Hugging Face's logo

Datasets:
LoveLogicAI
/
freestream

Dataset card Files
xet
 Community
freestream / docs /SPEC-DRIVERS.md
LoveLogicAI's picture
LoveLogicAI
Driver interface spec
71c3187 verified
preview code
|
Raw History Blame Contribute Delete
3.67 kB

SPEC-DRIVERS — backend interface specs

Every layer is a driver with a common interface, chosen per run. A driver never prints; it raises typed errors and the CLI maps them to exit codes.

Error contract

Exception Exit Meaning
ValidationError 2 bad input (URL, missing image, unknown driver)
DriverError 3 transport failure (HTTP, WebSocket, track error)
SessionError 4 session refused/failed/closed by remote
CapabilityShortfall 6 driver deps/config missing on this host; carries capability + remediation

PresenceDriver 

class PresenceDriver:
    name: str          # registry key: "browser" | "livekit" | "sip"
    audio: bool        # owns a mic track?
    video: bool        # owns a camera track?

    join(meet_url, platform, bot_name, image, voice,
         system_prompt, meeting_password) -> session_id
    status(session_id) -> "joining" | "ready" | "error" | "closed"
    video_connected(session_id) -> bool
    bot_connected(session_id)  -> bool
    leave(session_id) -> None

Contract rules:

  1. join returns a driver-global session id stable across CLI invocations (leave is a separate process).
  2. Sessions must survive the joining process — state lives in a driver registry on disk (.freestream/sessions/<id>/driver.json), not in memory.
  3. join must emit event-log entries (driver.join_attempted, session.created, session.status, …) when a log is attached.
  4. Missing host capabilities are CapabilityShortfall, never crashes.
  5. Sessions must survive process teardown — state lives in a driver registry on disk (.freestream/sessions/<id>/driver.json), not in memory.

livekit driver (MVP target)

  • Joins a LiveKit room via livekit-rtc with a local track: mic-in fed by the persona TTS output, camera-in fed by the avatar frame sink.
  • ASR subscribes to remote audio tracks; faster-whisper consumes them.
  • Lowest jitter of the free drivers; requires the meeting to run on LiveKit/Jitsi (self-hosted or open instance).
  • Config: LIVEKIT_URL, LIVEKIT_TOKEN (room-local token minted by freestream for bot-name).

browser driver (post-MVP)

  • Playwright/Chromium joins as a guest; virtual camera (v4l2loopback on Linux, BlackHole + AVFoundation on macOS) supplies avatar frames; virtual mic supplies TTS audio.
  • Needs per-platform join packs: URL validation, guest-button selectors, permission-grant selectors, pre-admit waiting-room handling. Selector drift is the maintenance hotspot; packs are versioned and pinned per platform build (see docs/PLATFORMS.md).
  • Commercial platforms (Meet/Zoom) are ToS gray zones — shipped last.

sip driver (audio-only)

  • Dials a conferencing bridge via SIP; near-zero cost, no video.
  • Fidelity rung caps at AUDIO_REACTIVE_WAVE (no camera track).
  • Config: FREESTREAM_SIP_URI, trunk credentials.

Persona drivers

  • VoiceBackend: piper (default, CPU real-time) → xtts-local (cloning). clone() returns the Pika-compatible dict and writes life/voice_id.txt
    • life/voice_config.json. Clones are local files, never uploaded.
  • AvatarBackend: selected by select_rung() from the fidelity ladder (FULL_PORTRAIT 4 → STATIC_CARD 1). Rung degrades downward at runtime if measured FPS < target for 5s; every downgrade is an event-log entry.

Cognition

  • MCP server (freestream/mcp_server.py): stdio JSON-RPC by default, HTTP when FREESTREAM_MCP_HTTP=host:port. Tools: session.observe, session.speak, session.listen, session.leave. The agent is the participant; FREESTREAM owns presence and persona only.