Capturing conversations

Conversations are captured with the turn() API: plain function calls, so it drops into any server framework and is independent of your wire protocol — you hand Eldros the role and content directly.

The turn() API

async def ws_handler(sock):
    while True:
        msg = await receive_user_message(sock)
        with eldros_sdk.turn(msg, session_id=session_id) as t:
            reply = await respond(msg)      # LLM span nests under the turn
            await send_reply(sock, reply)
            t.reply(reply)

Each turn() block records the user message on entry and the assistant reply via t.reply() — and because the turn is the active span while your handler runs, any auto-instrumented LLM call inside it nests underneath automatically. Transcript and trace land in one tree, with no ids to manage.

If the session id isn’t in scope where turns happen, call eldros_sdk.set_session(session_id) once at connect instead and omit the parameter.

Extra roles

Tool calls and other roles go through the same handle, or the low-level primitive:

with eldros_sdk.turn(msg, session_id=sid) as t:
    t.record("tool_call", "search_flights(origin='SFO')")
    t.record("tool_result", "3 flights found")
    t.reply("I found three options for you.")

Simulation traffic: continue the platform’s context

When Eldros simulates your agent, the connection it opens carries standard W3C traceparent + baggage headers with the test run’s identity. Continue them with trace_context — same handler, one extra line around it:

async def ws_handler(sock):
    with eldros_sdk.trace_context(sock.headers):   # handshake headers
        while True:
            msg = await receive_user_message(sock)
            with eldros_sdk.turn(msg, session_id=session_id) as t:
                reply = await respond(msg)
                await send_reply(sock, reply)
                t.reply(reply)

The same pattern works for any HTTP or WebSocket based agent — REST, A2A, JSON-RPC, or anything else. The protocol doesn’t matter; trace_context accepts any header mapping. HTTP / REST / A2A (per-request):

# FastAPI, Starlette, Flask, A2A executor — any HTTP handler
async def handle(request):
    with eldros_sdk.trace_context(request.headers):
        with eldros_sdk.turn(user_msg, session_id=session_id) as t:
            reply = await agent.respond(user_msg)
            t.reply(reply)

WebSocket (once at handshake, covers the whole connection):

async def ws_handler(sock):
    with eldros_sdk.trace_context(sock.headers):   # upgrade request headers
        while True:
            msg = await receive_user_message(sock)
            with eldros_sdk.turn(msg, session_id=session_id) as t:
                reply = await respond(msg)
                await send_reply(sock, reply)
                t.reply(reply)

Everything inside the block is stamped with the run’s episode.id and traffic_type="simulation", so test traffic stays out of your production views. On production connections the headers simply aren’t there and the line does nothing — keep one handler for both. Full model: Simulation & episodes.

Safe to leave in

No-op until init() runs — and init(enabled=False) hard-disables everything.
Exceptions inside a turn are recorded on the span and re-raised — capture never swallows your errors.

Getting started

Capture

Integrations

Concepts

Capturing conversations

The turn() API

Extra roles

Simulation traffic: continue the platform’s context

Safe to leave in

​The turn() API

​Extra roles

​Simulation traffic: continue the platform’s context

​Safe to leave in

The turn() API

Extra roles

Simulation traffic: continue the platform’s context

Safe to leave in