Agent guide

This page is the short version. The authoritative agent guide lives in AGENTS.md at the repository root — pipe it in at session start with:

clu brief

The agent loop

1. Identify yourself with -a <your-name>.
2. Discover work with clu ready --json -a <your-name> (or watch).
3. Atomically take work with clu claim -a <your-name>.
4. Comment your progress and decisions with clu comment add.
5. Close on success; cancel on irrecoverable failure.

Push-style task delivery

Don't poll. Use Claude Code's Monitor tool against a --watch invocation:

Monitor: clu ready --watch -a code-reviewer

clu suppresses unchanged ticks, so Monitor turns each new state into one notification. No while true, no diff-against-seen.

Things to avoid

• Don't fight other agents over the same lane. If you're filtering by agent name, claim from your lane. If you're a generalist, claim from the unassigned lane.
• Don't bypass checkpoints. A checkpoint:pending label means a human (or another agent) needs to clear it via clu approve (the approver list is informational, not enforced).
• Don't hand-edit .clu/data.sqlite. Use the CLI; it knows about labels, KV side tables, and the cascade walk.