Browser telemetry

Availability: published package on feat/browser-telemetry; not connected to the live player. This guide uses the package in its own iframe integration or isolated test harness.

Build the package

From the browser branch's apps/web directory:

npm ci
npm run typecheck:browser-telemetry
npm run test:browser-telemetry
VAPOUR_ENV=staging \
FIREBASE_PROJECT_ID=YOUR-STAGING-PROJECT-ID \
VAPOUR_DEVTOOLS_ENABLED=true \
VAPOUR_STAGING_PLAYER_ORIGIN=https://YOUR-STAGING-GAME-ORIGIN \
npm run build:browser-telemetry

Use the exact game iframe origin, not the surrounding dashboard origin. The build requires an HTTPS origin and the known staging project. Output is .staging/browser-telemetry/, outside Next's public/ and production output. These artifacts are not automatically served by the player.

Start a capture through the trusted loader

The integration guide describes the owner grant and memory hook. Default TTL is 300 seconds, hard-capped at 300. JS heap/document measurement and function profiling are both off by default.

const session = await telemetry.startBrowserTelemetry({
  toggledOn: true,
  authorization: verifiedOwnerGrant,
  options: {
    getMemory: () => actualGameMemory,
    ttlSeconds: 300,
    overlay: true,
    onMetrics: consumeOnThisMachine,
  },
});

These variable names are explicit loader inputs, not discovered globals. An absent memory accessor leaves Wasm metrics unavailable rather than measuring some other document.

Read the metrics correctly

• FPS/frame times: rAF presentation cadence, not GPU render time.
• Lows: inverse mean duration of the slowest 1%/0.1% of retained frames in the relevant sample window; the 1Hz metric buckets and whole-trace summary differ.
• Spikes: greater than twice the preceding rolling median, after warm-up.
• Main-thread busy: lower-bound long-task/long-animation-frame occupancy, never whole-process CPU percent. Only one observer source is used.
• Wasm memory: memory.buffer.byteLength, sampled at 1Hz. Several grows can coalesce into one observation.
• JS/document memory: optional and sampled at most every 30 seconds. Chromium heap estimates and user-agent document/worker estimates are labeled differently.
• Functions: opt-in JS Self-Profiling, interval at least 10ms, 10,000-sample cap. Policy denial/unsupported browsers are reported; no substitute samples are invented.

The overlay can be changed with session.setOverlay(true|false). No per-frame allocations are made in the collector's rAF callback; buffers are allocated on capture start and capped. Hidden-tab gaps are omitted rather than called spikes.

Asset records: the browser collector does not emit asset records, and its vendored schema predates them. See Find expensive assets and the two schema pins.

Export, analyze, stop

const trace = await session.dump();
const folded = await session.dumpFolded();
const offline = await session.analyze({ offline: true });
telemetry.stopBrowserTelemetry();

Save exports before off. Dumping function data ends the optional sampler; it is not silently restarted. Off/TTL/pagehide clears buffers, observers, timers, channels and the overlay, and aborts pending provider analysis. Another capture requires a document reload.