← ClaudeAtlas

omega-looplisted

Autonomous multi-model engineering loop (plan → build → isolated adversarial audit → verify → commit) with automatic Codex detection and Claude-only fallback. Use when the user invokes /omega-loop or asks for the OMEGA-LOOP protocol on a task.
priore/omega-loop · ★ 0 · AI & Automation · score 72
Install: claude install-skill priore/omega-loop
# OMEGA-LOOP (global) Tooling lives in `~/.claude/omega/`. Per-repo state lives in `.omega/` of the current repo (globally git-ignored). Task = skill args. Rules: the writer never reviews its own diff. Project CLAUDE.md rules still win (commit-per-phase confirmations, CHANGELOG, etc.). Never use `--no-verify`. ## Provider failover - Codex out of tokens/erroring: its role (architect, builder, auditor) moves to Claude immediately, without stopping the loop. `audit.sh` does this by itself (codex → claude). - Claude out of tokens: `audit.sh` falls back to Codex; the rest of the loop resumes from the terminal with `~/.claude/omega/codex_takeover.sh` (reads `.omega/` and continues). Claude cannot launch it once exhausted: ask the user to do it, and always keep STATE.json/DECISIONS.log current so resuming is possible. - Every provider switch must be written to DECISIONS.log. ## Model routing (runtime) Orchestrator = the model selected in the session (Sonnet, Opus, Fable). Everything else goes to local Qwen (Ollama :11434 / LM Studio :1234) when available. - Qwen unavailable: Sonnet replaces it; if Sonnet is already the orchestrator, Haiku is used. - Delegate read-only side tasks (searches, summaries, validations) with `OMEGA_MAIN_MODEL=<your model> ~/.claude/omega/subagent.sh "<prompt>"` instead of the Agent tool. `PROVIDER_STATE.json` shows `qwen` and `delegate`. - Audit order in `audit.sh`: codex, qwen, claude (Claude-only mode: qwen, claude, codex). Log every routing change i