fix: stabilize Android Maestro replay interactions

[thymikee]

Summary

Improve Android Maestro replay stability for React Navigation CI flakes around tap-triggered transitions and horizontal tab swipes.

• Route Android horizontal swipeScreen: direction left/right through content-lane coordinates instead of the native gesture preset.
• Confirm Android native assertVisible wait successes with exact Maestro snapshot matching so Albums does not pass on Push albums.
• Add failure-only Android recovery for immediate tap->assert and swipe->assert misses: retry the last relevant interaction once, cap the follow-up wait at 5s, and clear stale tap/swipe recovery state when another interaction supersedes it.
• Dedupe the Android assertVisible recovery confirmation path, model recoverable tap/swipe state as one discriminated union slot, and relax provider-scenario snapshot-count assertions so tests cover behavior instead of exact internal capture counts.
• Scope: 7 files, Maestro runtime/tests plus one Android provider scenario test; non-Maestro commands do not enter this runtime path.

Validation

Local checks:

• ./node_modules/.bin/vitest run --project provider-integration passed: 19 files, 48 tests.
• ./node_modules/.bin/vitest run src/compat/maestro/__tests__/runtime-interactions.test.ts src/compat/maestro/__tests__/runtime-assertions.test.ts src/daemon/handlers/__tests__/session-replay-vars.test.ts passed outside the sandbox: 3 files, 111 tests.
• ./node_modules/.bin/oxfmt --check test/integration/provider-scenarios/android-test-suite.test.ts passed.
• ./node_modules/.bin/oxlint . --deny-warnings passed.
• ./node_modules/.bin/tsc -p tsconfig.json passed.
• Earlier on this branch: pnpm build:all passed.
• Local React Navigation Android full suite under CI-like Pixel 5 dimensions passed: 38/38, 0 retries, 420.1s.

Device/CI evidence:

• Packed this branch and installed it into React Navigation PR 13126 probe branch.
• Focused React Navigation Android CI probe passed all 3 targeted flaky flows first attempt: Native Stack - Card + Modal, Stack - Card + Modal, Tab View - Custom Tab Bar: https://github.com/react-navigation/react-navigation/actions/runs/27496272320
• Full React Navigation Android CI suite passed: 38 passed, 0 failed, 1 flaky in 1335.5s: https://github.com/react-navigation/react-navigation/actions/runs/27496812112/job/81272293307
• The one flaky case was Drawer - Master Detail, which passed on attempt 3. The same flow passed locally 10/10 in isolation and 5/5 after the preceding CI test; the remaining evidence points to a low-frequency Android accessibility hierarchy refresh/state issue in that app screen, not this targeted replay change.

Performance/scope note:

• Normal passing Android Maestro assertions pay one extra exact snapshot confirmation after native wait success. Tap/swipe recovery only runs after assertion failure. Non-Maestro runs are performance-neutral because the changed runtime code is under src/compat/maestro/* and is only used by test --maestro replay.

[github-actions]

Size Report

Metric	Base	Current	Diff
JS raw	1.3 MB	1.3 MB	+2.7 kB
JS gzip	413.6 kB	414.3 kB	+725 B
npm tarball	547.9 kB	548.6 kB	+709 B
npm unpacked	1.9 MB	1.9 MB	+2.7 kB

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	26.8 ms	26.1 ms	-0.7 ms
CLI --help	54.0 ms	51.2 ms	-2.8 ms

Top changed chunks:

Chunk	Raw diff	Gzip diff
dist/src/session.js	+2.7 kB	+725 B

[thymikee]

Addressed the Android Maestro replay investigation with the branch now at a518ea564.

What changed:

• Android horizontal Maestro screen swipes use content-lane coordinates.
• Android native assertVisible successes are confirmed with exact Maestro snapshot matching, which prevents Albums from passing on Push albums.
• Failure-only Android recovery now handles immediate tap->assert and swipe->assert misses, with stale tap/swipe recovery state cleared when another interaction supersedes it.

Validation:

• Focused unit bundle passed: runtime-interactions.test.ts, runtime-assertions.test.ts, session-replay-vars.test.ts -> 110 tests.
• oxfmt on touched files, oxlint --deny-warnings, tsc -p tsconfig.json, and pnpm build:all passed.
• React Navigation focused Android CI probe passed all 3 targeted flaky flows first attempt: Native Stack - Card + Modal, Stack - Card + Modal, Tab View - Custom Tab Bar: https://github.com/react-navigation/react-navigation/actions/runs/27496272320

Residual risk:

• Full 38-file React Navigation Android suite has not been rerun after the final patch, so I left this as draft rather than calling it merge-ready. The focused CI now covers the previously failing Native Stack and Tab View patterns cleanly.

[thymikee]

Reviewed the change — design is sound (recovery state is consumed before any recursion, and tap/swipe remembers clear each other, so no double-recovery or retry loops). One non-regressive cleanup suggestion:

Dedupe the Android assertVisible recovery paths. retryRecentAndroidTapAfterVisibleMiss and retryRecentAndroidSwipeAfterVisibleMiss share a ~30-line identical tail (cap the follow-up wait, re-run native wait, return on success else fall back to single-snapshot), differing only in the retryTap/retrySwipe flag, and the two call sites both duplicate the tap-then-swipe sequence. Extracting a shared confirmVisibleAfterAndroidRecovery(params, args, recovery) and a recoverFromAndroidVisibleMiss(...) caller helper removes the duplication and the platform === 'android' guard repetition. With the file now at ~794 LOC (past the 500-LOC "extract before adding" guidance in AGENTS.md), this trims it by ~17 lines.

