agg-newlisted

# /agg:new — set up an AgenticGoGo loop for this project You are setting up **AgenticGoGo** (`agg`): a loop that runs fresh headless agent workers until a **Definition of Done** — expressed as judges — is met. Your job in this skill is to turn whatever planning material already exists into the files the loop reads, all under the mandatory `agg/` folder: - `agg/agg.yaml` — the whole config: `defaults` / `judge` / `steps` / `sequence` - `agg/judges/<name>.sh` (a script judge) or `agg/judges/<name>.md` (an LLM rubric judge) — one per clause of the Definition of Done. **A judge IS a goal.** - `agg/AGG.md` — the stable scope/goals/architecture each worker reads for orientation (committed) - `agg/state/STATE.md` — the forward "what to do next" advice the worker rewrites each session (gitignored; agg composes it + AGG.md + memory into a per-session `agg/state/INSTRUCTIONS.md` brief) **Core principle: do NOT replicate spec tooling.** Read what's already there and *translate* it into judges. Only ask the user for what's genuinely missing. **The model in one sentence: your judges ARE the gate — your Definition of Done, made executable.** Each judge decides one clause; `done_if` composes them into "done"; agg runs the judges, never the worker, so the loop cannot fake its own completion. `agg` can drive **Claude Code, OpenAI Codex, or GitHub Copilot** — chosen by the `agent:` keys you write into `agg.yaml`. They are **not interchangeable**, and `agg` REFUSES to start a run whose