Skip to content

engineering-status.md

Latest commit

 

History

History
64 lines (51 loc) · 7.67 KB

engineering-status.md

File metadata and controls

64 lines (51 loc) · 7.67 KB

Engineering Status

QwenVoice is the merged Apple-platform codebase that currently ships publicly as QwenVoice v1.2.3 on macOS and will ship its next macOS release under the forward Vocello brand. The repo carries a shared engine core, a macOS XPC-isolated runtime path, and an iPhone engine-extension path without reintroducing a secondary Python backend or standalone CLI surface.

The current milestone is operating on a macOS-first release track: macOS is the only public release target for the next ship, while iPhone remains a maintained compile-safe and deferred release surface.

Rescue Checkpoint

As of main commit 63a5e02 (Guard macOS UI observation boundaries), the repo is back on a green rescue baseline:

  • local main and origin/main are aligned at the same commit
  • GitHub Project Inputs passed for 63a5e02 in run 24973916220
  • GitHub Apple Platform QA Gate passed for 63a5e02 in run 24973916195, including contract, Swift, native runtime, hosted UI smoke, macOS build, iPhone build, and unsigned macOS release-artifact verification
  • local pre-push proof for 63a5e02 passed ./scripts/check_project_inputs.sh, ./scripts/qa.sh validate, git diff --check, ./scripts/qa.sh test --layer swift, and ./scripts/build_foundation_targets.sh macos
  • prior local full release proof passed from c6beacd with ./scripts/release.sh, ./scripts/verify_release_bundle.sh build/Vocello.app, and ./scripts/verify_packaged_dmg.sh build/Vocello-macos26.dmg build/release-metadata.txt; the hosted QA gate re-proved the unsigned release-artifact lane for 63a5e02
  • controlled local macOS acceptance on April 26, 2026 launched build/Vocello.app, switched Custom Voice / Voice Design / Voice Cloning, typed a Custom Voice script, generated a 2-second preview, played it through the sidebar player, and persisted the output under ~/Library/Application Support/QwenVoice/outputs/CustomVoice/

The next recovery work should keep this baseline stable: native SwiftUI only, no broad visual redesign, no speculative model work from screen mount, and no overlapping heavy build/test commands on the 8 GB local development machine. Manual app acceptance does not replace Instruments or signed-release proof, but it confirms the rescued app can complete the primary Custom Voice path locally.

Current Strengths

  • One shared Apple-platform codebase with explicit separation between UI orchestration and isolated engine execution
  • QwenVoiceCore now owns the repo’s shared engine semantics for requests, results, events, load state, clone state, lifecycle state, and capability negotiation
  • Shared manifest-driven contract for model, speaker, and platform-variant metadata
  • Shared app-layer playback and generation-persistence ownership now lives in Sources/SharedSupport/ instead of drifting across separate platform copies
  • Process isolation preserved on both platforms during generation and prewarm work
  • Shared host lifecycle/capability primitives now live in QwenVoiceCore, with both macOS XPC and iPhone extension paths negotiating through the same lifecycle vocabulary
  • The active macOS XPC helper now hosts MLXTTSEngine from QwenVoiceCore, so the repo no longer relies on a separate live macOS-native policy stack for load, prewarm, generation, and clone behavior
  • The iPhone host/runtime contract now runs through a monitor-backed extension manager that selects a preferred identity, replaces stale transports, and invalidates on teardown instead of leaving that lifecycle implicit in the UI shell
  • Explicit low-RAM policy surfaces for the iPhone path, including guarded and critical memory bands
  • The shared frontend-safe engine state surface now exists as TTSEngineFrontendState, with matching macOS and iPhone store adapters
  • Restored repo workflows for project inputs, the Apple-platform QA gate, macOS release packaging/notarization, and iPhone TestFlight packaging
  • Rebuilt scripts/qa.sh as the repo-owned QA orchestrator for validation, contract/source/native/iOS/UI test layers, diagnostics, and opt-in benchmarks
  • Maintained release scripts for signed/notarized macOS DMGs and iPhone archive/export flows
  • Deterministic local foundation paths now separate package resolution, build, archive, and export work into explicit roots with .xcresult evidence
  • Apple Platform QA Gate now treats .xcresult bundles as first-class artifacts for maintained build and archive/release lanes instead of depending on raw xcodebuild log tails alone
  • An explicit public-homepage posture that keeps GitHub landing-page messaging aligned with the currently shipped QwenVoice v1.2.3 build, with Vocello framed as the forward rebrand that lands with the next macOS release

Current Caveats

  • The iPhone target is Vocello-branded, but the macOS target graph still keeps several internal QwenVoice names and bundle paths for continuity.
  • The supported macOS minimum-hardware path is the 4-bit Speed lane on Mac mini M1, 8 GB RAM; Quality is selectable per generation mode when installed, but floor hardware defaults to Speed and Quality is not guaranteed there.
  • The repo compiles the iPhone app and engine extension, but official minimum-device proof still depends on real iPhone 15 Pro validation under load.
  • Owned-device iPhone validation currently centers on iPhone 17 Pro; that does not replace the separate iPhone 15 Pro proof obligation.
  • The restored iPhone TestFlight workflow still depends on real Apple signing materials, provisioning, and App Store Connect credentials when run outside local source-only validation.
  • The iPhone release/TestFlight path remains maintained, but it is intentionally deferred from signoff for the current macOS-first public release milestone.
  • The macOS and iPhone release verifiers now rely on the checked-in capability and entitlement matrix, but floor-device proof and live signed distribution proof are still separate evidence obligations.
  • The iPhone App Group remains intentionally narrow and file-based, but it is still a real cross-process dependency because model, output, voice, and cache state must be shared between the host app and the engine extension.
  • The legacy QwenVoiceNativeRuntime module has been retired (Session 6, May 2026). The active macOS helper path is QwenVoiceCore end-to-end, with regression coverage migrated to QwenVoiceTests/MLXTTSEngineMockBackedTests.swift and QwenVoiceTests/NativeStreamingSynthesisSessionTests.swift.
  • A plain signed xcodebuild -scheme QwenVoice build on shared local DerivedData can still be polluted by stale build output; the maintained deterministic compile-proof path is the isolated ./scripts/build_foundation_targets.sh flow.
  • Hosted UI smoke can still soft-skip macOS Accessibility/TCC or foreground-window issues; controlled release signoff must use QWENVOICE_E2E_STRICT=1.
  • Manual local app launches and Computer Use remain useful after the qa.sh and build gates, especially for visual polish, real model-load checks, and visible UI benchmark validation. UI benchmark scripts still capture deterministic macOS Accessibility/AppleScript probes, timing, traces, process/memory snapshots, screenshots, and audio-QC artifacts.
  • The public README is intentionally conservative during the refactor period, so public GitHub messaging is narrower than the internal repo architecture docs by design.
  • Preview, debug, and manual-verification helper surfaces still need a keep/refactor/delete pass so the cleanup tracker can close with explicit ownership.

Source Of Truth

When documentation and code drift, trust:

  1. Sources/
  2. project.yml
  3. scripts/ plus .github/workflows/
  4. docs/reference/current-state.md, docs/reference/engineering-status.md, and docs/reference/release-readiness.md
  5. other prose docs