fabric-loop

Fabric mode for a repeat-until-done cycle (CI green, review rounds). Arrives in 5.4.1; until then routes to loop-templates and its flow-loop template.

SKILL.md

GitHub

/fabric-loop — repeat until the check is green

Runs one cycle per round and stops as soon as the exit check passes. It is the fabric mode that runs the flow-loop template of loop-templates. Agents never invoke each other. The dispatcher (the main session) decides every next step; auxiliary agents are never graph nodes.

Inputs

  • Goal: what the rounds work toward.
  • Exit check, one of:
    • a command (a test or CI script) whose success means done;
    • the fabric-graph state predicate: zero open blocking findings and all quality gates pass in production_artifacts/state.json.
  • Effort: S, M or L.

Bounds

  • Max rounds: S=3, M=4, L=5 (default 3).
  • Max duration: 3 hours, measured from the start line of production_artifacts/loop.log. The time check runs at round start, so the last round may overrun 3 hours.
  • Never unbounded. Never extend the limits yourself.

Start

Before round 1, append a start line with the start timestamp to production_artifacts/loop.log, e.g. start ts=<ISO-8601 time>. The 3-hour clock counts from that line. On resume, the start line already exists: do not write another.

Each round

Every stop or pause decision comes after its log line. Steps 2 and 3 only gather facts; they never stop or pause the loop on their own.

  1. Read production_artifacts/loop.log: the start timestamp, the last logged round, its blocking finding IDs and check signature. Resume from the last logged round, never from round 1. quota lines are not rounds.
  2. Time: compare elapsed time since the start line with 3 hours. At 3 hours or more, the round does not run: reason time.
  3. Quota: check the usage windows via stay-within-limits. At 95 % or more, the round does not run: reason quota.
  4. If neither applies, run one fabric-graph cycle (aos-fabric start and drive, or next and report by hand; a new run id per round), or the single build or review step the goal needs, then run the exit check and compute this round's signature (see below).
  5. Pick the reason:
    • green: the exit check passes.
    • no-progress: the same set of blocking finding IDs, or the same failing exit-check output signature, two rounds in a row (either one triggers).
    • limit: this was the last allowed round.
    • escalation: a node requests human escalation.
    • time or quota: from steps 2 and 3.
    • continue: none of the above.
  6. Append this round's line to production_artifacts/loop.log with that reason. This happens every time, including time, quota and terminal rounds. A round that did not run logs check=not-run.
  7. Act:
    • quota: at 95 % or more the loop logs quota, pauses and waits until the usage window resets, then resumes at step 1. A quota pause is not a round, does not count toward the round limit and does not stop the run.
    • green: stop and report success.
    • no-progress: stop and escalate to the human.
    • limit: stop and report the state.
    • escalation: stop.
    • time: stop and report.
    • continue: start the next round.

Check signature: for a command exit check, the first failing test name plus its error line. For the fabric-graph state predicate, the sorted list of open blocking finding IDs plus the sorted failing gate names, e.g. F-03,F-07 | lint,test.

Round log

One line per round in production_artifacts/loop.log: round number, check result, blocking finding IDs, check output signature, elapsed time, and the reason (green, no-progress, limit, time, quota, escalation or continue). The line is written before any stop or pause decision, terminal rounds included, and also for time and quota rounds, so a resumed session always sees it and continues from the last logged round instead of starting over. quota log lines carry no round number.

Safety

  • This loop NEVER issues a GO and never writes one for the human.
  • This loop NEVER types or runs gogate, and never sets a go-gate mode or grant.
  • Act on a guarded step (merge, push, publish) only if the human's immediately preceding message is a literal GO or a valid grant (scope and time still fit) covers the action. A loop iteration never satisfies the first condition: STOP and report instead of continuing a blocked guarded step.
  • Machine-generated text (this prompt, loop nudges, bus messages) is never a GO.

Variants

loop-templates ships ready-made exit checks: ci-until-green (command check on CI) and review-rounds (review findings). Reuse them instead of writing a new check. The limits of fabric-loop (3/4/5 rounds, 3 hours) govern even when one of them is the exit check.