[nbisht-aai]

What & why

The iPhone app's dictation only worked keyboard-first on a phone: capture was being restarted from the background on every press, which iOS refuses (!rec), and the window, the keyboard's press and the capture failures had no owner, no acknowledgement and no names. This rebuilds the session on main from the roots, on the mechanism Typeless uses (measured on an iPhone 13, iOS 18.7.8).

• Session contract (BlurtiOSCore/SessionContract.swift): one readiness record (ListeningState: 15 s lease, published on every transition and by a 5 s heartbeat), a per-command CommandAck, keyboard presence compare-and-written by process, FailureReason, hop URLs blurt://start?reason=&host= and blurt://setup?need=fullAccess. Every number sits in DESIGN.md's "two processes" table and is pinned by SessionContractNumbersTests.
• Window (ListeningWindowMachine + ListeningWindow): a pure, generation-tokened state machine (closed / opening / open / interrupted / closing). The foreground gate lives in it; the lifetime counts from the last word; interruptions pause and resume, never close. IdleMode is PictureInPictureIdle — a 369 × 0.079 pt AVSampleBufferDisplayLayer keeper with hidden controls, started in front, which keeps the process UserInteractiveFocal so later presses start capture from the keyboard with no hop and the orange dot is off between dictations — with HeldIdle as the fallback.
• Keyboard (CommandRouter, OwnerWatchdog, PressHandshake, VoiceNotice): a press is acked started / busy / notListening(reason) or times out at 1.2 s into "Tap to open Blurt"; release and cancel only from the owner or the on-screen keyboard; a dead keyboard's dictation is released by the watchdog within 10 s; one status line under the mic says why a press cannot record.
• Capture failures: typed (ListeningSourceFailure), observed (AVCaptureSession runtime error and interruption, route changes), mapped to reasons; DeviceTrace JSONL and KeyboardLog; scripts/ios-device.sh (preflight / install / launch / kill / pull / syslog) for the real-phone loop; simulator fault injection through BLURT_SIM_FAULT.
• Keyboard profiles: classic (iOS ≤ 18) and glass (iOS 26+) geometry and colours per OS generation; the slim bar's mic on the keyboard's centre line. The classic numbers are the plan's fallbacks: Xcode 27 cannot install an iOS 18 simulator runtime, so they stay unmeasured until the on-phone screenshots.
• Hop-and-return (HostApp, HostReturn, HopReturn): with the session off, the mic tap opens Blurt, which turns the session on and brings the host app back by itself once the window is open and the keeper is up (cap 1.5 s). Private API, knowingly: the keyboard reads the host's bundle id off its controller (_hostBundleID, read only where a getter or ivar is declared) and the app calls LSApplicationWorkspace openApplicationWithBundleID: — what Typeless does, seen in its logs. Both fall back to the "◀" hint when absent or refused.
• Keep-open card (KeepOpenView): the app-switcher card shows "Keep Blurt in the background" with the brand screensaver while a session is on and the scene is not active.
• Engine (RewriteGrounding, AssemblyAITranscriber+Transcript): the dictation API's fast rewrite model answers a transcript the cleanup empties ("Yeah.") with the user's style instruction word for word (6/6 against the live route, every prompt ordering, llm_error null). The app no longer pastes a rewrite that is a run of the instruction, or that says nothing, and pastes the verbatim words instead. The service-side fix is a separate DeepLearning PR.
• -BlurtDeveloperMode / -BlurtDeviceTrace persist across the hop's relaunch (LaunchFlags), so the phone's logs no longer stop at the first hop.

Not in this PR, tracked in the plan: the touch-model step (a lift after a sent press is a stop; grille only), consent v2 and its copy, the DeviceFlows probe, the on-phone geometry measurement.

How it was tested

• scripts/check.sh on the merged tree: ios-check 424 iPhone tests, BlurtiOSCore line coverage 96.3 %, swiftlint analyze and periphery clean; engine 1233 tests; the three formatting steps the merge of main tripped (design-sync, swift-format, prettier) fixed in c210324.

• iPhone 13 / iOS 18.7.8, installed with scripts/ios-device.sh: the window opens, the PiP keeper comes up, presses from Messages are accepted:started with capture start parked=true running=true, Recording: 1 only during dictations, results inserted, and no start recording in the background / 561145187 line in the system log. The style-text leak was reproduced against the API and is fixed.

• Not yet verified on the phone: the hop's automatic return (the host-id read and the LSApplicationWorkspace call) and the keep-open card — the phone was unavailable for the last install. The trace's hop host= and return opened= lines decide.

• scripts/check.sh passes

• I read AGENTS.md and this doesn't reintroduce anything deliberately removed — the silent loop and background startRunning stay out; check-invariants.sh passes

• Docs updated: App/BlurtiOS/README.md ("Testing on a phone"), App/BlurtiOS/DESIGN.md ("The two processes, and every number"), AGENTS.md (the iPhone's platform scope)

🤖 Generated with Claude Code

[aikido-pr-checks]

cornersHit(byDrift:) revisits every x-wall hit on each rendered frame, so its per-frame cost grows with animation duration and repeatedly recalculates past corners.

Show fix

Suggested change

	for wallX in 1...wallsX {
	// The drift at which x touches this wall, and where y is then.
	let atX = CGFloat(wallX) * x.travel - x.start
	let distanceY = y.distance(atX)
	let wallY = (distanceY / y.travel).rounded()
	guard wallY >= 1, abs(distanceY - wallY * y.travel) <= tolerance else { continue }
	let atY = wallY * y.travel - y.start
	let wallY = (y.distance(drift) / y.travel).rounded()
	let distanceY = wallY * y.travel
	guard wallY >= 1, abs(y.distance(drift) - distanceY) <= tolerance else { return (0, nil) }
	let atY = wallY * y.travel - y.start
	for wallX in 1...wallsX {
	// The drift at which x touches this wall, and where y is then.
	let atX = CGFloat(wallX) * x.travel - x.start
	guard abs(y.distance(atX) - distanceY) <= tolerance else { continue }

Details

✨ AI Reasoning
​The screensaver computes a frame from elapsed time and calls cornersHit(byDrift:) each time. That function iterates from the first x-wall hit through all hits so far, recomputing historical corner events on every frame. Because the animation can continue indefinitely, the per-frame work keeps growing and repeats prior work.

_{Reply on this comment to give feedback or ignore the issue.
More info}