perf: collapse app-driving startup turns

[thymikee]

Summary

• make open <app> --foreground preserve normal app/device selection, work for both fresh and existing sessions, and return the initial interactive snapshot in one call; bare iOS auto-discovery remains fail-closed
• make the compact skill and canonical CLI guidance start ordinary app-driving tasks directly, without mandatory help/version/session probes, with a conformance test pinning the default recipe
• extend same-scope iOS tap corroboration from 5s to a bounded 15s so ordinary model turn latency can still prove a tap landed after XCTest reports failure

The wider corroboration window deliberately accepts some additional organic-label-drift exposure. It remains limited to XCTEST_RECORDED_FAILURE, requires the same backend and presentation, and returns an explicit re-observation warning; a false corroboration could still record a tap that did not land, so this is an accepted bounded tradeoff rather than a general success fallback.

Validation

• pnpm check:affected --run — all runnable checks passed, including 2,559 affected Vitest tests, 100% changed-line coverage, Node integration, package, architecture/progress, and replay compatibility
• red-before: reverting the review fixes makes the existing-session explicit-app foreground test fail with INVALID_ARGS and the skill conformance test find both foreground and plain-open recipes; both pass on this head
• focused suite: 11 files / 245 tests passed; pnpm build passed
• live iOS at exact head: open xyz.blueskyweb.app --foreground --platform ios --udid C4578E05-BB63-4A3D-AB53-B011EFA2D981 --session review-1693 --json returned a healthy interactive snapshot on a fresh session, then the identical command succeeded again on the existing session with a new refs generation; the session was closed afterward
• controlled GPT-5.4 runs with isolated config/binary: Element-14 improved from 75.3s / 14 calls to 59.3s / 7 (Argent: 53.6s / 10); Element-03 improved from 174.8s / 30 to 85.4s / 12 (Argent: 122.8s / 18)
• Bluesky reproduced the XCTest recorded-failure path and the changed post-action capture correctly corroborated the landed tap; timing was excluded because local golden verification proved that benchmark fixture was logged out

[github-actions]

Size Report

Metric	Base	Current	Diff
JS raw	2.00 MB	2.00 MB	-207 B
JS gzip	646.5 kB	646.4 kB	-92 B
npm tarball	785.0 kB	785.2 kB	+147 B
npm unpacked	2.74 MB	2.74 MB	+396 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	27.2 ms	27.6 ms	+0.4 ms
CLI --help	65.8 ms	66.3 ms	+0.5 ms

Top changed chunks:

Chunk	Raw diff	Gzip diff
dist/src/registry.js	-117 B	-72 B
dist/src/session.js	-91 B	-16 B
dist/src/cli-help.js	0 B	-5 B
dist/src/interaction.js	+1 B	+1 B

[thymikee]

[P1] Keep the compact default loop self-consistent. skills/agent-device/SKILL.md tells ordinary tasks to use open <app> --foreground only when the prompt explicitly says the app is already foregrounded; its Otherwise recipe is plain open <app>. The following guidance then tells the agent to continue from current refs, but production composes the initial snapshot only when foreground === true, so plain open <app> supplies no refs. This recreates an open -> snapshot/targetless-action turn on the common path and contradicts the canonical top-level help, which now makes open <app> --foreground the default. Make the known-app default recipe use --foreground regardless of whether the prompt states foreground state (normal open resolution still owns app/device selection), and add a skill/default conformance case that would fail if the recipe regresses to plain open without a snapshot.

[thymikee]

Addressed in d28221adb.

• The compact known-app recipe is now unconditionally agent-device open <app> --foreground; the contradictory plain-open branch is gone. A structural conformance test reads the shipped skill and fails if that default loses --foreground or stops promising the initial snapshot.
• Explicit-app foreground resolution now precedes the existing-session rejection. Bare auto-discovery still fails closed with an existing session, while open <app> --foreground works uniformly as open/switch + snapshot.
• Red-before proof: reverting those two fixes makes the existing-session test fail with INVALID_ARGS and the skill test find both foreground and plain-open recipes. The focused suite passes 245/245 on the fix.
• Exact-head iOS proof: the fully scoped Bluesky command returned a healthy interactive snapshot, and repeating the identical command on the existing session succeeded with a new refs generation. The session was closed afterward.
• The PR body now explicitly records the accepted 15-second corroboration drift tradeoff. The stale Haiku tuning comment and remaining RFC branding were also removed/replaced with current evidence.

The full affected gate passed before push, including package validation, 2,559 affected Vitest tests, changed-line coverage, Node integration, architecture/progress, and replay compatibility.

[thymikee]

Re-reviewed exact head d28221adb: the compact-skill P1 and the existing-session foreground concern are fixed. The shipped skill has one known-app recipe (open <app> --foreground) with a structural conformance guard; explicit-app foreground resolution now works on an existing session while bare auto-discovery remains fail-closed. Red-before proof and exact-head live iOS fresh/repeated-open snapshot evidence are adequate, and the bounded 15-second corroboration tradeoff is documented. No actionable code finding remains; ready for human review.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-08-09 08:10 UTC