pr-triage

PR Triage is a CI cost-optimization skill (documented in Japanese) that analyzes the contents of a pull request and emits a structured .pr-triage.json for a later code-review phase to consume. It solves the problem of spending expensive review compute indiscriminately: designed to run on a lightweight model such as Haiku, it does the cheap classification and surface-level checks first so the heavier review only runs where it matters.

Using the gh and git CLIs read-only, it fetches PR metadata and the changed-file list, then conditionally pulls the diff (skipping the full diff and analyzing by filename and PR description alone when 15 or more files changed). It classifies files as added, modified, or deleted, detects languages and frameworks, and determines change categories (authorization, database, PostgreSQL RLS, API, tests, config, skill changes). Based on those categories it decides which reference guides a reviewer will need (TypeScript best practices, general authorization review, Postgres RLS review, skill review, and always the GitHub PR-review-actions guide). It detects minor/suggestion surface issues such as use of the any type, var, empty interfaces, magic numbers, naming-convention violations, and missing tests, and writes a bounded diff summary and focus areas. An incremental mode reads a prior .pr-review-state.json, uses git diff to find files changed since the last reviewed commit, carries over unchanged files' issues, re-checks only changed files, and records resolved issues, with an explicit guard to fall back to a full triage if that state file is missing or malformed.

The target users are teams running AI-assisted code review in CI who want to reduce token cost and structure the review pipeline. It performs only read-only repository inspection and JSON output; it makes no changes to the code or repository.