These Playwright tests run against a production Vite build and seed the browser mock backend with deterministic localStorage fixtures. They measure interaction settle time, first visible page content, visible image load/decode, source-reported page commits, invoke count/bytes, Long Tasks, mounted conversation rows, and total DOM nodes without requiring a Tauri database.

Run the complete release-gate suite (the package script sets PERF_ENFORCE=1 on every platform):

pnpm test:perf

The direct Playwright command is metrics-only unless PERF_ENFORCE=1 is set. In metrics-only mode, tests still enforce correctness: the page must render without an uncaught exception, the native/virtual branch must match the expanded row count, virtualization must keep the mounted row count bounded, the 159/160 row boundary must keep the first-screen Ant UI geometry within 1 px and the exact pixel difference within 0.5%, Chat/Drawing scroll restoration must stay within 1 px, and 100 warm Chat/Drawing cycles must not retain listeners, intervals, or cycle-proportional DOM nodes. Warm page samples also reject repeat conversation, provider, message-version, or drawing-history IPC loads. Metrics are emitted as per-test JSON attachments and as test-results/performance-results.json.

Collect metrics without wall-clock/heap budgets on an unstable shared runner:

pnpm exec playwright test --config=playwright.perf.config.ts

The enforced budgets are 50 ms for an active-conversation switch, 100 ms for a warm Chat/Drawing switch, and 150 ms for a Roles transition. Module navigation records every Long Task with its start/end offsets and commit overlap, while the explicit 1,000-row sidebar scroll gate rejects Long Tasks of 50 ms or more. The suite also requires React commit P95 to stay within 50 ms and GC-retained Chromium JS heap growth after 100 warm cycles to stay below 10%. Do not use the release-gate command on shared or heavily loaded CI hosts.

Navigation duration ends only after the target page has committed and its active content selector is present. The probe deliberately avoids geometry/style reads in that commit frame so measurement itself cannot force layout of the retained Activity tree.

pageCommitMs, reactCommitMs, and invoke byte counts come from the application's conditional window.__AQBOT_PERF__ instrumentation. Unsupported CDP heap collection is recorded as null rather than treated as a successful measurement.

To reuse a separately started production preview, set PERF_BASE_URL; otherwise the configuration always runs pnpm build followed by pnpm preview:

PERF_BASE_URL=http://127.0.0.1:4173 \
  pnpm test:perf

The default project uses Playwright's bundled Chromium. A machine that has not run pnpm exec playwright install chromium can opt into an installed Chrome without changing the checked-in configuration:

PERF_BROWSER_CHANNEL=chrome \
  pnpm test:perf