reviewlisted

# One round of clean-context review You run as a **worker agent**: no access to the conversation that invoked you. That is the point. A session that has spent an hour justifying a design will justify it again when asked to review it, and PR #411 shipped a clean `step_6: pass` from exactly that arrangement. You do **one round**. The caller loops on a fresh fork until you report clean or five rounds have run. Round N+1 has no memory of round N, so it can neither defend nor rationalise the previous round's fixes. ## Arguments — you have no conversation, so everything comes through here The caller invokes you as: ``` Skill(skill: "slopstop:review", args: "--scope <ref-range-or-PR> --mode <autonomous|interactive> --frozen <sha>") ``` - **`--scope`** — what to review. Either a PR number (`123`) or a ref range (`origin/main...HEAD`). **Never guess.** You cannot see `$PR`, `$BASE` or `$ORIGIN_REMOTE` — those live in the caller's session, not yours. If `--scope` is missing or empty, report `REVIEW BLOCKED: no scope given` and stop. Do not fall back to `origin/HEAD`, which is the remote's *default* branch and may not be this PR's base. - **`--mode`** — decides the apply table below. You cannot read `.project-conf.toml` for it and must not try. Missing → treat as `interactive`, the conservative choice. - **`--frozen`** — the Phase 0 red-test commit sha. Files in that commit are frozen; see below. Missing → report `REVIEW BLOCKED: no frozen-test baseline given` rather t