Files
paperclip-docs/scripts/screenshots/README.md
Aron PrinsandClaude Opus 4.8 ae046f058b docs(screenshots): refresh all 98 stale shots + harden pipeline isolation
Recapture every screenshot flagged in SCREENSHOTS_PENDING.md (98 entries; 233
PNGs updated across light/dark) against parent `master`, and stamp registry.json
with the new captured_sha. Mark SCREENSHOTS_PENDING.md as caught up.

Also fix two isolation gaps that made `screenshots:refresh` unusable while a real
local Paperclip is running:

- Free embedded-Postgres port: onboard hard-codes port 54329 (no env/flag
  override; the port comes only from config.json), which collides with a real
  local instance. run.mjs now generates the config in external-DB mode (keeping
  the scratch data dir pristine) and rewrites database.embeddedPostgresPort to a
  scanned free port before boot.
- Parent .env neutralized: the server loads PARENT_REPO/.env from its cwd via
  dotenv (override: false), which leaked DATABASE_URL (real external DB on 5432)
  and SERVE_UI=false into the "isolated" instance. instanceEnv() now pins
  DATABASE_URL="", DATABASE_MIGRATION_URL="" and SERVE_UI="true" so these win.

seed-execution-workspace.mjs now throws instead of falling back to 54329, so the
direct-DB seed can never connect to a real database.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
2026-06-27 11:36:27 +02:00

6.6 KiB
Raw Permalink Blame History

Screenshot pipeline

Automated capture of light + dark UI screenshots at 1440×900 @2x for use in the docs. Output PNGs land under docs/user-guides/screenshots/{light,dark}/ and are reviewed in a PR — they are never auto-pushed.

Prerequisites

  1. Parent repo present. The pipeline spins up a real Paperclip instance from source. Set PAPERCLIP_REPO to its absolute path, or ensure it lives at the default location ~/Documents/PaperclipAI/paperclip (the default resolved by config.mjs). See paperclip-localdev-runbook.md conceptually — the runner follows the same local_trusted / private / loopback isolation pattern.
  2. Playwright Chromium installed. Run once after npm install:
    npx playwright install chromium
    
  3. pnpm available. The runner starts the parent's onboarding entrypoint via pnpm paperclipai onboard --yes --run.

One-command flow

npm run screenshots:refresh

What happens under the hood:

  1. Isolated instance. run.mjs spawns pnpm paperclipai onboard --yes --run inside the parent repo with a throw-away HOME / PAPERCLIP_HOME under os.tmpdir(), PORT=3197, PAPERCLIP_BIND=loopback, PAPERCLIP_DEPLOYMENT_MODE=local_trusted, and PAPERCLIP_DEPLOYMENT_EXPOSURE=private. The child process receives only a small allowlist of runtime environment variables, so real database URLs, provider keys, and GitHub tokens are not forwarded.
    • Parent .env neutralized. The server also loads a .env from its working directory (the parent repo) via dotenv with override: false, which would otherwise re-inject the developer's real settings and defeat the allowlist. instanceEnv() therefore explicitly pins the keys that matter: DATABASE_URL="" (and DATABASE_MIGRATION_URL="") so the instance uses its own embedded Postgres instead of the parent's external database, and SERVE_UI="true" so the UI is served even if the parent's .env set SERVE_UI=false. Because dotenv runs with override: false, these explicit values win over the parent .env.
    • Free embedded-Postgres port. Before that, run.mjs runs a short config-generation pass (onboard --yes pointed at an unreachable external DATABASE_URL, so it writes a config without initializing the embedded cluster) and rewrites database.embeddedPostgresPort to a guaranteed-free port discovered by scanning upward from 54329. This matters because onboard hard-codes embedded Postgres to 54329 with no env/flag override, and that's the same default a developer's real local Paperclip uses. Pinning a free port means a screenshot run started while your real instance is up can never collide with — or, worse, connect into and seed — your real database. (The generation pass must stay in external-DB mode: if it ever ran initdb on the scratch data dir, a half-finished cluster would make the real run skip init and fail with role "paperclip" does not exist.) The direct-DB workspace seed (seed-execution-workspace.mjs) reads the live port back from the instance's postmaster.pid / config.json and refuses to guess 54329.
  2. Health check. Polls http://127.0.0.1:3197/api/health until 200 (up to 120 s).
  3. Seed. seed.mjs creates a minimal demo company ("Acme Robotics") plus a manager agent, a regular agent, a project, a goal, a routine, an issue, and a workspace. The company's issuePrefix is auto-derived server-side (first three uppercase letters of the name → ACM) and cannot be set via REST, so the real prefix is captured and written to .seed-ids.json alongside the entity IDs (gitignored). routes.mjs reads companyPrefix from there, so routes always resolve to whatever the server assigned. Idempotent — re-running looks up the existing company first.
  4. Sync registry. sync-registry.mjs merges route and depends_on metadata from CAPTURE_TARGETS into registry.json without overwriting other fields.
  5. Capture. capture.mjs opens Chromium via @playwright/test, navigates each route in both light and dark themes (1440×900, deviceScaleFactor 2), and saves PNGs to docs/user-guides/screenshots/{light,dark}/<name>.png. The board theme is seeded via both colorScheme emulation and a localStorage init script so the UI picks it up consistently.
  6. Stamp. After capture, registry.json entries are updated with captured_sha (parent git HEAD) and captured_against (branch/tag name).
  7. Teardown. The server child process is killed and the scratch home is removed (--keep skips the cleanup for debugging).

Overhaul mode

After a large UI/UX redesign — when every screen has changed — recapture everything in one pass:

npm run screenshots:refresh:all

This passes --all to capture.mjs, which ignores staleness and re-shoots every entry in CAPTURE_TARGETS.

Individual scripts

Script Command What it does
seed.mjs npm run screenshots:seed Seeds demo data against a running instance at BASE_URL. Writes .seed-ids.json.
capture.mjs npm run screenshots:capture Captures screenshots against a running instance. Flags: --all, --only <substr>, --theme <light|dark|both>, --stale <file,...>, --base-url <url>, --keep.
sync-registry.mjs npm run screenshots:sync-registry Syncs route + depends_on from routes.mjs CAPTURE_TARGETS into registry.json. Safe to run standalone.
run.mjs npm run screenshots:refresh Full orchestrator: starts isolated instance → seeds → syncs registry → captures → tears down. Forwards --all, --only, --theme to capture.mjs.

Route map

The full list of screenshots to capture, their routes, and their depends_on mappings is maintained in docs/user-guides/screenshots/SCREENSHOTS_TODO.md. routes.mjs is the machine-readable version of that same list — keep them in sync when adding new pages.

Rules

  • Never run against live data. The pipeline always talks to the seeded demo instance on loopback port 3197. PAPERCLIP_BIND=loopback ensures no external traffic.
  • Safe to run alongside your real local Paperclip. The screenshot instance pins its embedded Postgres to a free port (see "Isolated instance" above), so you do not need to stop a real instance first. It will not touch the real database on 54329.
  • Output goes to a PR, never auto-pushed. After screenshots:refresh, open a PR with the changed PNGs for human review. The pipeline does not push.
  • Always capture both themes. The renderer swaps variants at runtime — providing only one breaks dark- or light-mode users.