hermes-agent/evals/desktop-pane-fixture/README.md

Saved-pane fixture import-delay control

From a root npm ci installation, run in apps/desktop:

npx vitest run --config ../../evals/desktop-pane-fixture/vitest.config.mts --reporter verbose
npx vitest run --project ui src/store/session-pane-focus.test.ts --sequence.shuffle.tests --sequence.seed=12345 --reporter verbose

The Vite plugin appends an 11-second asynchronous wait to the real pane-tree store's evaluation. It does not mock pane functions, inspect source text in an assertion, lower a deadline, or alter the two test bodies. The UI test deadline remains 15 seconds and the default hook deadline remains 10 seconds. This is a scheduling-controlled reproduction of import latency, not a claim that a local machine reproduced the original CI load unassisted. It is opt-in, not a slow CI test.

For A/B, apply this eval directory to the parent revision and run the same command, then run on the fixture fix. Base evaluates the graph inside each beforeEach; fixed evaluates it once during test collection. The delay marker appears twice on base and once on fixed. Expected: base fails with Hook timed out in 10000ms; fixed runs both behavioral assertions successfully.

Measured on the initial base fb3446a281e:

Control	Result
Base + evaluation delay	2 hook timeouts, 20.04s test phase
Fixed + identical delay	2 passed, 5ms test phase; 11.84s import phase
Fixed normal and reversed test order (seed 12345)	2 passed in each order
Store directory, 8 workers	123 files / 1491 tests passed
Complete UI project, 8 workers	753 files / 7355 assertions passed, but exit 1 from an unrelated Radix focus-scope timer's wrong-realm Event in statusbar-visibility.test.tsx

Mutation controls, applied separately to session-states.ts and then restored:

• Remove revealTreePane(paneId) in focusOpenSession: the overlay re-adoption case fails (null instead of canonical-chat); the miss case passes.
• Make focusWorkspaceOwnerSessionTile return the stored id regardless of focusOpenSession's result: the miss case fails (canonical-chat instead of null); the overlay case passes.

The fixture uses the real registry, tree watcher, pane mirror, overlay, tile store, and focus helpers. Install app-lifetime watchers once per isolated Vitest file graph. Discard the fixture tile through the real action to clear both the reactive source and its persisted in-memory bucket; reset layout, focus, preset, selection, read baseline, and localStorage between cases. The mirror observes the discard and unregisters its contribution. No production API needs a test-only reset or disposer.

If the host has exhausted inotify watches, prefix commands with CHOKIDAR_USEPOLLING=true; this changes file watching, not test deadlines. The recorded store/full-suite and final focused controls used that workaround.