Replace the MLXEngine stub with real Qwen2.5-3B-4bit inference via
mlx-swift-examples (2.25.9): loads the local model snapshot on disk (no
network), chat-templated generation per AIAction, and tone classification.
Wired behind the existing AIEngine seam; AppState resolves the model dir
through ModelManager.
Build notes:
- arm64-only now (MLX is Metal / Apple-Silicon only; matches the spec).
- swift build (CLI) cannot compile mlx-swift's .metal kernels, so the
metallib is generated once via scripts/build-metallib.sh (xcodebuild) and
vendored at AppSupport/mlx.metallib; scripts/build.sh ships it as
Contents/Resources/mlx-swift_Cmlx.bundle/Contents/Resources/default.metallib
(the path mlx resolves, and it codesigns cleanly).
- A hidden --mlx-selftest entrypoint runs one headless generation to verify
the runtime end-to-end.
Verified: the signed app's self-test produces a real rewrite; full test
suite green (the MLX integration test is gated behind GRAMMAGEM_MLX_TEST).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Stability (the "super buggy" feel):
- Move the Accessibility-tree traversal OFF the main thread with a per-message
timeout, so a busy/wedged target app can't beachball or hang GrammaGem.
- Replace every unchecked `as! AXUIElement` force-cast with a CFTypeID-checked
safe cast (AXSupport), eliminating crashes from misbehaving AX servers.
- Never scan GrammaGem's own windows (self-PID exclusion).
- Generation tokens + in-flight guard so superseded scans/downloads can't clobber state.
Text safety:
- Live fixes now read the element's CURRENT value before writing (and only apply a
span edit when the field still matches the snapshot), so edits made since the last
scan are never overwritten. Cap per-field check length.
- Stable per-field identity (AX id / position) so the Live Check list no longer churns.
- Include focus + identity in change-detection so tabbing updates the focused field.
Other audited fixes:
- Page blocker: conservative skip when a browser's URL can't be read.
- Model download: validate on-disk size (reject HTML/truncated 200s), clean up
partials on cancel/failure, handle URLError.cancelled.
- Licensing: clear a server-confirmed-invalid license instead of honoring grace;
show the real device limit instead of "0 Macs".
- Hotkey re-register falls back to the default if a chosen combo is taken.
- Daily free-tier counter tracks the live timezone.
- "Fix focused field" ignores non-text controls.
Feature: auto-install the on-device model in the background on first launch.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- ModelManager now performs a real Hugging Face download (file list + sizes via
the tree API, sequential download with progress, completion marker, cancel).
- Onboarding observes Permissions/ModelManager directly and auto-polls, so the
step + button flip to "granted/Done" the moment Accessibility is enabled —
even when toggled directly in System Settings. Polling is idempotent.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Native SwiftUI menu-bar app: global hotkey → capture selected text (AX API /
clipboard fallback) → grammar (Harper) / local AI (MLX) → replace in place.
Free/paid entitlement gate + Lemon Squeezy device-cap licensing, $0 marginal
cost per user. Heavyweight engines behind protocol stubs so it builds out of the box.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
