WebSocket agents

1. Add traces

Call init() once at startup. Wrap each turn in the connection handler:

import eldros_sdk

eldros_sdk.init(traffic_type="prod")

async def ws_handler(websocket):
    await websocket.accept()
    session_id = websocket.query_params.get("session_id")

    while True:
        user_msg = await websocket.receive_text()

        with eldros_sdk.turn(user_msg, session_id=session_id) as t:
            reply = await agent.respond(user_msg)   # LLM call nests under the turn automatically
            await websocket.send_text(reply)
            t.reply(reply)                          # records the assistant turn

Pass the same session_id on every turn to group the whole conversation together.

2. Connect simulation

When Eldros runs a test scenario it opens a WebSocket connection with W3C traceparent + baggage headers on the HTTP upgrade request. Continue them once at connection start — they cover all turns on that connection:

async def ws_handler(websocket):
    await websocket.accept()
    session_id = websocket.query_params.get("session_id")

    with eldros_sdk.trace_context(dict(websocket.headers)):   # ← add this, once per connection
        while True:
            user_msg = await websocket.receive_text()

            with eldros_sdk.turn(user_msg, session_id=session_id) as t:
                reply = await agent.respond(user_msg)
                await websocket.send_text(reply)
                t.reply(reply)

On production connections (no Eldros headers) trace_context is a no-op — same handler for both. When the headers are present, every span is stamped with:

episode.id — links the trace to the specific test run
traffic_type="simulation" — keeps test traffic out of your production views

For simulation, t.reply() is optional — Eldros already has the transcript from the platform side. turn() is still required so each LLM/tool span is linked to the correct turn, giving you a structured trace even without the client-side transcript.

3. Integration modes

Simulation + production observability (recommended)
Full transcript and traces for both production and eval runs:

eldros_sdk.init(traffic_type="prod")
# handler: trace_context + turn() + t.reply()

Simulation only
Only export during Eldros-driven test runs. Production connections go through normally but nothing is sent to the backend. t.reply() optional — platform has the transcript:

eldros_sdk.init(traffic_type="prod", simulation_only=True)
# handler: trace_context + turn() — t.reply() optional

Production observability only
No simulation integration. Full transcript required since there is no platform-side record:

eldros_sdk.init(traffic_type="prod")
# handler: turn() + t.reply() — no trace_context needed

Traces without turn() — in any mode — cannot be correlated to specific conversation turns. The transcript is what gets judged; traces explain the verdict.

Getting started

Integrations

Concepts

WebSocket agents

1. Add traces

2. Connect simulation

3. Integration modes

​1. Add traces

​2. Connect simulation

​3. Integration modes

1. Add traces

2. Connect simulation

3. Integration modes