Thermo-Nuclear Review Loop v3

You are the fixer: the agent running this skill in the host session (Claude Code or Codex). You spawn one reviewer, independently verify its findings, fix the justified ones, and repeat until it explicitly approves. There is no separate fixer agent.

Remedies are your call. Scope is the human's. When a finding needs the human to decide whether a scenario matters at all, the gate sends that question to them instead of letting either agent assume the answer.

Establish the scope

1. Read the repository instructions (AGENTS.md, CLAUDE.md, rules, README).
2. Take the review target from the user: a commit, branch, ref, range, the current working-tree changes, or a described historical feature. Without a target, ask one question and wait. Ask for the original request, ticket, or prompt that produced the change when one exists; it is the stated requirement, and the reviewer receives it verbatim.
3. Resolve the baseline: a commit's parent, a branch's merge base, the explicit range, the repository baseline for current changes, or the commits that introduced a historical feature. Trace a historical feature into the current tree and edit the current tree.
4. Inspect status before editing. Preserve unrelated user changes and generated artifacts. If the tree already holds unrelated uncommitted changes, the loop runs uncommitted (see Commit).
5. Record git diff --shortstat <baseline> for the scope. The final report compares against it.

If the ref cannot be resolved or the scope is ambiguous, stop and ask.

Spawn the reviewer

Model: an Opus-like model, meaning the highest or second-highest model on the default plan for that agent (Claude: Opus; Codex: the top or second reasoning model). Name the tier, never a specific model id.