Overview

Project Overview

Cosmo was Created by Gabriel S Passos.

Zenith Cosmo 42 is a local-first voice assistant runtime featuring:

• Event-driven architecture with asynchronous priority queue event bus.
• Offline speech recognition using Vosk (Kaldi) with Portuguese support.
• Persistent memory with SQLite-backed repositories for users, conversations, memories, events, and system state.
• Personality system with runtime parameter adjustment and YAML profile management.
• Local command handling for deterministic operations without LLM overhead.
• Provider-based TTS with Piper synthesis and experimental Espeak support.
• WebUI observability dashboard providing read-only runtime monitoring.
• Wakeword detection stabilized around exclusive microphone ownership by runtime mode.
• Experimental local vision pipeline with camera capture, image quality analysis, and baseline face detection.
• Prompt injection with persona, personality parameters, relevant memories, and conversation history.
• Concurrency protection for critical paths: wake word, capture, thinking, and speaking.

The system language and command vocabulary are primarily Brazilian Portuguese. The application entry point is cosmo/main.py, which calls bootstrap.start() inside asyncio.run().

Current Implementation Status

Implemented Features

• RuntimeStateManager with 7-state machine: idle -> wake_detected -> listening -> transcribing -> thinking -> speaking -> cooldown.
• Conversation/TTS concurrency protection with guard methods.
• Local fallbacks standardized across error paths.
• ConversationManager history limit of 10 messages.
• Safe handling of incomplete personality commands with fallback responses.
• Lightweight persona/personality parameter persistence through a JSON state file.
• Persistent memory with SQLite repositories for users, memories, conversations, events, faces, system_state, local_commands, and personality metadata.
• WebUI as read-only observability dashboard inside the Cosmo process.
• SQLite database layer with repository pattern for data access.
• Async event bus with priority queue and metrics collection.
• Wakeword energy optimization with idle sleep and silence grace.
• Logging to both console and SQLite.
• Event persistence to SQLite EventRepository.
• Diagnostics/runtime snapshots through DiagnosticsManager.
• Database-backed local commands with fallback phrases.
• Database-backed personality command aliases and number words.
• Provider-based TTS factory with Piper and Espeak implementations.
• Vosk wakeword detection with continuous monitoring.
• Vision system implemented as an experimental local camera pipeline.
• CameraManager camera capture with fresh frame reading, snapshot saving, and persistent camera session while auto-capture runs.
• Auto-capture periodically captures, saves, and analyzes frames; capture_interval controls periodic processing, not video FPS.
• Frame freshness uses buffer flushing through frame_flush_reads and frame_flush_delay.
• VisionAnalyzer computes brightness_mean, brightness_std / contrast, dark_ratio, bright_ratio, overexposed_ratio, blur_score, backlit_score, image_quality, and face_ready.
• FaceDetector uses OpenCV Haar Cascade as a baseline detector with optional CLAHE preprocessing, geometric filters, eye validation, and raw/accepted/rejected detection counters.
• WebUI /vision displays the latest snapshot, image quality, face detection status, reject reason, and bounding boxes overlay.

Experimental / In Tuning

• Vision is active but experimental; quality classification and Haar-based detection are being tuned for lighting, distance, blur, pose, and false positives.
• Face detection reports raw faces, accepted faces, rejected faces, filter reasons, and the largest accepted face, but it is still a baseline detection stage.

Pending Features

• Face enrollment, face recognition, identity matching, and tracking are not implemented yet.
• Robotic abstraction layer.
• API, CLI, and WebSocket interfaces; placeholder directories exist.
• Task scheduler/planner; placeholder directories exist.