Harness handoff
Skill SedyBenoitPeace/agent-harness-kit/skills/harness-handoff
Use when ending a work session, switching to another AI agent (Codex, Cursor, Gemini, another Claude), or asked to "hand off", "prepare the repo for the next agent", or "end the session properly" on a repo using the long-running-agent harness — verifies the session-end ritual (clean tree, green gate, logged session) via a deterministic script and prints a paste-ready prompt telling the next agent exactly which one feature to build.From its SKILL.md
npx -y skills add SedyBenoitPeace/agent-harness-kit --skill harness-handoffAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 3 stars3 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
2.4 KB, as published. Nobody here has run it
Harness Handoff
Session-end ritual check + prompt generator for the next agent, whatever
vendor it is. The mechanical checks and the prompt live in
scripts/handoff.sh — run it, never re-derive either by hand.
Announce at start: "Using harness-handoff to verify the session-end ritual and generate the next agent's prompt."
Workflow
- Run
bash scripts/handoff.sh(path relative to this skill) from the target repo root.- Exit 3 → harness not initialized: offer the harness-setup skill, STOP.
- Exit 2 → a harness file is broken: relay the message, suggest the harness-audit skill, STOP.
- Exit 1 → BLOCKED: relay each BLOCKED line and help the human finish the ritual (commit the tree, fix the gate, log the session), then re-run. Never hand off around a blocker.
- Exit 0 → relay the prompt verbatim in a fenced code block so the human can copy-paste it into the next agent. Do not embellish it.
- Tell them how to feed it: paste into any agent chat, or from a
terminal
codex "<prompt>"/codex exec "<prompt>"— Codex and most CLIs also read AGENTS.md natively, so the prompt is sufficient. - Offer
--skip-gateonly when the human explicitly accepts an unverified baseline (it prints a WARNING; the next agent must run the gate first). - Handoff is read-only. Fixing blockers is the human's call — offer, don't act unilaterally.
Red flags
| Thought | Reality |
|---|---|
| "I'll write a nicer prompt myself" | The script's prompt is the contract. Relay it verbatim. |
| "Tree is dirty but it's just docs" | Blocked means blocked. Commit first. |
| "I'll quietly commit their changes to unblock" | Surface the blocker; the human decides. |
| "The next agent is Codex, let me add vendor tips to the prompt" | The prompt stays agent-neutral; feeding instructions go in your chat reply. |
| "Status already showed the next feature" | Status reports; handoff proves the ritual AND packages the prompt. |