What We Shipped

Redesigned modal with smooth container scrolling, concise card styling, and interactive customization drawer for switching prompt personas on the fly.

Seamlessly enabled instant access for authenticated Pro/Elite accounts across desktop and web interfaces.

Implemented offline HMAC token verification and remote license activation via the official Algeris licensing gateway.

The senior-review-fix commit landed with a commit message and changelog entry describing provider state follow-ups, but its patch delta was empty (0 files changed). To keep git history strictly truthful and prevent ghost entries in release audits, this empty commit was formally reverted.

Added native-backed macOS/Windows crash-evidence capture with a user-facing toggle in Settings > Developer, recording peak memory, abort reason, and memory snapshots to debug background crashes without user friction.

Built end-to-end evaluation suite measuring retrieval accuracy, needle-in-a-haystack recall, and citation faithfulness across resume, job-description, and company-profile document stores.

Settings and status indicators now reflect real provider health via live probe checks, disabling model options when dependencies (e.g., Ollama daemon, Deepgram key) are unreachable.

Added option to hide dock/taskbar presence entirely, running Haggle purely from system tray / menu bar with global shortcut activation for zero-footprint operation.

Fixed race condition during Just-In-Time document ingestion where fast audio streams could trigger prompt construction before embedding generation completed.

Extracted job-description bullet citations are now surfaced directly in teleprompter footnotes with source-line verification.

Teleprompter debug drawer shows exact retrieval scores, context-window token counts, and embedding latencies in real time.

Enforced strict validation on runtime prompt assembly to prevent unverified context injection.

Isolated multi-turn conversation memory buffers so earlier question context does not leak into unrelated objections.

Users can now delete, edit, or disable custom tactical skills directly from the Persona Modes manager.

Consolidated BYOK provider integrations through an internal resilient LiteLLM gateway with automatic fallback to secondary models on 429/500 errors.

Added preflight architecture check on startup to guard against running x64 native addons under Apple Silicon Rosetta translations.

Resolved Electron main-process initialization hang caused by blocking synchronous IPC calls during initial window hydration.

Removed verbose diagnostic logging from production bundle for improved memory footprint and cleaner developer tools.

Fixed persistent macOS screen-recording / microphone permission prompts by caching system authorization states.

Modal settings sheets now reliably dismiss when clicking outside the backdrop or pressing the Escape key.

Resolved issue where Deepgram and custom Whisper API keys were not correctly written to encrypted local keytar storage.

Included lightweight offline fallback model for basic teleprompter operation when internet connectivity drops mid-negotiation.

SQLite databases now perform clean Write-Ahead Log checkpoints on application exit to prevent corrupted local cache.

Enabled prompt prefix caching for Anthropic Claude 3.5/3.7 and OpenAI models, cutting latency by 60% on subsequent turns.

Automatic graceful degradation from primary cloud model to local Ollama if network timeout exceeds 1500ms.

Added clean user notification and fallback suggestion when DeepSeek returns payment required errors.

Increased context window threshold for group discussions and multi-speaker lecture recordings.

Cleaned up dangling event listeners in audio capture streams and WebSockets on conversation unmount.