Daeseon Yoo

System shape

[HotKey ⌥Space / NSStatusItem]
        │ records trigger cursor monitor
        ▼
[TriggerPanel] ── user pastes instruction ──┐
                                            ▼
                              [AnalyzeCoordinator]
                                            │
        ┌───────────────────────────────────┼───────────────────────────────────┐
        ▼                                   ▼                                   ▼
[ScreenCaptureService]              [OCRService]                        [AXService]
 (ScreenCaptureKit)                 (Vision framework)                  (AXUIElement)
        │                                   │                                   │
        └────────────► [ElementMatcher: OCR + AX spatial fusion] ◄───────────────┘
                                            │
                                            ▼
                          [LLMDispatcher] → Gemini / Claude / (Qwen local, in progress)
                            FallbackDispatcher on 429
                                            │
                                            ▼
                          [HUDController → HUDOverlayWindow / HUDOverlayView]
                          red box + bubble on the cursor monitor
                                            │
                                            ▼
                          [SessionAuditLog] JSON per session
                          [SecretMasker] regex mask before LLM

Key libs / modules

Sources/ScreenBridge/AnalyzeCoordinator.swift — orchestrates capture → OCR/AX → LLM → HUD.
Sources/ScreenBridge/ScreenCaptureService.swift + ScreenCapture.swift — ScreenCaptureKit-based monitor capture.
Sources/ScreenBridge/OCRService.swift + OCRBox.swift — Vision-framework OCR with bounding boxes.
Sources/ScreenBridge/AXService.swift + AXElement.swift + ElementMatcher.swift — Accessibility tree probe + spatial fusion with OCR for deterministic coordinates.
Sources/ScreenBridge/LLMDispatcher.swift with GeminiDispatcher.swift, ClaudeDispatcher.swift, FallbackDispatcher.swift, QwenLocalDispatcher.swift — pluggable vision-LLM backends; FallbackDispatcher swaps on 429.
Sources/ScreenBridge/HUDOverlayWindow.swift + HUDOverlayView.swift + HUDController.swift — transparent overlay window on the captured monitor.
Sources/ScreenBridge/SecretMasker.swift + SessionAuditLog.swift — pre-LLM secret regex masking and per-session audit JSON dumps.

Why these choices

Swift native over the earlier Tauri stack — Tauri attempt broke on multi-monitor and macOS Spaces traps; swap rationale recorded in DECISIONS.md and the tauri-archive branch.
Package.swift targets macOS 14 because ScreenCaptureKit (12.3+), Swift 6 concurrency, and mlx-swift-examples (Apple Silicon Metal) all need that floor.
MLX dependency ml-explore/mlx-swift-examples from 2.21.0 pinned in Package.swift for the local Qwen2.5-VL-3B 4-bit path (Phase 9.0 Week 1, scaffold only per commit 26a1286).
OCR + AXUIElement fusion instead of LLM coordinates alone — Phase 6.1/6.2 commits show LLM-only coords were unreliable on icon-only Dock items, so deterministic AX matching is layered on top.
FallbackDispatcher exists because Gemini 2.5 Flash hit daily-quota 429s (commits 8db0290, ea77843, 33fd37e) — Claude dispatcher serves as the auto-fallback.

Boundaries

Product intent / why: PRODUCT.md.
Behavioral spec: SPEC.md.
Current build state: STATE.md.
Decision history (incl. Tauri → Swift swap): DECISIONS.md.
Phase history: PROJECT_TIMELINE.md.
Debug learning assets: TROUBLESHOOTING.md (project-local) and docs/troubleshooting.md.
Transferable playbooks: docs/playbooks/latency-optimization.md, docs/playbooks/security-for-screen-llm-agents.md.
Tauri-era v0.1 build artifact narrative: BUILD_REPORT.md.

State of completion

Phases 1–7.3 are landed: capture, OCR+AX fusion, multi-target overlay, continuation hotkey, completion pill, session audit, secret masking, and Claude/Gemini dispatch with 429 fallback are all in Sources/ScreenBridge/. Phase 9.0 (local Qwen2.5-VL via MLX) is at Week 1 skeleton only — QwenLocalDispatcher.swift exists, the mlx-swift-examples dep is pinned, but the local path is not yet wired into AnalyzeCoordinator. No production .app bundle yet; dev loop is swift run.