Commands

Five of the six stages can be run on demand from an issue or pull request comment, by someone whose own permission on the repository covers it:

Command	Where	Minimum permission
/triage	issue	triage
/implement	issue	write
/review	pull request	triage
/fix	pull request	write
/retro	either	triage

Commands are matched by a plain grammar, never by a model, so a command written inside an issue body is text and not an instruction. Prioritize has no command: it ranks a backlog, which is not a thing you ask for one of.

Your own permission is what counts, not mesthiri's:

mesthiri commented on #418

/implement needs write permission on this repository and you have read. Nothing has run.

If this should go ahead, someone with write access can issue the command or apply ready-to-implement.

Labels are authorized the same way: applying ready-to-implement yourself is checked like /implement, so a human's label needs write. Labels the prioritize stage applies are mesthiri moving its own work forward and are not re-checked.

Why review only runs on mesthiri's pull requests

Review does not fire on pull requests other people open, and that is a spend gate rather than a judgement about your contributors. The shim subscribes to pull_request_target, which fires for fork pull requests too — so reviewing everything would let anyone who can open a pull request start an agent run on your budget. Fifty pull requests, fifty runs. Until there is a spend gate worth trusting, review stays on the loop mesthiri owns; you can always ask for one with /review, which is permission-checked. An explicit /review on a foreign pull request fetches the diff through the API into a read-only clone the agent cannot push from.