LiveKit Agents

import eldros_sdk
from eldros_sdk.integrations.livekit import observe_session

eldros_sdk.init(traffic_type="prod")            # once, at startup

async def entrypoint(ctx: agents.JobContext):
    session = agents.AgentSession(...)
    observe_session(session, ctx=ctx)           # session_id defaults to the room name
    await session.start(room=ctx.room, agent=MyAgent())

​

What gets captured

• Every finalized user ↔ assistant exchange (from the session’s conversation_item_added events) becomes a turn span — input.value/output.value plus structured eldros.message events, the same shape as every other Eldros capture path.
• Session grouping: the LiveKit room name is the conversation id (override with session_id=...). Because observe_session runs before session.start(), the session identity propagates into the tasks LiveKit spawns — so your auto-instrumented LLM calls carry the same session.id.
• LiveKit’s own telemetry (agent spans, metrics) is routed to the same pipeline when available (wire_telemetry=True, the default; harmlessly skipped on versions without it).

​

Notes

• Non-conversation items (system messages, heartbeats) are skipped by design.
• If a session ends with zero recognized turns, the SDK logs one warning — capture problems fail loudly, never silently.
• All hooks are guarded: an Eldros failure can never break your agent session.