Example tasks
Tasks grouped by the problem they solve. Each names the trigger that starts its runs, what a run does, and what it triggers next. Most factories start as one of these and grow by chaining the next.
Untriaged tickets pile up
Explore task
- Trigger: a ticket is created, or a human labels it.
- Does: checks the ticket against the codebase and sorts it into a category (easy call, needs a human, not a bug).
- Triggers: an implementer run for easy calls. Anything else ends with the ticket labelled as waiting for a human.
Ready tickets wait for someone to build them
Implementer task
- Trigger: a ticket is labelled ready, by a human or an explore run.
- Does: builds the ticket until the automated checks pass, and opens a draft PR.
- Triggers: an automated review run.
PRs wait for review
Automated review task
- Trigger: a PR opens with no automated review yet, or a human labels it.
- Does: reviews the diff against the codebase's standards and the ticket it came from.
- Triggers: a rework run on a fail. On a pass it ends, and the PR waits for human review before a human merges it.
Failed automated reviews sit unfixed
Rework task
- Trigger: an automated review run fails a PR.
- Does: fixes only what the automated review named, and pushes.
- Triggers: an automated review run. After a set number of fails, it ends with the PR labelled as waiting for a human.
Specs are too big for one run
Decompose task
- Trigger: a spec ticket is labelled.
- Does: splits the spec into tickets, each recording which tickets it depends on.
- Triggers: an implementer run.
The implementer run here picks the next ticket whose dependencies are all closed, and triggers itself again until no ticket is left.
Migrating to a stricter automated check
A new lint rule (or a stricter type setting) is switched on, but only some files pass it. The rule enforces on the passing files, and the failing ones sit on an exception list.
Ratchet task
- Trigger: a daily clock.
- Does: takes one file off the exception list, fixes it until the stricter check passes, and opens a PR.
- Triggers: an automated review run. Once the exception list is empty, the ratchet run has nothing to take and the migration is done.
One small PR a day keeps human review cheap, and the check only ever gets stricter.
Users report bugs nobody reads
A feedback button in the app is a ticket source: it files each report as a ticket with the explore label already on it. The explore run does the rest.
Production problems are found late
A monitor (slow queries, error rates) is a ticket source. It files a ticket per problem, and comments on the open ticket instead when it sees the same problem again. After a fix merges, it reopens the ticket if the problem persists.
The factory's own runs fail
Fix task
- Trigger: the orchestrator reports a failed run.
- Does: files the failing log as a ticket, and attempts a fix as a PR when the failure is fixable.
- Triggers: an automated review run. A failure in a fix run's own work triggers nothing, so fixes never chain. Because the fix changes the factory itself, it always gets human review.
The codebase drifts
Architecture task
- Trigger: a nightly clock.
- Does: scans recent changes for refactoring candidates and files them as one ticket.
- Triggers: nothing. A human picks a candidate and labels it, which starts the usual chain.
The same mistakes keep coming back
Retro task
- Trigger: a PR merges, throttled to one merged PR in ten.
- Does: reads the PR, its ticket, its automated reviews and its run logs, and proposes changes to the factory's own instructions as a PR.
- Triggers: an automated review run.