hides them from the model in Claude Code and Antigravity; Codex shows them to the model, so each description says to use it only when the user invokes it.

The handoff records which agent wrote it and, when known, the command to resume that session (claude --resume <id>, codex resume <id>, agy --conversation <id>).

Develop

tests/check.sh checks that the manifests agree on name and version, the skill frontmatter, and runs claude plugin validate and agy plugin validate when those agents are installed. Bump the version in every manifest together.

Keep skill text agent-neutral: no tool names or paths that only one agent has.

License

MIT