Measures the numbers from docs/design-decisions.md §8 against an isolated DEV_BROWSER_HOME temp dir and a warm headless daemon. Everything is cleaned up at the end (daemon stopped, temp home removed), also on error.

bun run bench/run.ts [--check] [--bin PATH] [--runs N]

Uses dist/dev-browser when it exists (bun run build), else bun run src/cli/main.ts (slower cold start and per-call client time). --bin overrides. The chosen binary is printed first; --bin dev forces the bun run path.
--runs N is the sample count per item (default 15); the table shows medians.
--check exits 1 if any measured item is over its target. Items that are skipped because the API is not implemented yet count as pass.

item	what is timed	target
`1+1`	wall time of one `dev-browser --headless -e '1+1'` process	25 ms
`getPage+title`	wall time, `getPage("bench")` + `title()`	40 ms
`evaluate(()=>1)`	wall time, `getPage` + one evaluate	40 ms (informal)
`snapshot small`	`page.snapshot()` timed inside the script, ~3 KB page, ~60 elements	30 ms
`snapshot large`	same, SERP-like page (~1500 nodes, 300 links, 100 buttons)	150 ms
`shot viewport`	`page.shot()` timed inside the script	120 ms
`per-call overhead`	(50x `await page.evaluate(()=>1)` in a script / 50) minus the same loop via raw `puppeteer.connect` to the daemon's Chrome	0.1 ms
`cold start`	stop daemon, remove socket, time the first `1+1` (daemon spawn + Chrome launch), median of 3	1500 ms

Fixtures live in bench/fixtures/ (small.ts, serp.ts) and are served by a local Bun.serve on 127.0.0.1. bench/fixtures/fixtures.test.ts pins their size (bun test bench/fixtures).

In-process micro-harness

bun run bench/inproc.ts [--runs N]

Calls runScript from src/daemon/run.ts directly with a fake emit and a BrowserManager: no socket, no client process. Compare its 1+1 median with the 1+1 row of bench/run.ts to see how much of the end-to-end time is the client (process start + socket round trip) vs the daemon.

CI

CI runs bun run build && bun run bench/run.ts --check on Linux (headless Chrome, --no-sandbox is applied automatically when the sandbox is unavailable). Medians are noisy on shared runners; bump --runs rather than the targets if it flakes.