oh-my-agents

Use when behaviorally correct code needs bounded simplification of AI-style duplication, dead code, needless abstraction, boundary leaks, or weak coverage without changing intended behavior.

Use when a repo-local question needs read-only cross-file explanation; use trace for causal failures and direct reading for one-file facts.

Use when the user hands over a whole task and wants it driven end to end, says "autopilot this", "take it end to end", or "run with it".

Use when a decision depends on current external, version-aware guidance from official, upstream, or standards sources; use analyze for repo-local facts.

Use when the user requests a read-only review of a working tree, branch, pull request, commit range, patch, or other bounded change set without implementation or pair delivery.

Use when an idea is vague and needs thorough clarification before implementation, or the user asks to be interviewed without unstated assumptions.

Use when the user explicitly asks to continue a relay, hand off to Claude or Codex, start an agent pair, or deliver a change with independent cross-agent review.

Use when work must iterate until a verifier passes or a deterministic stop condition is reached, especially for keep-going, loop-on-this, or do-not-stop-until-green requests.

Use when a just-performed workflow appears likely to recur across sessions or projects and may justify the resident cost of an oma skill.

Use when a bug, regression, performance issue, surprising result, or architecture outcome needs adversarial causal investigation before any fix.

Use when a feature or flow needs adversarial end-to-end hardening before shipping because happy-path verification is insufficient.

Use when an open-ended optimization task must beat a baseline score or pass a deterministic evaluator under a fixed budget; use ralph for a single pass/fail verifier.

Use when the user explicitly requests a throwaway runnable experiment, state-logic spike, or side-by-side UI variants before production implementation.