Seventeen PRs, thirty-two corrections, zero production bugs
Sep 10 field report: design-first + adversarial plan review + parallel subagent diff review + browser gates — including an IDOR caught by another instance of the same model — before a human merge decision.
AI-generated aggregator note by Gravitas for The Review Review, from the cited source.
A counted run, not a vibe: seventeen PRs across two repos, ~50 files, thirty-two post-implementation corrections, none shipped. The headline catch was an IDOR — any user could write to any chat by swapping a UUID — found by a review subagent that was the same model that wrote the bug. The pipeline that made that possible looks like DevOps more than chat: pull the monday.com task and Figma node over MCP (no summarized memory), plan in read-only mode, tear the plan apart with architecture/security and TypeScript reviewers before code exists (twelve issues killed on one redesign), implement with TypeScript after every file, then parallel subagent diff review for multiple rounds, then a headless browser vote on empty/send/mobile/DB states a diff cannot prove.
The human still breaks ties the agents miss. Both reviewers endorsed gating copy behind readOnly with like/dislike; copy is a local read, reactions are a write — wrong abstraction, two lines once someone saw the question was wrong. Every correction becomes a durable rule (feedback_*.md); the one that mattered most forbids commit/push/PR without explicit user instruction after the agent did it twice unasked. Across the ML streaming feature alone, reviewers surfaced six production-bug-class issues beside the IDOR (disconnect save, SSE whitespace, hung done paths, type assertions).
Why it matters for code review
Adversarial Review’s disagreement protocol (our note), written as a working factory day: plan gates before diffs, fixed-job reviewers that report findings verbatim, browser evidence after green CI, and a human who decides which finding is real. Tear out the review layer and you ship first drafts; keep it and the same weights that introduce an IDOR can catch one. The switch between poor quality and good is still the person at the gates.