Skip to content

ADR-010: E2E UI Smoke Test with the Real WASM Engine

Status: Accepted Date: 2026-07-05

Context

ADR-009 added orca-wasm/scripts/smoke-test.mjs, a Node script that loads slicer.js/slicer.wasm directly and calls _orc_init/_orc_slice to catch an engine that compiles but traps on a real slice. It only runs in build-wasm.yml, which is path-filtered to orca-wasm/** — a normal UI PR never triggers it, and it never touches the actual app path a real user exercises: FileUpload → App.tsx's queue state machine → worker-singleton.ts → slicer.worker.ts's message protocol → the WASM module. A regression in that glue code (a broken postMessage payload, a worker that never spawns, a UI state that gets stuck on loading-wasm) would compile fine, pass the Node-level smoke test, and still ship broken.

Decision

Add a Playwright E2E test (e2e/slice.spec.ts) that drives the real browser UI end-to-end against a real compiled engine:

  1. Load the app, upload the real Voron Design Cube v7 STL (e2e/fixtures/voron-design-cube-v7.stl, vendored, GPL-3.0 — see below and NOTICE.md) — the exact model that historically triggered two production crashes in the Arachne wall generator (orca-wasm/patches/apply.py sections 8/8c–8f; also referenced in ADR-009's context). Using this real, previously-crash-triggering mesh is a stronger regression guard for the UI path than a trivial synthetic primitive would be.
  2. Navigate to the Slice tab and click "Slice All".
  3. Wait for the queue item to reach Done (cold engine load + a real slice, generous timeout) and assert the download button appears.
  4. Assert no browser console errors occurred during the whole flow.

data-testid attributes were added to the file input, tab buttons, slice button, queue item status, and download button (src/components/FileUpload.tsx, src/App.tsx) so the test doesn't depend on copy text.

Wired into a new .github/workflows/e2e-smoke.yml, triggered on every pull_request (no path filter, unlike build-wasm.yml) — a UI-only change is exactly what this test exists to catch. It downloads the latest published engine release via npm run setup (the same script local dev uses) rather than building from source: engine-source correctness is already gated by build-wasm.yml's own smoke test before anything is ever published, so this test only needs a real, working binary to drive the UI against, not the one the current PR might be building.

Runs against npm run dev (not vite preview) because the required COOP/COEP headers (vite.config.ts) are configured under server, not preview.

Why vendoring this file is safe (unlike ADR-009's synthetic mesh)

ADR-009 deliberately avoided vendoring a third-party STL for orca-wasm/scripts/smoke-test.mjs, to "sidestep any question about redistributing someone else's model." For this E2E test we vendor the real Voron Design Cube v7 (e2e/fixtures/voron-design-cube-v7.stl) instead, because the actual blocker was provenance/attribution, not license incompatibility:

  • Voron Design's hardware repos (including VoronDesign/Voron-2, which hosts this file) are licensed GPL-3.0.
  • This repository is AGPL-3.0-or-later. GPLv3 and AGPLv3 both contain a clause (§13 in each) that the FSF added specifically so GPLv3-covered and AGPLv3-covered works can be combined — they are not in conflict.
  • Attribution is recorded in NOTICE.md (source URL pinned to the commit the file was fetched from, license, copyright), satisfying GPL-3.0's notice requirement.

orca-wasm/scripts/smoke-test.mjs itself is unchanged and still defaults to a synthetic mesh — this doesn't reopen or reverse ADR-009, it only settles that vendoring a compatible-licensed, properly-attributed fixture is different from vendoring an unknown/incompatible one.

Consequences

  • Positive: Catches breakage in the worker/UI integration layer that the Node-level smoke test structurally cannot see, on every PR rather than only ones touching orca-wasm/**.
  • Positive: Runs against a real engine binary, not a mock — closer to what a user's browser actually does.
  • Negative / accepted scope: Tests against the latest published engine release, not a WASM change made in the same PR (that PR's build-wasm.yml run validates the engine itself; the two checks are complementary, not redundant).
  • Negative: Adds a Chromium download + a dev-server boot + one slice to every PR's CI — a few minutes, not the multi-hour cost of build-wasm.yml.
  • Negative: Adds a ~170 KB third-party binary (voron-design-cube-v7.stl) to the repository, with an attribution obligation tracked in NOTICE.md.