opencode/packages/session-ui/performance/highlighting

Diff Highlighting

Manual Chromium benchmark of the production File component, normalize, Pierre pool, and bundled workers. No OpenCode server, session, or stored data is used.

From packages/session-ui, set HIGHLIGHT_BUNDLE to an external artifact directory, then run:

bun performance/highlighting/build.ts
bun x playwright test --config performance/highlighting/playwright.config.ts --repeat-each 20

Each isolated page measures a cold mount, unmount/remount of the identical patch, and an edited patch under the same filename. The fixtures contain 120 or 1,200 TypeScript route handlers with one changed line per ten handlers. Exact byte and line counts, bundle hash, source revision, and individual samples are reported through the existing performance reporter.

readyMs ends when the viewer has called onRendered, its pool has settled, and onPostRender has committed the final worker result containing the expected edit. firstReadyMs records the earlier production callback, which can represent plain first paint. Worker request counts and round-trip time describe the mechanism, not CPU time or desktop memory. Full reconstructed content is checked separately after timing. No forced GC or machine-dependent thresholds are used.

HIGHLIGHT_CORRECTNESS=1 runs the deterministic *.correctness.ts checks against the same fixture instead of the timed benchmarks: cache reuse across remounts, invalidation on content, theme, and worker option changes, and plain rendering of large diffs with complete reconstructed content.

Optional output settings: HIGHLIGHT_RESULTS, HIGHLIGHT_SCREENSHOTS, OPENCODE_PERFORMANCE_RUN_ID, and OPENCODE_PERFORMANCE_TRACE_DIR (separate diagnostic runs, not clean timing). HIGHLIGHT_PORT defaults to 4793. Preserve the baseline build before product edits, then point HIGHLIGHT_BUNDLE at either frozen build to compare the same workload without rebuilding.

HIGHLIGHT_RETENTION=1 enables a separate post-unmount, forced-GC renderer-isolate heap diagnostic. Do not use that run for timing or describe it as desktop RAM; it excludes worker heaps and native/browser-process memory.