Agent loop orchestrator

Each cycle runs four steps: triage, dispatch, monitor, report. This skill runs one cycle. Repetition comes from the harness, not from you.

Scheduling

Never sleep in a shell between cycles. Use the harness's recurring-run facility (Claude Code: /loop <interval> /agent-loop:resume, or the schedule skill). Without one, run one cycle, report, and tell the user how to rerun.

Cycle

1. Triage. Follow ../agent-loop-triage/SKILL.md for the configured scope. It returns items classified as autonomous, needs_review, defer, or blocked.
2. Dispatch. Spawn one background subagent per autonomous item, at most max_concurrency at a time. Give each one the item, the repo, and pointers to ../agent-loop-autoreview/SKILL.md and ../agent-loop-browser/SKILL.md. Never batch two items into one worker. Workers touching the same repo must use separate branches or worktrees.
3. Monitor. Collect each worker's result: merged, pr_opened, failed (retry at most twice), or needs_human (with what blocks it).
4. Report. Update the state file, append a cycle summary to .agent-loop/logs/, and print it: repo, counts per class, one line per worker result, duration.

Stop early if the state file has shutdown_requested: true. Let running workers finish, write final state, and print the summary.

Configuration