[ashtonanderson]

Adds PostHog events that show what real visitors wait for on the analysis page, and when it dies, so that #334 and later changes can be judged on real devices rather than emulated ones.

Events

• analysis_engine_load: one per visit, for the first position: Maia, the chart, Stockfish's first depth, d12, d18, lines settled, search done.
• analysis_engine_steps: a summary of the moves stepped through, sent when the page is hidden or left, and every 30 moves: medians, and how many reached each stage.
• analysis_page_ended_unexpectedly: the last analysis page in this tab was never left, as when a phone runs out of memory, with whether it was on screen.
• stockfish_engine_ready and _failed: now also give engine and memory_retries.
• maia_engine_ready and _failed: new here, with restarts. The tracking functions existed but were never called.

All the timing events are split by engine, engine_build, is_phone, cores, release and PostHog's own $os and $browser.

LOADING.md, round 11, explains how to read them and has SQL for PostHog's editor: crash rate, what visitors wait for, and Stockfish before and after #334.

Things to know

• Before and after Stockfish 19's small net, and the best line first #334: only stockfish_engine_ready, stockfish_engine_failed, $web_vitals and $exception existed before it went live. The new events have no "before", so they compare later changes against it.
• Crashes are undercounted: a crash is reported only when the visitor comes back to an analysis page in that tab.
• release: it relies on Vercel exposing NEXT_PUBLIC_VERCEL_GIT_COMMIT_SHA. That isn't confirmed; locally it's empty.
• Not checked yet: events reaching the real PostHog project. I have no PostHog access, so please look at Activity on this preview.

Testing

• Unit tests: npm test (85), npx tsc --noEmit and npm run lint pass.
• In a browser, with PostHog's requests captured: the events and values arrive. A normal reload reports no crash, and a page that loses its exit handlers is reported on the next load with was_visible true. This run found and fixed a bug where every normal reload would have counted as a crash.
• Speed, 6 paired loads per device against the same code without the events: no slowdown within the noise. Test browsers' events are dropped by PostHog, so sending itself isn't counted.
• Not run on a real iPhone.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
maia-platform-frontend	Ready	Preview	Nov 5, 2026 4:10am UTC