Run fan-out implementation experiments after askquestionsspec or planning. Generates candidate solutions, verifies them with explicit hard gates and soft checks, narrows to the best variants, and iterates until one satisfies the verifier contract or the loop hits its limits. Use when users ask to try multiple implementations, keep iterating until tests pass, prove an implementation, or search for the best verified outcome.
Use this skill after requirements are known. Prefer it when a task has multiple plausible implementations, risky regressions, or a strong verifier contract.
/experiment-loop <objective>/experiment-loop spec <path>/experiment-loop plan <path>At least one of:
spec_pathplan_pathAnd a verifier contract with at least one hard gate.
scripts/build_manifest.ts if one does not already exist.POST /v1/fleet/experiment-loops with the manifest-derived request.GET /v1/fleet/experiment-loops/:id until terminal.Useful request overrides:
round_timeout_msstagnation_limitdispatch_policyEach run should leave artifacts under:
.artifacts/experiment-loops/<loop-id>/manifest.json.artifacts/experiment-loops/<loop-id>/candidates/*/verifier-summary.json.artifacts/experiment-loops/<loop-id>/candidates/*/review-summary.mdTerminal failure reasons to report verbatim:
loop_timeoutmax_rounds_reachedstagnation_limit_reachedswarm_failed_before_verifierno_candidates_remainingreferences/verifier-contract.mdreferences/loop-strategy.mdreferences/prompt-recipes.mdscripts/build_manifest.ts - normalize objective/spec/plan + verifier into manifest.jsonscripts/run_verifier.ts - execute hard/soft checks and write verifier-summary.jsonscripts/score_round.ts - score a round from verifier outputsscripts/derive_next_round.ts - generate repair/optimize children from scored candidates