ReviewMint

A proactive pull-request reviewer that reads the full change, runs the test signal, creates a fix branch, and learns how your team reviews.

Developer Tools track - OpenAI Build Week 2026

What it does

ReviewMint is designed for a 3-50 person engineering team where senior engineers are the code-review bottleneck. On a pull request event it gathers the changed files, test output, and learned team conventions, then returns only high-confidence findings. For substantive problems, it creates a sibling fix branch, reruns checks, and supplies a one-click apply action.

The local experience is intentionally complete and judge-testable: it contains a representative PR, a failing test signal, full-file review context, personalized conventions, line-level feedback, and a verified patch. It runs without accounts or API keys.

Run locally

cd mergeclaw
pnpm start
# open http://localhost:3000

No install is required for the demo because it uses Node's built-in HTTP server. Use Node 20+.

pnpm test

Production flow

GitHub pull_request webhook
  -> verify SHA-256 signature
  -> queue review context
  -> isolated sandbox runs repository tests
  -> GPT-5.6 reasons over full files + diff + test output + team style
  -> GitHub line comments and sibling fix branch
  -> tests rerun on fix branch, then PR summary card

src/webhook/ contains a tested, framework-neutral signature validator and event parser. The demo's local reviewer (src/services/rule-engine.js) is deterministic so judges can reliably exercise the complete product surface. In deployment, replace that adapter with the GitHub API, a queue, and a sandbox runner while preserving the reviewPullRequest contract.

GPT-5.6 integration

GPT-5.6 is the review reasoning layer, not a text formatter. Its production prompt receives:

full changed-file contents alongside the diff, so it can reason beyond a single hunk;
sandbox test and coverage results, so comments are grounded in evidence;
historical team conventions, so it avoids generic nitpicks and catches what a team actually blocks;
a structured response contract for line comments, risk score, verdict, and executable patches.

The UI exposes this context to make the model's decision-making legible. Configure OPENAI_API_KEY and OPENAI_MODEL=gpt-5.6 in .env when wiring the production adapter.

Safety and trust

Every sandbox run is isolated and has no shared repository state.
Generated patches go to sibling branches; ReviewMint never pushes to a contributor's branch.
The default action is a review, not a merge. A human still applies the verified patch.
The review engine reports only evidence-backed findings and retains the test output that informed them.

Codex build log

This repository was built in the primary Codex Build Week session. Codex was directed to turn the implementation plan into a judge-testable product, then to add a deterministic review engine, a signature-verification boundary, tests, and a responsive GitHub-native review workspace. The current /feedback session ID should be added by the submitter before Devpost submission.