A repository protocol agents can actually follow.

Principles

1. One issue, one branch, one pull request.
2. A worker claims exactly one claimable issue before editing code.
3. A reviewer never approves work produced by the same agent.
4. The repository and GitHub API are the only coordination state.
5. Every role has explicit GitHub permissions and explicit prohibitions.
6. Failed checks and requested changes stay with the current issue.

Role boundaries

Every session adopts exactly one role. The orchestrator decides which role a new session should adopt by reading repository state; every other role then works within its own boundary alone.

Labels

Labels are state, and their names are configurable in .rallypoint/config.json. An open issue is claimable unless something holds it back: no assignee and none of status:claimed, status:blocked, status:deferred or needs:human. There is no readiness label.

Install

Rallypoint’s CLI reads its own checksummed manifest and writes each file to an explicit repository-root path. It stops on conflicts rather than overwriting anything silently. An existing AGENTS.md gets special handling: it is left untouched, the rest of the loop still installs, and the CLI suggests adding a one-line @AGENTIC_LOOP.md reference to it.

init requires the GitHub CLI, installed and authenticated. It syncs the configured integration branch to the repository’s default branch, creates the labels the loop reads and writes, and reports — without changing — whether the default branch has protection rules.

Prompting

A role is one file, so at minimum, prompting an agent to become that role means pointing a fresh session at AGENTS.md and exactly one file under .rallypoint/roles/. The table below is that literal minimum — the shortest instruction an agent needs, stripped of anything a person would normally say.

Don’t know what should happen next? Prompt the orchestrator first. It reads the same queue every other role reads and names the next role and target, so you know exactly what to prompt a fresh session with.

How a person actually asks

Nobody talks like the table above. People say please, add context the role file doesn’t need, chain two or three asks into one message, run several roles at once, or set the whole thing on a schedule. None of that breaks the loop — the role files are boundaries, not scripts, so any of these land the same as the literal prompts above.

• A simple, polite ask

  Could you grab the next ready issue and ship it? Thanks!

• With extra context the role file doesn’t need

  A couple of users got stuck on the auth step in yesterday's onboarding call — can you check the backlog reflects that, add whatever's missing, and reprioritize anything blocking a fix?

• Chained into one message

  Run the planner to top up the backlog, then have a worker pick up whatever's highest priority, and kick off review as soon as that PR is open.

• Several roles at once

  Spin up three workers in parallel on the three oldest ready issues, and have a reviewer standing by for whichever one finishes first.

• On a schedule

  Set this up to loop: run the janitor every morning to clear stale claims, and have the orchestrator check in every few hours to tell me what's next.