Behavior is unchanged — the recovery kind is preserved via a computed [recovery]: true key, so the existing tests pass without edits. I pushed the change to claude/pr-805-review-xy3ewd (stacked on this PR's head) for reference; cherry-pick if useful.

Validation on that branch: focused bundle 110/110, and oxfmt / tsc -p tsconfig.json / oxlint --deny-warnings all clean.

Also rename assertVisibleRetryTapTimeoutMs → assertVisibleRetryTimeoutMs, since it now covers both tap and swipe recovery.

Nit (skip if not worth the churn): MaestroRetryTapTarget.point is typed { x: number; y: number } where the imported Point would do.

---

Generated by Claude Code

[thymikee]

Addressed the cleanup review in commit 8aaaaa6df:\n\n- deduped Android assertVisible tap/swipe recovery through a shared recovery dispatcher and confirmation helper;\n- renamed the retry timeout helper from tap-specific to generic;\n- reused the shared Point type for retry tap coordinates.\n\nChecks after the change: focused Maestro replay bundle 110/110, oxfmt on the touched file, oxlint --deny-warnings, and tsc -p tsconfig.json all pass. I am rerunning the full React Navigation Android suite against this new head now.

[thymikee]

Final validation update for 8aaaaa6df:\n\n- Addressed the cleanup review: shared Android assertVisible recovery confirmation for tap/swipe, renamed the retry timeout to generic naming, and reused the shared Point type.\n- Static/local checks still pass: focused Maestro replay bundle 110/110, oxfmt on touched file(s), oxlint --deny-warnings, tsc -p tsconfig.json, and pnpm build:all.\n- Local React Navigation Android full suite under CI-like Pixel 5 dimensions passed 38/38 with 0 retries in 420.1s. A wider local emulator run was 37/38 because Drawer - Master Detail can hit a layout-specific ambiguous Go back; the same drawer flow passes under Pixel 5 dimensions.\n- Full React Navigation Android CI probe passed: 38 passed, 0 failed, 1 flaky in 1335.5s: https://github.com/react-navigation/react-navigation/actions/runs/27496812112/job/81272293307\n- The one CI flaky case was Drawer - Master Detail, passing on attempt 3 after two Article tap misses. The originally targeted Native Stack/Card Modal, Stack/Card Modal, Material Top Tabs, and Tab View cases passed in the full CI run.\n\nScope/perf: changes remain under Maestro replay runtime/tests. Non-Maestro runs do not enter this code path. Passing Android Maestro assertions add one exact snapshot confirmation after native wait success; tap/swipe recovery is failure-only. I marked the PR ready for review with the drawer suite flake called out as residual risk.

[thymikee]

Merge-readiness cleanup pushed in 64809fec0.

What changed:

• Updated the Android provider-scenario tests that failed CI to assert snapshot-count ranges instead of exact internal capture counts. This matches the adjacent existing range assertion and keeps the behavior checks intact while allowing the extra exact-match snapshot introduced by Android Maestro assertion verification/recovery.
• No runtime behavior changes in this commit. Branch scope is now 7 files: Maestro runtime/tests plus one provider scenario test.

Validation after the cleanup:

• ./node_modules/.bin/vitest run --project provider-integration passed: 19 files, 48 tests.
• Focused Maestro/runtime bundle passed outside the sandbox: 3 files, 110 tests.
• ./node_modules/.bin/oxfmt --check test/integration/provider-scenarios/android-test-suite.test.ts passed.
• ./node_modules/.bin/oxlint . --deny-warnings passed.
• ./node_modules/.bin/tsc -p tsconfig.json passed.

Notes:

• pnpm format could not run locally because the pnpm shim tried to verify/download pnpm 11.1.2 and the registry is unavailable in this sandbox; I ran the local oxfmt binary directly for the touched file.
• A local full vitest run --coverage attempt progressed past the fixed provider assertions but timed out three unrelated Android fill tests under coverage. Those tests passed in the previous PR CI log, so I am relying on the restarted PR Coverage job for the final signal.

[thymikee]

Zoom-out refactor pass pushed in 1b1c89b3c.

What I changed:

• Replaced the separate recentTap and recentSwipe WeakMaps with one MaestroRecoverableInteraction discriminated union slot. That matches the runtime invariant better: only the last recoverable Maestro interaction should be eligible for assertion recovery.
• Recovery now consumes one union value and dispatches on kind, so tap/swipe mutual exclusion is enforced by the data shape instead of by paired clear calls.
• Successful fuzzy taps and percentage-point taps now clear the recoverable interaction slot, preventing an older tap from being retried after a newer non-recoverable tap path.
• Added a regression test for stale tap recovery after a fuzzy tap.

Validation:

• ./node_modules/.bin/vitest run src/compat/maestro/__tests__/runtime-interactions.test.ts src/compat/maestro/__tests__/runtime-assertions.test.ts src/daemon/handlers/__tests__/session-replay-vars.test.ts passed outside the sandbox: 3 files, 111 tests.
• ./node_modules/.bin/vitest run --project provider-integration passed: 19 files, 48 tests.
• ./node_modules/.bin/oxlint . --deny-warnings passed.
• ./node_modules/.bin/tsc -p tsconfig.json passed.

I did not split a new module in this pass. The next architecture step, if we keep growing this area, should be extracting Android assertion recovery from runtime-assertions.ts into a focused runtime-recovery/runtime-android-recovery module.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-06-14 13:46 UTC