- tts.py: replace piper subprocess with HTTP POST to madcat-tts /v1/audio/speech (chatterbox voice cloning)
- chat.js: replace whisper server upload with browser Web Speech API (webkitSpeechRecognition)
- chat.css: style persona picker — appearance:none select, themed with CSS vars, mobile responsive
- main.py: default TTS voice → bt7274-en
Part 1 — Transport swap:
- Replace anthropic.AsyncAnthropic streaming with httpx SSE client
calling opencode's OpenAI-compat /v1/chat/completions on sin:4096
- Auth: basic auth opencode:$OPENCODE_PASSWORD
- Env: OPENCODE_URL (default http://sin:4096), OPENCODE_PASSWORD
- Sidecar binding (sin:4098) consulted per message to resolve active
persona; voice read from binding → cart → env default
- Helper _session_id_for_user: deterministic sha256 slug per email
so sidecar binding survives WebSocket reconnects
- anthropic dep retained in pyproject.toml (not removed — P4 may use it)
Part 2 — Persona switcher:
- PERSONAS dict: bt7274, friday, samantha (slug → voice/backend/prompt)
- POST /api/persona — bind persona via sidecar, maps slug → full config
- GET /api/persona/current — return current binding
- GET /api/personas — list available personas
- chat.html: persona <select> in topnav with server-rendered active state
- chat.js: onChange → fetch /api/persona, update __personaName +
status badge + system message in conversation feed
TODO: add CSS polish for .topnav__persona-wrap (inherits base styles for now)
marauder-actual