Clean Code Review

review_pull_request is public GitHub only, but review_pasted_code takes private work: a unified diff, files each under a // file: path comment, or a single file, up to 1,000,000 characters. The hosted page and MCP server talk to the models anonymously, so if privacy is the deciding factor the better move is self-hosting the Next.js project, where the deployment authenticates to the AI Gateway with the project's OIDC identity and stores no API key.

No. The limit is 24 code files per review, taken largest first, with 64,000 characters per file read in windows of 16,000. Images, binaries, lockfiles, minified and generated files are skipped entirely, and anything not judged comes back in the response with a reason rather than silently disappearing. Prose files such as Markdown and AsciiDoc are shown read-only, up to ten of them, and neither model sees them.

The deployment carries brakes at every layer because of it: Vercel Firewall rate limits per client IP, a $15 weekly AI Gateway budget, a per-session spend cap, per-address limits on new sessions and MCP calls, and a shared model budget per hour and per UTC day counted in the Runtime Cache. A refused turn says the budget is spent and when it resets, answers already on screen stay, and a review answered wholly from cache is still served.

No. The decision and paragraphs are model output shaped by the submitted code, and the README is explicit that they are advice, never authorisation to merge. The structure is meant to make them checkable rather than authoritative: Jev answers with probabilities and scores instead of sentences, and the question set is per file, with the test question asked only on test paths and the Boy Scout question only on diffs.

One durable eve session per browser tab. A judge turn sends the files as JSON and fans out one evaluate() call per file to Jev (typesafe-ai/jev through the AI Gateway) in parallel; a summarize turn runs Luna (openai/gpt-5.6-luna-fast) one call per six files plus one overall and streams the review. Cancelling the turn aborts the calls.

Every judged file's answers by question id, Luna's decision and paragraphs, the prose files, the files that were not judged with a reason, the model ids and the cost. A fresh review takes 2 to 8 seconds, a 24-file pull request included, and up to about 60 seconds when Luna is slow.

A yes/no answer is a finding at or above its question's own cutoff rather than at 50%, and the questions Jev read most loosely carry definitions of yes and no; both were fitted against blind, code-quoting reviews of 200 public pull requests and checked on 25 others. The set is per file: the test question is asked only on test paths and the Boy Scout question only on diffs.