| name | loop-goal |
| description | Keep one agent task progressing linearly until an objective completion condition passes, with fresh state checks and hard resource limits. Use when the user has one long-running goal that does not need parallel PR orchestration or a changing workflow graph. |
Loop Goal
Keep one thread moving toward one outcome. Do not turn it into a swarm unless the work genuinely branches.
Workflow
- State the outcome as a verifiable predicate and record the current evidence.
- Set iteration, elapsed-time, and repeated-failure limits plus any required human gates.
- Work on the smallest next action that can change the predicate.
- At the end of every turn, use fresh evidence to evaluate completion. Do not accept self-reported completion without a check.
- If incomplete and progress was made, persist the new state and continue through the product's goal or wakeup mechanism.
- If incomplete and the same failure repeats without new evidence, change approach once, then escalate at the configured limit.
- Stop immediately for permissions, destructive requirements, ambiguous intent, or a human approval gate.
- Finish with the predicate, evidence, iterations, elapsed time, work completed, and remaining blocker.
Prefer this shape for one deep task. Use $loop-project when the goal becomes several independently reviewable work streams.