Loop Referee

Loop Referee

A reviewer running on a different vendor's model catches what the writer cannot see. The two can also disagree politely forever, at your expense. The stop rules loops ship with today are "clean" and a round cap. Real loops die in ways neither covers. This skill is the stop switch and the receipt.

How The Loop Actually Runs

No extra software. The loop lives inside one agent session, and that agent is the referee. It runs alongside the loop, not instead of it. Example with Claude Code writing and Codex reviewing:

1. You: "implement X, then run a review loop with codex."
2. The agent writes the contract (below), then implements.
3. Reviewer round: the agent runs the reviewing tool on the diff, for example codex exec "review this diff for must-fix bugs: $(git diff)", or the Codex plugin, or a review bot on the PR.
4. The agent appends a ledger row, checks the stop rules, then either fixes and loops or stops and says which rule fired.

Works the same with the roles reversed, or with any writer/reviewer pair. The stop rules bind on countable events (a reopened finding, two rounds of diff growth, the round cap), so the orchestrating agent cannot easily talk itself into one more round.

The Contract

Write these down before round 1. Changing them mid-loop is how loops run away.

• Max rounds: 3, unless the human raises it in writing.
• Severity floor: which findings the writer must fix (for example bug or security). Findings below the floor get logged, not fixed, and never extend the loop.
• Scope: the files the loop may touch. A fix outside scope ends the round.
• Budget: a token or cost cap if the platform supports one.

The Ledger

Append one row after every round. Never edit old rows.