Timeline Text Projection Benchmark

Run from packages/app, against a production bundle. No OpenCode server is needed. The fixture owns its HTTP event source and mocks all other API responses.

bun run build -- --config e2e/performance/timeline/projection.vite.config.ts
PLAYWRIGHT_BUILD=1 playwright test --config e2e/performance/timeline/projection.playwright.config.ts --repeat-each=20 --workers=1 --retries=0 --reporter=line

Set PLAYWRIGHT_PORT to an unused local port. Set PROJECTION_BUNDLE to preview a previously frozen bundle, PROJECTION_OUTPUT for test artifacts, and PROJECTION_REVISION to label results. Keep all three workloads separate: 40 historical user/assistant pairs, 320 pairs, and 320 pairs with six additional completed read tools per assistant. The existing mixed fixture includes Markdown, reasoning, and edit/write/patch output. Results include serialized history bytes, message/part counts, and event bytes.

Each case opens the real application timeline with a busy session, waits for the initial Markdown to be ready, then sends the same 160 text deltas at a 25 ms source cadence. The Node HTTP source does not wait for renderer acknowledgements. Results retain every source emission time and animation-frame interval. Completion means the final marker is in ready Markdown and has reached an animation-frame callback; it does not claim a compositor presentation timestamp. Row replacements count the active streaming row, not normal virtualizer mounts while scrolling.

Use PROJECTION_COUNTERS=1 for separate diagnostic runs. The benchmark build instruments the full-history constructor with its call count, input entries visited, and synchronous elapsed time. The probe is absent from normal builds and disabled in clean timing runs. OPENCODE_PERFORMANCE_TRACE_DIR enables the existing Chrome trace collector. Set PROJECTION_SCREENSHOT to an output path prefix for screenshots after timing. Do not compare diagnostic timing with clean timing.