ai-* labels from ai-ready to ai-merged, and the labels themselves are the delivery record.
Orbi Cloud keeps no separate dashboard of truth. The labels on your Issue, the comments Orbi leaves, and the pull request it opens are the complete state of a delivery — which means you can read, audit and correct everything from GitHub alone.
The state machine
What each label means
Two labels are set by people, never by Orbi:
Timing
Orbi works on a timer rather than reacting instantly, so pickup takes up to five minutes after you addai-ready. The same tick interval applies to every transition: a delivery that needs a repair round waits for the next tick before resuming.
How long a delivery takes after pickup depends entirely on the size of the task. A small, well-specified change can finish in minutes; a large one takes longer and may go through several repair rounds. The status page shows runtime, model requests and resumption count per work item once a delivery reports usage.
The independent review
The session that reviews the pull request is not the session that wrote it. It reads the diff fresh, runs the tests again, and either returns a clean verdict or files findings. When it files findings, Orbi repairs them and the pull request goes back for another round. This loop is bounded: if rounds are exhausted without reaching a clean verdict, the Issue is markedai-blocked rather than looping forever. The pull request, branch and workspace all stay intact when that happens, so nothing is lost.
Only a clean verdict reaches the merge gate, and the merge is of the exact commit that was reviewed — not a rebuilt or rebased version of it.
Steering a delivery that is already running
You do not have to wait for a delivery to finish to correct it. While an Issue isai-in-progress, add a comment with the new decision, or edit the Issue body. Both channels are equivalent, and the running session restarts with your correction — normally within a minute.
This is the intended way to redirect work that is going the wrong way; you do not need to cancel anything.
Two limits are worth knowing:
Steering is bounded. A delivery accepts 3 steering restarts by default. After that, further corrections do not restart it — let the delivery finish and put the revised scope into a follow-up Issue.
Only trusted commenters can steer. Comments from the repository’s owner, maintainers, members and collaborators steer a run; a comment from an arbitrary public commenter is ignored by design, because otherwise anyone who can comment on your Issue could redirect what the agent does.
ai-blocked is a decision point, not a failure
ai-blocked is a terminal state that means a human must decide what happens next. The comment Orbi leaves on the Issue states why automatic recovery was not possible.
“Decide the next step” genuinely means any next step: change the code, rewrite the Issue’s acceptance criteria, or close the ticket. Removing the label alone does not restart anything — fix the underlying problem first, then re-apply ai-ready to put the Issue back in the queue.
See troubleshooting for how to read a blocked comment.
Task types
ai-ready is the execution switch for every kind of task. A second label chooses which playbook runs:
One more label affects ordering rather than behaviour:
p0 marks an Issue urgent, and Orbi picks up p0-labelled Issues before bug-labelled ones, which come before plain ones.
This page describes the lifecycle as you experience it on Cloud. The engine’s own reference — every transition, the exact resume semantics, and the full config surface — is at docs.orbi.build/workflow.