v1.2.0 — Multi-Provider AI

What's New

Multi-Provider AI

• Switch between Gemini, Claude, and OpenAI mid-game from a dropdown in the overlay
• Claude screenshot support — attach the current game frame for visual analysis
• No extra API keys — Claude and OpenAI use your existing CLI subscriptions

DX12 Stability

• Hardened command queue fallback (filters non-DIRECT queues, frequency tracking)
• New hook_delay config for games with long DX12 init (e.g. Horizon Forbidden West)
• catch_unwind + 5s retry on DX12 hook build

Performance

• Zero mutex locks on render hot path when overlay is hidden
• Atomic fast-paths for capture state, health check, and init status
• Static HTTP client (was per-request allocation)

Deep Review (26 fixes)

• Cancel now actually kills proxy subprocesses (was no-op due to generation race)
• Child process leak on error paths fixed
• Stale screenshot after cancel+resend race prevented
• Model name validation in proxy (prevents URL path traversal)
• SSE stream capped at 2MB (OOM protection)
• Unicode sanitizer for ImGui Latin-1 font (em dashes, smart quotes, TM/C/R symbols)
• Translation disabled by default (F10 conflicted with game keybindings)

Dependency Updates

• tokio 1.51.0, toml 1.0.7, TypeScript 6.0, Svelte 5.55, Vite 8.0.3

Full changelog: v1.1.0...v1.2.0