Weekly scheduled cloud agent for the Figma §22 full-library audit

Context

The §22 audit (9 categories: accessibility, page display, variables, Text Styles, states, variants, in-page documentation, links, code↔Figma parity) has existed since 2026-07-08, but only ever triggers on a newly created/modified page, or on explicit human request ("audit", "check everything"). It has never run against the *existing* library as a whole.

That gap caused a real incident: on 2026-07-09, 10 of the library's 11 Text Styles had zero Figma Variables bound to their fontSize/fontFamily/ fontWeight/lineHeight — literal values only, invisible unless someone opened each style and inspected its bindings. It was caught by a human manually comparing styles by eye, not by any systematic check, because no audit was running on pages nobody had recently touched.

Decision

A weekly scheduled cloud agent ("routine"), created via the schedule skill/RemoteTrigger API, runs the full §22 checklist against the *entire* Figma library every Monday at 06:00 America/Toronto (10:00 UTC, cron 0 10 * * 1):

• Trigger ID: trig_015wTFo2gJNztALS2bvFZ54d
• Scope: every page and component/ComponentSet in the file

(uXgPVB6cMLwAPqSwoa0dGq), not a subset — the exact gap that let the 2026-07-09 incident go undetected

• Environment: the only cloud environment available on this account

("Portfolio web", env_01GrYNDKPKbnLc4HE3dvaftU) — not a deliberate choice, there was no alternative to pick from

• Reporting: silent if clean; a GitHub issue (label figma-audit) if any

violation is found, grouped by the 9 categories with exact node/style names; a GitHub issue is also filed if the run fails outright (Figma unreachable, MCP error), so a failed run can never be mistaken for a clean one

• Read-only: the routine never modifies the repo or the Figma file — audit

and report only, exactly like the existing playwright-reminder.yml pattern for visual regressions

Rationale

A recurring, unattended check is the only way to catch drift on pages nobody is actively touching — the 2026-07-09 incident's whole failure mode was that nothing was looking at *already-shipped* library content. Weekly matches how the library actually changes: in bursts around specific chantiers, not continuously, so daily would be noisy and monthly would let drift sit for too long between publications.

Rejected alternatives

• A GitHub Actions workflow, like playwright.yml. Rejected: §22's checks

run inside the Figma Plugin API sandbox (use_figma/Figma MCP tools) against a live Figma session — there is no headless, CI-runnable equivalent the way Playwright drives a browser. A cloud agent with the Figma MCP connector attached is the only mechanism available that can actually execute this.

• Trigger only before each Community File publication. Rejected (per

discussion with the Design System Lead): no publication is currently scheduled (the Community File chantier is still Backlog), so a publication-triggered-only cadence would mean the audit effectively never runs until then — reintroducing the exact "ad hoc" gap this ticket exists to close.

• Always report, even when clean. Rejected: the Design System Lead chose

issue-only-on-violation, matching playwright-reminder.yml's existing philosophy of only surfacing actionable problems.

Consequences

• Drift on any page, however old or untouched, surfaces within at most a week

instead of only by chance human discovery.

• The routine's prompt is self-contained (the cloud session starts with no

memory of this conversation) and points at .claude/instructions/figma-components.md §22 as the single source of truth for what counts as a violation — if §22's checklist changes, the routine's behavior changes with it automatically, with no separate copy to keep in sync.

• First real-world validation (does gh issue create actually work from this

cloud environment, does the Figma MCP connector reliably enumerate the whole library) only happens at the first scheduled run (2026-07-27) — this ADR will need a follow-up note if that run reveals a configuration gap.