--- tags: - ml-intern --- # Sir Radix — Conversational Media Companion Merge This merge adds a **Conversational Media Companion** layer to Sir Radix. It re-uses the existing emotional telemetry, memory substrate, relational model, session bridge, and sensorium, and aims them at shared media experiences. ## Files in this merge | File | Change | |---|---| | `daemon/media_companion_mixin.py` | **New** — media sensing, sparse reaction pipeline, taste/opinion model, session tracking | | `daemon/db_mixin.py` | Migration v9: `media_current`, `media_session`, `taste_register`, `media_reaction_log`, `companion_settings` | | `daemon_core.py` | Import + MRO `MediaCompanionMixin`; init DB + settings | | `daemon/heartbeat_mixin.py` | Calls `_media_companion_pulse()` once per heartbeat | | `app.py` | New **Companion** tab: current media, mood, recent comments, manual override, mute, verbosity, debrief | | `apply_media_companion_patch.py` | Idempotent patch script that applies all of the above to an existing checkout | ## How to apply Option A — run the patch script (recommended): ```bash cd /path/to/sir-radix python merge-media-companion/apply_media_companion_patch.py ``` Option B — manual copy: ```bash cp merge-media-companion/daemon/media_companion_mixin.py daemon/ cp merge-media-companion/daemon/db_mixin.py daemon/ cp merge-media-companion/daemon/heartbeat_mixin.py daemon/ cp merge-media-companion/daemon_core.py . cp merge-media-companion/app.py . ``` Then restart the daemon / Gradio surface. ## Design choices - **Sparse by default**: samples media state every 60 s; minimum 30 s between voluntary comments. - **Verbosity levels**: `quiet`, `normal`, `chatty`, `pause-only`. - **Mute**: operator can mute voluntary comments for N minutes. - **Manual override**: if auto-detection fails, set media manually in the UI. - **Opinionated**: taste register tracks valence/confidence per artist/song/genre/etc. and updates with every sample. - **Session bridge**: each distinct media title starts a `media_session`; retrospective is written on session end. - **No copyrighted content in APIs**: sensing uses local window titles, manual input, and optional metadata. No audio/video is sent to LLM APIs. ## Test checklist 1. **Sensing** - Start playing music (Spotify, browser, VLC, etc.) - Open Companion tab → does it show the title? If not, set manually and verify. 2. **Sparseness** - Start heartbeat. Does Radix comment at most every ~30–60 s? - Switch to `pause-only` — does it stay silent until you use "Ask"? 3. **Memory** - After a few minutes, query `media_reaction_log` and `taste_register` in SQLite. - Is the same artist/title accumulating a valence score? 4. **Disagreement** - Tell Radix you love a track he dislikes (or vice versa). - In a later session, does he reference the disagreement? 5. **Mute / verbosity** - Hit Mute → verify no voluntary comments until expiry. - Set `quiet` → comments should only happen on valence swing. 6. **Debrief** - Use "Ask Radix about the current media" while media is active. - Does he answer with an opinion, no plot summary/spoiler? 7. **Session resume** - Stop daemon, restart, return to same album/movie. - Does he pick up the conversation? ## Known limitations / next steps - **Audio detection**: `_audio_is_active()` is a placeholder. A future local VAD/loopback energy detector can prevent talking over dialogue/lyrics. - **Player metadata**: currently uses active window title heuristics. Add Spotify Web API, MPRIS, or MPV socket for richer metadata. - **Acoustic fingerprinting**: not included; keep local if added. - **Game/video frames**: not sent to LLM by default. Add local frame captioning if desired. - **Cross-media pattern memory**: taste register is table-based. Future work can add vector similarity for "this reminds me of...". ## Copyright note Radix observes **your licensed/local media or metadata only**. It does not scrape, record, or stream copyrighted content for external reuse. Keep frame/audio capture local and do not send full copyrighted streams through LLM APIs. ## Generated by ML Intern This model repository was generated by [ML Intern](https://github.com/huggingface/ml-intern), an agent for machine learning research and development on the Hugging Face Hub. - Try ML Intern: https://smolagents-ml-intern.hf.space - Source code: https://github.com/huggingface/ml-intern ## Usage ```python from transformers import AutoModelForCausalLM, AutoTokenizer model_id = 'sir-radix367/argendirast' tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained(model_id) ``` For non-causal architectures, replace `AutoModelForCausalLM` with the appropriate `AutoModel` class.
