opencode/packages/session-ui/performance/markdown-lifetime

Completed Markdown Job Lifetime

This manual browser benchmark bundles the production Markdown component, cache, sanitizer, and worker with Vite's production mode. It does not connect to an OpenCode server or use user data. Run from packages/session-ui.

Set MARKDOWN_BUILD_DIR and MARKDOWN_RESULTS_DIR to absolute artifact directories:

bun --bun x vite build --config performance/markdown-lifetime/vite.config.ts
bun --bun x playwright test --config performance/markdown-lifetime/playwright.config.ts --repeat-each 20

The three independent scenarios render one deterministic long completed answer with 36 TypeScript fences, explanatory prose, links, inline code, and tables:

• mounted: keep the consumer mounted through result delivery.
• leave: unmount the last consumer before delivering its parse result and render a short completed answer in the current destination.
• shared: unmount one of two consumers with the same cache key, preserving the other consumer's result.

The fixture intercepts only the real worker's message delivery. Admission and worker-result arrival enable the Continue button. Continue disposes the departing consumer, then delivers the held result. No parser work is replaced. This makes the unmount/result ordering deterministic without wall-clock delays. Cold worker startup and parsing happen before the measured interval; this benchmark measures main-thread work after the worker has produced a result, not worker CPU savings.

Readiness is the production component's data-markdown-ready state, observed by a MutationObserver. A MessageChannel task marks completion of the released result's promise microtasks, including cache postprocessing. CDP reports renderer task/script duration and renderer JS heap, not total browser or desktop RAM. The cache-character metric confirms whether the obsolete result was sanitized and stored. It is a mechanism check, not the performance result by itself.

The initial frozen baseline called this field cacheBytes, although it counted HTML characters. The runner normalizes that frozen-build field to cacheChars; the renderer probe operation is unchanged.

Use MARKDOWN_ASSERT_DISPOSAL=1 to assert that the abandoned result is not cached. Use MARKDOWN_RETAINED=1 in separate runs for a post-GC retention check; do not mix these samples with natural-GC heap measurements. OPENCODE_PERFORMANCE_TRACE_DIR enables the repository's Chrome trace collector. MARKDOWN_SCREENSHOT optionally captures each final scenario after timing. All artifacts belong outside Git.

Freeze the baseline build before changing production source and reuse the same fixture, runtime, viewport, and readiness contracts for the candidate. Do not use machine-dependent latency thresholds. The scenarios test real components, but do not reproduce full session navigation or establish Electron process-memory gains.