version: "1.0.2" name: loop description: "Use when benchmarking and evolving prompt assets through a fixed-suite optimization loop" effort: high disable-model-invocation: true

Resolve runtime paths before file or script operations.

Iron Law

A loop without a benchmark is meaningless.

Red Flags

• "Let's just run it and see what happens" → STOP, because the Iron Law says a loop without a benchmark is meaningless — a fixed suite with numeric scoring must exist before any run.
• "The new variant feels better" → STOP, because evaluator output must expose a numeric score that can be parsed deterministically — subjective improvement is not a valid promotion signal.
• "Baseline is close enough, let's promote the candidate" → STOP, because winners are promoted only when the score beats baseline by at least min_delta — close is not enough.
• "I'll mutate the docs and test files too" → STOP, because .mjs, docs/, README*, and tests/ are excluded from mutation scope — only allowed targets in the suite allowlist may be changed.
• "One more generation might improve it" → STOP, because the loop must stop when budget is exhausted, min_delta is not met, or scores plateau — open-ended iteration wastes tokens without measurable gain.

Loop

Run a Karpathy-style optimization loop over the plugin's own prompt assets.

When to Use

• Maintainers want to improve skills/**/SKILL.md, agents/*.md, commands/*.md, or templates/*.md
• A fixed benchmark suite with numeric scoring already exists
• Candidate variants should compete under the same time budget and the best variant should be promoted to an isolated branch

Subcommands

Workflow

1. Use node scripts/loop-runner.mjs show-suite <name> to load the suite and verify the allowed targets.
2. Use node scripts/loop-runner.mjs run <name> ... to create the isolated codex/loop-... branch and run worktree.
3. Inside that run worktree, create 3-5 prompt variants only within the selected targets.
4. Evaluate every candidate with the suite's fixed case commands. Parse scores from JSON or review-style output only.
5. Keep the top 1-2 elite candidates per generation. Stop when budget is exhausted, min_delta is not met, or the loop plateaus.
6. Only promote a winner when all hard gates pass and the score beats baseline by at least min_delta.

Hard Gates

• Selected targets must stay within the suite allowlist and the global v1 mutation scope.
• Contract/runtime smoke checks must still pass when those files exist in the repo.
• Evaluator output must expose a numeric score that can be parsed deterministically.
• Winners are promoted only into the run branch/worktree, never into the current workspace.

State

• Active state: <plugin-data>/state/loop-active.json
• Artifacts: .captures/loop-<run_id>/
• Bundled suites: benchmarks/loop/*.json

Gotchas

• Do not add routing for this skill in prompt-detect.mjs for v1. It is slash-command only.
• Do not mutate .mjs, docs/, README*, or tests/ as loop targets.
• Do not declare a winner when baseline remains best or when min_delta is not met.
• Do not write to the main worktree directly. The isolated loop branch is the output surface.