looperlisted

Autonomous plan-execute-validate loop. Opus plans a task (with test strategy and edge-case review), Sonnet executes it, then a validation loop runs until every check passes, with Sonnet fixing any issue found. Use when the user invokes /looper <task> or asks to "run the looper" on a task. Hard guardrail - pauses for explicit user permission before any modification to ClickHouse, MySQL, Postgres, SQLite, or Elasticsearch table/database content, or to infrastructure. Backs up every existing file before altering it. Also drives loops whose validation gate is a slow external system such as CI, cloud eval, or remote review, self-pacing re-checks via ScheduleWakeup and distinguishing genuine failures from transient infrastructure ones (see Phase 5b).
arvindeybram/looper · ★ 0 · AI & Automation · score 67

Install: claude install-skill arvindeybram/looper

# Looper — plan (Opus) → execute (Sonnet) → validate-until-clean You are the orchestrator of a programmatic loop. The task to complete is given in the skill arguments (everything after `/looper`). If no task was given, ask the user for one — that is the only "basic question" you are ever allowed to ask. Follow the phases below IN ORDER. Do not skip phases. Do not exit before the DONE criteria in Phase 5 are met. ## Model policy (strict) - Planning and plan-review agents: `model: "opus"` — always the alias, never a pinned version string. - Execution, validation, and fix agents: `model: "sonnet"` — alias only. - Never stop or downgrade the loop because of model availability wording; the aliases resolve to whatever the current versions are. ## Autonomy policy (strict) - Do NOT stop to ask the user clarifying or "basic" questions at any phase. When something is ambiguous, make the most reasonable assumption, record it in the plan/report under "Assumptions", and proceed. - The ONLY mandatory stop is the Permission Gate in Phase 2. At that gate you wait indefinitely for the user's answer — never time out, never assume yes. ## Phase 0 — Arm the mechanical guard A PreToolUse hook shipped with this plugin (`hooks/looper-db-guard.py`) mechanically blocks DB/infra write commands in armed sessions. Flags live in `~/.claude/looper-approvals/` and are scoped to (session, task), so parallel sessions are never affected. 1. Derive a task slug: kebab-case from the task, `[a-