Rams Design Review — the judge in your agent loop

Score UI changes 0–100, emit structured findings and git-applyable patches, and gate merges on critical design defects. Rams judges; your tools repair.

PR gate

- uses: actions/checkout@v4
  with: { fetch-depth: 0 }
- uses: rams-design/rams-action@v1
  with:
    api-key: ${{ secrets.RAMS_API_KEY }}   # free at rams.ai/app/settings
    fail-on: critical                       # or score<80, or never

The loop (scheduled, agents repair)

on:
  schedule: [{ cron: '0 9 * * 1,4' }]      # Mon + Thu
jobs:
  design:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with: { fetch-depth: 0 }
      - id: rams
        uses: rams-design/rams-action@v1
        with:
          api-key: ${{ secrets.RAMS_API_KEY }}
          mode: since-last-run              # reviews only what changed since the last run
          fail-on: never
      # hand the findings to YOUR agent step — Claude Code, Codex, anything:
      #   rams-review/issues.json  — structured findings
      #   rams-review/patches/     — git-applyable fixes
      # then commit .rams/last-reviewed-ref to advance the window

Outputs

output	meaning
score	0–100 (any confirmed critical caps at 59)
critical-count	confirmed criticals
issues-json	path to structured findings
patches-dir	one .patch per unambiguous fix — git apply ready

Quota: each run consumes one review (shared with the GitHub App and MCP). Engine: the same 291-rule judge as every Rams surface — rams.ai.