Skip to content

design:os · docs

Live-canvas verification checklist (owner-manual)

The one verification layer CI cannot reach: everything below needs a real Figma Desktop with the plugin loaded, because it either drives the canvas or opens an editor (FigJam / Slides) a headless test can’t. Consolidated across absorption phases #10/#13/#26/#28 so it runs in a single connected session instead of four scattered PR bodies.

Setup (once)

Figma hotloads code.js/ui.html from disk — a normal rebuild propagates to the running plugin with no re-import. BUT manifest.json is read only at plugin load, and the absorption phases changed it (editorType widened to ["figma","figjam","slides"]). So this one time you must re-register so Figma picks up the widened manifest AND its new path (the plugin moved out of the monorepo; the old dev entry points at the deleted location):

  1. Figma Desktop → Plugins → Development → re-import the manifest from /Users/jang/Products/design-os-figma-plugin/plugin/manifest.json. (Only needed because the manifest itself changed — code-only changes after this hotload as usual.)
  2. Confirm the panel opens and reads “design:os by JANG”, status shows Connected.
  3. figma-agent status from a terminal → plugin.editorType: "figma", no errors. If it still reports the old manifest (no figjam/slides), Figma didn’t pick up the re-import — remove the dev plugin entry and add it again.

A. Design file (regression — MUST be unchanged)

The absorption phases widened editorType to ["figma","figjam","slides"]. This section proves the existing design-file path is byte-unaffected — the one real risk of the manifest change.

  • ☐ Open a normal Figma Design file. Panel connects, editorType: "figma".
  • ☐ Run a known-good command (e.g. ui figma reconcile --apply on a bound file, or a small figma-agent exec-js that reads a node) — behaves exactly as before the phases.
  • ☐ ui.componentSet(...) / ui.slot.* still work in a design file (the reverse guard must NOT wrongly refuse here).
  • ☐ ui.figjam.* and ui.slides.* REFUSE in the design file with an actionable message naming the editor they need (forward guard fires).
  • ☐ Gap-fill / sync still capture design-file edits normally (no regression from the widen).

B. FigJam (new surface — phase-03)

  • ☐ Open a FigJam board. Panel connects, figma-agent status → editorType: "figjam".
  • ☐ ui.figjam.sticky(...) creates a sticky; ui.figjam.connector(...), .shape, .table, .codeBlock, .section each create their node; reply names what was made (no fabrication).
  • ☐ ui.figjam.board() reads the board; ui.figjam.connections() lists edges — an unresolved endpoint shows unresolved: true, never a dropped edge.
  • ☐ ui.figjam.arrange(...) with one bad id → the bad id lands in skipped[], the rest arrange.
  • ☐ A design-only command (ui.componentSet, ui.slot.*) REFUSES in FigJam, naming the editor.
  • ☐ Make a manual edit (move a sticky) → it appears in the change feed with nodeType: "STICKY" verbatim (honest-by-construction — never a fabricated design-file fact).

C. Slides (new surface — phase-04)

  • ☐ Open a Figma Slides deck. Panel connects, editorType: "slides".
  • ☐ ui.slides.list() / .grid() / .content(slideId) read the deck honestly.
  • ☐ ui.slides.background(slideId, color) sets the slide fill directly (no rectangle appears on the slide — it uses setFillsAsync, method slide-fill).
  • ☐ The single-slide-view trap (the load-bearing correctness test): switch to single-slide view, focus slide X, then ui.slides.addText(targetSlide=Y, ...) → the text lands on Y, NOT the focused slide X. (This is the silent-misattribution guard; if it lands on X, that’s a real bug to report.)
  • ☐ ui.slides.reorder(...) with a grid that drops or duplicates a slide → REFUSES, naming the missing/duplicate id (never silently drops a slide).
  • ☐ ui.slides.setTransition(...) applies; ui.slides.transition(slideId) reads it back.
  • ☐ Undo honesty: run an exec-js that navigates (ui.slides.focus/viewMode) then throws → rolledBack: true covers the CONTENT, but you are left on a different slide (navigation is not undoable — this is documented, verify the message is honest about it).

Report back

For each section: pass / fail-with-detail. Any failure in A is a regression (blocks — tell the team). Failures in B/C are new-surface bugs (file as board issues). All-pass on A + a working sticky/slide in B/C = the cross-editor absorption is live-verified.