v0.5.0: Leaner, Faster, Sharper AI Assistant

v0.5.0 is a major refactoring release that removes dead services, tightens real‑time inference, and sharpens tool selection with a live mode‑gate. The codebase sheds over 10,000 lines of legacy code while adding targeted features that make Chalie more responsive and reliable.

The mode‑gate classifier now runs in production, promoting capability tools (schedule, goal pursuit, research and more) per turn based on an 8‑head multi‑label model. After successful shadow‑mode verification the gate is live, and the plan‑head fire threshold was lowered from 0.8 to 0.5 to stop productivity tools from being silently demoted on common natural‑language requests.

ONNX inference is hardened against concurrent‑call out‑of‑memory crashes with a single FIFO worker queue, GPU auto‑detection with CoreML strip for models exceeding Metal’s texture limit, and NaN/Inf guards on all paths. Classifier heads are now pre‑shipped with the repository so fresh clones classify on the first turn without a model download.

A new SubconsciousWorker daemon fires an idle cognition tick every 5 minutes — consolidating super‑episodes, running decay, extracting behavioral patterns, and updating the user synthesis — but only after 30 minutes of user inactivity and when the background LLM queue has headroom.

The chat interface gains drag‑and‑drop image upload with OCR preview and blocking text injection into the prompt, a polished glass‑surface design, and a wake‑lock for voice recording and playback. The system prompt now mandates proactive memory recall before offering recommendations, and the telemetry world‑state section renders dynamically from heartbeat data.

Under the hood, the release deletes KnowledgeService, Goals, AmbientInference, AdaptiveLayer, ConsequenceClassifier, and several daemon threads. Static thinking‑levels are replaced with a continuous deliberation score, the user summary is moved to the user prompt for recency weight, and the delete‑all privacy path now wipes every user‑data table.

Mode‑gate classifier live: 8‑head cognitive intent model promotes capability tools per turn, plan threshold lowered to 0.5
Inference serialized into a single worker queue with GPU auto‑detection and CoreML texture‑limit guardrails
SubconsciousWorker drives idle cognition cycle: consolidate, decay, extract behavioral patterns, synthesise user summary
Image drag‑and‑drop with OCR preview; proactive memory recall mandated before recommendations
Massive cleanup: KnowledgeService, Goals, AmbientInference, AdaptiveLayer, ConsequenceClassifier and more removed
Deliberation score replaces static thinking level; user summary moved to user prompt for recency anchoring