agent:ready starts an agent that writes a failing test, fixes the bug, runs your fast checks and opens a pull request that closes the issue. The Reliant bot opens it, so you can approve and merge it yourself.
This is autopilot-issue-worker, one of the Autopilot built-ins. Its own description says: “Never merges, never approves.”
Before you begin
- The
reliantCLI, v1.21 or later, signed in (Sign in and tokens), andgh. - The Reliant GitHub App installed on the repository, and push access to it on your own account (GitHub integration).
- A connected machine with a clone of the repository (Run a machine from the terminal). The agent works in a worktree of that clone.
- In an organization: an organization AI account that can run the authoring model (the
flagshiptag by default). See What this does not do.
1
Create a project for the clone
Run this in the clone on the machine. The project’s path must be the clone itself, or a folder holding it as a subdirectory named after the repository.
terminal
2
Turn on the issue worker
--machine takes the machine’s hostname or daemon ID. gate_command is the check that must pass before anything is pushed. The default runs gofmt and task test:short or make test-short, so set it for your repository.terminal
trusted_ids can start work; anyone else’s label or mention is skipped. Other settings, such as max_turns and lease_minutes, are in Autopilot settings.3
Create the label
The workflow creates its other labels on its first run, but that run starts from this one, so create
agent:ready yourself:terminal
4
Label an issue
Write the issue the way you would for a teammate: what is wrong and what done looks like.A
terminal
launched row with a chat ID means the agent started. Commenting @reliant on an open issue starts it the same way.5
Watch it work
The run takes a lease on the issue: a comment from the bot and the
agent:working label, so a second label cannot start a second run. Then it:- creates a worktree from the default branch;
- runs an agent whose shell has no credentials to write a failing test and the smallest fix;
- runs
gate_command, and retries the agent once with half the turns if it fails; - pushes an
autopilot/issue-<number>branch, never the default branch and never with--force.
reliant workflow follow <chat> or under Workflows → Runs.6
Review the pull request
terminal
reliant-labs[bot], says Closes #2, and carries agent:authored, agent:in-review and a size label (size:S under 60 changed lines). Its footer reads “Never merged by Autopilot; a human reviews and merges.” The issue swaps agent:ready for agent:in-review and gets a comment linking the pull request.If the agent could not finish, the issue gets agent:punted (it needs your decision) or agent:blocked (it tried and failed) and a comment saying what is missing. Add agent:ready again once you have answered it.Review the change, then approve and merge it yourself. Because the bot is the author, a branch rule that requires another person’s approval works as it does for a teammate.7
Turn it off
Pass
--project-id. Without it, deactivate pauses the issue worker in every project you own.terminal
What this does not do
- It never merges or approves. If you want a pipeline that merges, you compose it yourself: a workflow you write can call
github/pr.merge@1in an action node, pinned to the reviewed head commit (Merging is a workflow step). - It does not pick issues. Only issues you or a trusted account opted in are worked. On-call files issues as
agent:triage, and a person promotes them. - It does not touch CI, secrets or deploy configuration. It punts those issues, and changes over about 400 lines, back to you.
- It has no budget of its own. One attempt runs up to 120 agent turns, plus a retry of up to 60 (What it costs).
- In an organization with two or more members, runs spend only the organization’s AI accounts by default (
Unattended funding: org_onlyinreliant trigger get). The CLI cannot widen that. A person allows your own accounts on the automation in the app, or an admin adds an organization account.
Troubleshooting
This automation runs only on its organization's AI accounts, and none of them can run the flagship model
This automation runs only on its organization's AI accounts, and none of them can run the flagship model
reliant trigger events shows the firing as failed with no chat, and reliant trigger get shows Health: degraded. Nothing was written to the issue, and a second failure in a row pauses the activation. An admin can add an organization API key or turn on Reliant credits, or you can set model to one an organization account can run. Then label the issue again.'agent:ready' not found
'agent:ready' not found
The repository has no such label yet. Create it first, as in step 3.
No checkout of owner/name was found in the worktree
No checkout of owner/name was found in the worktree
The project’s path does not hold a clone whose
origin is this repository. Point the project at the clone, or at a folder holding it in a subdirectory named after the repository.The repo's fast gate failed, so nothing was pushed or opened
The repo's fast gate failed, so nothing was pushed or opened
gate_command failed after both attempts; the issue comment shows the end of its output. Check that it passes in a clean clone on the machine.Next steps
Review every pull request
Have the PR reviewer read what the issue worker opens.
Answer review comments
The review responder pushes fixes for feedback on
agent:authored pull requests.Autopilot reference
Every setting, label, lease and cost of the five Autopilot workflows.
Who an automation acts as
The Reliant bot, your own account, and the push check behind both.