mirror of
https://github.com/SawyerHood/dev-browser.git
synced 2026-09-20 13:44:44 +08:00
Sawyer Hood
16dec89cfe
* doobie: skeleton — client, daemon, vm runner, page helpers, build Co-Authored-By: Claude <noreply@anthropic.com> * doobie: snapshot engine, waitForLoad, docs, tests, bench, Bun-native fast client Co-Authored-By: Claude <noreply@anthropic.com> * doobie: fix round 1 from adversarial review (transport backpressure, profiles per mode, page mutex, bringToFront, zombie runs, shot DPR, dialogs, snapshot names/frames/boxes, docs) Co-Authored-By: Claude <noreply@anthropic.com> * docs: handoff notes Co-Authored-By: Claude <noreply@anthropic.com> * launch: mark clean exit to skip session restore, cache --no-sandbox; untrack build artifact Co-Authored-By: Claude <noreply@anthropic.com> * snapshot: drop [cursor=pointer] on inherently interactive roles, no row content names (HN interactive 29k -> 21k chars) Co-Authored-By: Claude <noreply@anthropic.com> * ci: bench tolerance on shared runners; release: npm publish only with NPM_TOKEN Co-Authored-By: Claude <noreply@anthropic.com> * mcp: stdio MCP server over the daemon frames (doobie mcp); ci: explicit test timeout Co-Authored-By: Claude <noreply@anthropic.com> * docs: bb integration note (socket source contract) Co-Authored-By: Claude <noreply@anthropic.com> * fix round 2: front lock (no stale cache), run gate covers handles/frames/popups, --connect extends only touched tabs, doobie chrome verifies launch, self-healing shim + devDependencies, snapshot name fallbacks/pointer inheritance/shadow refs, docs Co-Authored-By: Claude <noreply@anthropic.com> * daemon: drain active requests before exit; tests: realpath-safe and node-optional packaging tests; chrome: share sandbox helpers Co-Authored-By: Claude <noreply@anthropic.com> * relative file paths resolve against the caller's cwd; readFile("downloads/<name>"); TimeoutError for ref waits Co-Authored-By: Claude <noreply@anthropic.com> * runtime: share host Error constructors with the script realm Co-Authored-By: Claude <noreply@anthropic.com> * shim: sh/JS polyglot so bun-only machines run it; v0.1.1 Co-Authored-By: Claude <noreply@anthropic.com> * launch: automation profile prefs (leak-detection dialog off, no password/autofill UI) — fixes dead input after logins in new headless Co-Authored-By: Claude <noreply@anthropic.com> * v0.1.2 Co-Authored-By: Claude <noreply@anthropic.com> * client: retry when racing a shutting-down daemon (flaky stop/status on CI) Co-Authored-By: Claude <noreply@anthropic.com> * feat!: make doobie the dev-browser 1.0 runtime * ci: use Node 24 GitHub actions * fix: isolate authenticated CDP sessions * fix: close run gate escape paths * fix: preserve UTF-8 across protocol chunks * test: stabilize browser context gate coverage * test: isolate browser context gate coverage --------- Co-authored-by: Sawyer Hood <kirbyhood@gmail.com> Co-authored-by: Claude <noreply@anthropic.com>
dev-browser benchmarks
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-browserwhen it exists (bun run build), elsebun run src/cli/main.ts(slower cold start and per-call client time).--binoverrides. The chosen binary is printed first;--bin devforces thebun runpath. --runs Nis the sample count per item (default 15); the table shows medians.--checkexits 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.