Lots of UI and UX fixes

• Tried to make the everything more user friendly for setup and configuration changes
• Conversational trip-ups and hanging responses on certain query combinations fixed

Conversation mode, Pocket TTS, Better Obsidian

• Voice experience: added exclusive single-satellite conversation mode, better contextual follow-ups, speech probing/barge-in during TTS, and browser TTS buffering limits for more responsive interruption.
• New integrations and voice options: introduced experimental CPU Pocket TTS with one-shot WAV voice cloning, plus an Obsidian editor bridge/plugin for reading and appending notes.
• Assistant behavior/configuration: expanded configurable personalities and prompt handling, improved Home Assistant follow-up/media resolution, and model-source documentation now in MODELS.md.

GGUF and Higgs Experimental Models, Unified Model Serving

• GGUF models can be loaded in for ASR and TTS
• Higgs TTS v3 (GGUF) version can be loaded in if you want to experiment with emotions
• Local SLM models are now all loaded in through the 'local' unified in process llama server
• Lots of other fixes, changes etc. etc. etc.

Ease-of-use fixes and headless satellite

• Improved first-run setup: added blocking model-download preflight, clearer HTTPS URL/banner guidance, and moved Home Assistant/SearXNG setup into an optional Connect step.
• Deployment hardening: added single-port HTTP/HTTPS TLS sniffing, /ready health endpoint and Docker HEALTHCHECK, plus first-boot data ownership handling.
• Expanded and stabilized clients: introduced a configurable headless satellite client; fixed browser/iOS microphone, loudspeaker, TTS muting, endpoint duplication, and barge-in-related behavior.

multi-satellite, VAD and Spotify Search

• Proper multi-satellite available now, multiple browsers can login to server and run voice/chat without crashing everything. You can set what room your in as well.
• The changeover to WebSocket for audio lost our neat VAD end-of-speech detection. It is back and latency is way better again!
• Also brought back our own Spotify tool. This now links in with Home Assistant for playing to Sonos devices in different rooms, but give you the proper "smarts" in your searches, so you don't have to give exact track names.

WebSocket and more simplifications

• Moving towards making this simpler to setup, no more sounddevice dependencies or docker compose files. Just launch the container and test.
• Brought my own Spotify tool back into the mix after being disappointed in the HA implementations. Fallsback to HA options if not configured within Fulloch.
• Lots of testing of models to see what might work on other machines, CPU only etc. Still a lot of work here to be done.

OpenAI endpoint and CPU Tier

Big Release! Here are just a few of the changes:

• Web UI becomes central configuration point, simple docker compose up and then launch the browser to get up and running with the setup wizard.
• CPU Tier is back, lots of testing and the best model configuration has been chosen (at least for my current setup).
• OpenAI API can be used instead of the in-built LLM model. Offload your requests to an already running LLM (Note: you will lose the forced JSON GBNF ruling on this setup, so higher chance of bad tool calls etc.)
• Too many other fixes and prompt tweaks to mention...

Snappier, cleaner and better responses

• Focus in the prompts, intent examples and tool descriptions for better AI interpretation
• Speculative detection of speech ending rather than waiting 1.5s after every turn, snappier response times
• Full stopping, if user says "Atticus, stop!" it just stops everything and doesn't keep listening for follow up, dashboard also includes a stop button now
• Filter out ASR hallucinating biased terms and wakewords on coughing or tapping noises
• Lots of little agent and tool fixes and prompting tweaks to squeeze more out of the 9B model

Improved dashboard, entity blocking, compacted history, more fast paths

• Dashboard looks a lot better and now includes a new entities tab, where users can decide to disable Fulloch access on selected devices in their Home Assistant setup
• History is now compacted with bloated tool call results being cut, so the agentic loops can run on history that matters without having unneeded context
• More Regex fast paths. The agentic setup meant simple lights off commands were taking up to 5 seconds. Regex fast path means these now happen in half a second.

VAD for end speech detection and better notes searching

• SileroVAD is now properly used to detect when a user stops talking, this should work way better in noisy environments than the previous RMS energy detection
• Notes are now searched by both text and semantic in one go, improving the results on first look