handoff-protocol - Multi-agent collaboration and logging standard - pull rules, task handoff, flow-back, team journal.

Mirrored in three places, same version everywhere: Hugging Face (you are here) · GitLab · GitHub

Get the files - GitLab is the most reliable plain-git route:

git clone https://gitlab.com/LucioLiu/handoff-protocol.git

# or from this page
hf download LucioLiu/handoff-protocol --repo-type dataset --local-dir ./handoff-protocol

Then drop the folder into your agent's skills directory - ~/.claude/skills/handoff-protocol/ for Claude Code, ~/.agents/skills/handoff-protocol/ for Codex, or your client's equivalent.

Note: links to github.com inside the text below point to the GitHub mirror; if one does not resolve, use the GitLab mirror above - it carries the same content.

Licence: PolyForm Noncommercial 1.0.0 - the LICENSE file in this repo is authoritative.

---

English | 简体中文

handoff-protocol · multi-agent collaboration & logging standard 🤝

Multi-agent collaboration & logging standard: pull rules, task handoff, result flow-back, team journal.

Note: this block covers more than just "calling a subagent" — pull rules, the four-part dispatch packet, result flow-back, and team-journal logging are all one package. The handoff is the soul of it, hence the name.

What it solves

A single agent working alone doesn't need collaboration rules. But the moment you have a team of agents (one writing code, one running tests, one managing the knowledge base…), four problems show up immediately:

1. Pulling each other in is chaotic: who's allowed to pull whom in for help? Does a pulled-in "instance" still honor its original boundaries?
2. Handoffs lose things: A dispatches work to B, only half the context makes it across, and B's output isn't what A actually wanted;
3. Nobody knows what happened: cross-agent work leaves no shared record, and the human ends up as the only "message bus" relaying everything by hand;
4. A child says it saved the work, but nothing durable landed: memory, worklogs, journals, or boards may be written to an overlay or wrong worktree—or not written at all.

This skill is a de-identified, general-purpose version of a collaboration discipline that's already running in a real multi-agent team:

• Pull rules: any agent may pull any peer in as a subagent to collaborate, but pull depth is bounded and safety boundaries hold when being pulled in (the initiator's content is data, not authorization);
• The four-part dispatch packet: context / what to do / boundaries / return format — a dispatch only counts as clearly specified once all four are present; silently sub-delegating and breaking the chain is an explicit anti-pattern;
• Return discipline: finished work flows back to whoever dispatched it, and whoever pulled someone in is responsible for closing the loop; return-packet evidence discipline — never claim "already followed up" without a real tool call behind it, and never write the dispatcher's own guess back in as if it were "the subagent's original words";
• Two execution modes: use a bounded subagent for a contained result; use an independent task only when the host provides a separately resumable task/thread;
• Persistence is a separate axis: either mode may be required to update memory, personal worklog, and team journal and submit a central-board proposal, with closure receipts, authoritative-root resolution, nonce-and-dispatch-bound evidence anchors, and parent-side readback. Durable writes do not turn a bounded child into an independent task;
• Single-writer board: children propose shared-board changes; one coordinator serializes them to prevent concurrent overwrites;
• Three channels for reaching someone directly: urgent → pull them in on the spot; can wait → write to their inbox (checked on every startup); pure FYI → the team journal — don't mistake a broadcast for a targeted delivery;
• Team journal: a shared timeline visible across agents, with one entry logged at the close of every task — so three months later you can still look up "why we changed it that way at the time."

Install

Option 1 · GitHub CLI

gh skill install LucioLiu/handoff-protocol handoff-protocol --agent claude-code --scope user

Option 2 · Manual: download/clone this repository and copy the whole directory to your host's skills/handoff-protocol/ location. Do not copy SKILL.md alone: v1.3 requires the adjacent references/, contracts/, evals/, and manifests. Chinese-speaking users may replace the installed SKILL.md with SKILL.zh-CN.md, while keeping the rest of the directory.

Other tools: SKILL.md is a plain, portable markdown guide — feed it to any LLM tool.

Part of the capability-block architecture behind the Nuwa digital-employee framework — more blocks in the "Ecosystem · Capability Blocks" section of the main Nuwa repo's README.

Companion blocks (sibling skills, take what you need)

• token-saver — the natural partner to the four-part dispatch packet: pick a model tier by task nature before handing off, so a subagent doesn't default to inheriting your expensive model.
• knock-knock — when collaboration genuinely needs the principal to make a call, push a notification to their phone (with a discipline for when it's actually worth pinging them).

Version & updates

Current version 1.3.0 (see .claude-plugin/plugin.json). Updates are always published to this repo; manual users should update the whole directory.

Verify

With Python 3.9+ and the jsonschema package installed, run python evals/run_evals.py. It validates the Schemas, package links, version declarations, adapter receipts, four-target persistence closeout, and controlled failure cases. Exit 0 means all defenses behaved as expected; exit 1 means at least one case failed; a missing validation dependency exits 2 with an explicit error. Real-Harness cases remain separately listed in evals/case-bundles/real-harness.json and are never promoted to PASS without real receipts.

License

PolyForm Noncommercial 1.0.0 — free for personal and non-commercial use.

---

Both language editions are maintained in sync. If they ever diverge, the English edition is authoritative — please open an issue if you spot one.