feat(om-review-loop): review-and-fix with fresh-eyes subagents until it goes quiet

[jtomaszewski]

Adds skills/om-review-loop/SKILL.md — an Open Mercato review-and-fix loop that keeps running om-code-review until N consecutive rounds turn up nothing new.

Why

om-auto-review-pr --autofix already loops, in-process, and posts the result. The context that wrote a fix then re-reviews it against its own intent: excellent at catching self-contradiction, structurally blind to what it never considered. This skill is the subagent, fresh-eyes version of that loop — every round is read by contexts that have never heard the argument for the code — and it is local only.

The skill states the three axes where the two differ rather than restating the pipeline: reviewer context (new subagents vs. one iterating context), exit condition (N consecutive quiet rounds + the full gate vs. one actionable-empty pass + CI), and side effects (none vs. reviews, labels, claims, handoff).

The design points it pins down

• Reviewers never learn the loop exists. No round number, no ledger, no fix rationale — "round 5, nothing since round 3" is an instruction to find nothing. Prompts are identical every round except the diff; dedupe happens orchestrator-side, after the review returns, so it can never leak in as a hint.
• An append-only ledger, rejected findings included. Drop them and each round re-raises them from a reviewer who never heard the refutation, the counter never advances, and the loop runs until the budget dies. Fingerprints key on file + enclosing symbol, never line numbers — those move with the first commit, after which dedupe silently stops matching.
• Verify before fixing. One adversarial refute-by-default verifier per new finding, accepting either a concrete failure path or a fact confirmed in the repo. Each finding costs a code change, and a code change on a wrong premise is a new defect. The refutation rate is reported, because converging by dismissal looks identical to converging by fixing.
• The validation gate, split three ways. Codegen prefix once at the start; a per-round subset chosen from validation.commands and keeping every prerequisite of what it keeps (subset below the prerequisites and you invent failures the loop then "fixes"); the full gate in configured order exactly once, when the counter first reaches the threshold. A failure there is a blocker by om-code-review's own rule and resets the counter to zero — there is no finishing around it.
• Judgement calls go up, not into the diff. Design disagreements, scope questions, out-of-diff findings — handed to the user, never decided silently and never dropped. "Quiet" means nothing new, not nothing left, and the report carries the still-open ones so the exit can't read as a clean bill.
• om-code-review's severity scale and verdict rule verbatim. No second rubric.

Exit reports the findings-per-round curve, and hitting the round cap is reported as non-convergence rather than dressed up as a result.

What it claims

The report's closing section says N passes by independent fresh contexts raised nothing new and the gate passes — and then says what that isn't: not "this change is correct" (the reviewers share a model and therefore a blind spot), no runtime evidence, and no prediction about what a human reviewer will find.

Assumptions

Consuming repo is Open Mercato with the OM skills installed and .ai/agentic.config.json present. Missing either → the skill stops and says so. It never installs anything and never runs om-setup-agent-pipeline. Nothing here goes upstream to open-mercato/skills.

Also in this PR

README skill-table row, and the minor plugin.json bump AGENTS.md requires (1.3.0 → 1.4.0).

🤖 Generated with Claude Code