cast-loop

A loop is only as good as its dispatcher and its gate. The task board is the dispatcher: ranked, deduplicated, visible to the human. Verification is the gate. Merge stays a human decision unless the human said otherwise.

Queue

Confirm there is a queue worth looping over: cast task ready --plan <id> or --project <ref> or -q "<query>". Each task must be finishable by one session alone with acceptance criteria in its description; fix the ones that are not before starting, or the loop produces eight half answers.

Arm

cast trigger add - --every <interval> --spawn --title "Loop: <name>" \
  --precheck 'cast task ready --plan <id> --json | jq -e "any(.[]; (.stale | not) and ((.assignee // \"agent:\") | startswith(\"agent:\")))" >/dev/null' <<'BRIEF'
Claim the next ready task with `cast task ready --plan <id> --claim` (it starts the
task and binds this session, so two overlapping runs never take the same one; if
it claims nothing, complete the run saying so), work in a worktree (`cast ws acquire <task id>`), and run
/cast-verify before closing it. Commit on the task's branch and open a pull
request with /cast-ship; do not merge. Queue anything that needs a person with
`cast decide`. Finish with `cast trigger complete <trigger id> --summary "<what landed>"`.
BRIEF

The precheck runs before every firing and skips the run unless a ready task is one the run's claim can take (touched in the last 30 days, and unassigned or handed to agents), so an idle loop costs no session. A run claims rather than reads the top of the list because the next firing can start before this one finishes. The brief is the run's entire context; a spawned session has none of this thread.

Watch and stop

cast trigger ls shows the loop; cast trigger log <id> the last run. Pin this session dormant naming the trigger. Stop it with cast trigger cancel <id> when the queue drains, when two runs in a row fail the same way, or at the --for limit; a loop that keeps retrying a broken task is spending the human's quota on nothing.