perf: speed up iOS replay runner#557
Conversation
thymikee
force-pushed
the
codex/maestro-benchmark-rebase-guard
branch
from
ecba590 to
955a1f8
Compare
|
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ecba5906b8
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| while (Date.now() - startedAt < timeoutMs) { | ||
| const result = await queryDirectIosSelector(params, params.session!, selector); | ||
| if (result?.found) { | ||
| const payload = { | ||
| kind: 'selector', | ||
| selector: selector.raw, | ||
| waitedMs: Date.now() - startedAt, | ||
| selectorChain: [selector.raw], | ||
| directSelector: true, | ||
| }; | ||
| recordIfSession(params.sessionStore, params.sessionName, params.req, payload); | ||
| return payload; | ||
| } | ||
| await sleep(DIRECT_WAIT_POLL_INTERVAL_MS); | ||
| } | ||
| throw new AppError('COMMAND_FAILED', `wait timed out for selector: ${selector.raw}`); |
There was a problem hiding this comment.
For iOS simple selectors this direct path now owns the whole wait loop, but queryDirectIosSelector returns null on runner errors/ambiguous matches and found: false for matches the runner filters out as non-hittable; in both cases this loop keeps polling until it throws instead of falling back to the snapshot selector runtime. That regresses cases the existing waitForSelector path supports, such as a selector that matches multiple nodes or a visible non-hittable/container node, because findSelectorChainMatch only needs at least one snapshot match while the runner-side findElement requires a unique hittable element.
Useful? React with 👍 / 👎.
thymikee
force-pushed
the
codex/maestro-benchmark-rebase-guard
branch
4 times, most recently
from
6326daf to
c68fd4c
Compare
thymikee
force-pushed
the
codex/maestro-benchmark-rebase-guard
branch
from
c68fd4c to
33a7f5e
Compare
1 participant
Summary
Speeds up the iOS replay runner while keeping the reliability fixes that fell out of the Maestro comparison.
High-level changes that moved the needle:
pnpm build:xcuitestleaves the runtime cache in the fast-path state instead of forcing one follow-up scan.id,label,text,value) through direct runner queries/taps forclick,get,is, and the firstwaitprobe, while preserving snapshot fallback for misses and transport failures.wait/findsnapshot polling fresh, so the 750ms snapshot cache cannot return stale UI during selector polling.appBundleIdis known, avoiding the visible helper-app hop.Touched scope: 64 files. Scope expanded beyond the original benchmark prototype because the comparison exposed iOS runner startup, selector fallback, gesture, and text-entry reliability issues that affected the measured path.
Validation
Validated with cache-focused unit coverage, selector/interaction unit coverage, tooling checks, XCTest builds, and repeated iOS benchmark runs.
Checks that passed:
pnpm exec vitest run src/platforms/ios/__tests__/runner-client.test.ts src/platforms/ios/__tests__/runner-xctestrun.test.ts src/platforms/ios/__tests__/index.test.ts src/daemon/handlers/__tests__/snapshot-handler.test.tspnpm exec vitest run src/daemon/handlers/__tests__/interaction.test.ts src/daemon/handlers/__tests__/snapshot-handler.test.ts src/utils/__tests__/interactors.test.tsnode --check scripts/write-xcuitest-cache-metadata.mjspnpm build:xcuitestpnpm formatpnpm check:quickzsh -n .tmp/maestro-prototype/run-ios-benchmark.zshgit diff --checkCache reliability notes:
pnpm build:xcuitest.Latest serial benchmark, after dropping one local close outlier:
Parallel benchmark mode is also available for visual side-by-side inspection, but it is not the fair timing source on this machine: under three-simulator load the local runner replay stayed stable around 13.7s while app readiness moved between
openand explicitwait.