Team rollout

How a design-conscious team goes from zero to a shared, enforced design gate. Every step is deterministic: findings cite file:line, re-run identically, and no LLM sits in the enforcement path.

Day 1 — one person runs init

npm i -g @memi-design/cli
cd your-app
memi init --team
git add memoire.policy.json .memoire/baseline.json .gitignore
git commit -m "chore: memi design gate (policy + accepted baseline)"

What that did, honestly:

• memoire.policy.json — the committed rulebook (preset memi-recommended). Thresholds and gate severity live in the repo, not in someone's shell. Every report stamps the policy hash, so a score is only ever compared against runs under the same rules.
• .memoire/baseline.json — every finding that existed today, accepted loudly as known debt. From now on only new findings gate. Suppressed counts stay visible in every report; a baseline hides nothing, it just stops old debt from blocking new work.
• .gitignore — a managed block keeps .memoire/* workspace state local while sharing baseline.json. If you had a plain .memoire/ ignore line, init warns you: git cannot re-include a file inside an ignored directory, so the old line must go.
• An agent kit — .agents/skills/ (or --kit claude-code, cursor, …) so every teammate's coding agent uses the same design tooling. --kit none skips it.

Day 1 — everyone else

npm i -g @memi-design/cli
memi init --team        # detects the committed policy + baseline, installs the kit, changes nothing shared
memi doctor             # Team gate section: policy ✓ baseline ✓ gitignore ✓

Teammate runs are idempotent: existing policy and baseline are reported, never overwritten. If the baseline was accepted under a different policy hash, init says so instead of pretending the numbers are comparable.

Week 1 — wire the CI gate

See CI_RECIPES.md. Short version:

# .github/workflows/design.yml
jobs:
  design:
    runs-on: ubuntu-latest
    permissions:
      contents: read
      security-events: write
    steps:
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
        with: { fetch-depth: 0 }
      - uses: memi-design/memi@5fcbf39e1255af0c14c5a17ba6bde8cf1206e525 # v2.7.9

PRs now get SARIF annotations on the exact changed lines, a step-summary scorecard, and a design-health report artifact. Whole-tree stats are always computed (thresholds stay valid); the PR is only blamed for files it touched.

Ongoing — burn the debt down

memi baseline status     # suppressed / new-active / stale-safe-to-prune
memi diagnose --trend    # score over time, same-policy runs only
memi report --badge      # shareable design-health.html + SVG badge

• Fix accepted debt → the fingerprint goes stale → memi baseline status lists it as prunable → re-accept to shrink the file.
• Tighten the gate when ready: switch the preset to strict in memoire.policy.json (fails on medium, zero regression budget, stale research citations fail audits). The policy hash changes, so trend lines honestly restart.

What this does NOT do

• It does not grade taste. It gates the mechanical layer — token discipline, contrast, scale drift, accessibility signals — the part that is checkable, so review time goes to the part that isn't.
• It does not assess what it cannot see. Reports mark tenets/dimensions as "not-assessed" rather than inventing a score; provenance on every finding says how it was produced (static-scan today).
• It does not compare incomparable runs. Different policy hash or a scoped scan → trend and regression checks say "not comparable" instead of fabricating a delta.