feat(openai): route OpenAI Realtime (WebSocket) through LLM Gateway

[cosminacho]

What & why

Adds OpenAI Realtime API (WebSocket) support, routed through the UiPath LLM Gateway. UiPathOpenAI / UiPathAsyncOpenAI now expose client.realtime.connect() exactly like the stock OpenAI SDK:

client = UiPathAsyncOpenAI(model_name="gpt-realtime")
async with client.realtime.connect() as conn:
    await conn.session.update(session={"type": "realtime", "output_modalities": ["text"]})
    await conn.conversation.item.create(item={
        "type": "message", "role": "user",
        "content": [{"type": "input_text", "text": "Say hello!"}],
    })
    await conn.response.create()
    async for event in conn:
        if event.type == "response.output_text.delta":
            print(event.delta, end="")
        elif event.type == "response.done":
            break

How it works

The Realtime API speaks over a WebSocket, so it can't use the existing httpx routing. A .realtime cached-property override returns a gateway-aware resource wrapper that, on connect():

• points the client's websocket_base_url at the gateway passthrough realtime path (.../vendor/nativeopenai/model/<model>/realtime),
• sets the S2S bearer token as api_key (read straight from LLMGatewayS2SAuth.access_token; the SDK sends it as Authorization: Bearer on the WebSocket upgrade),
• injects the X-UiPath-* routing headers on the upgrade.

Completions/embeddings are unaffected — their auth comes from the httpx auth pipeline, and the realtime URL is built lazily only when .realtime is accessed (so platform/agenthub construction is untouched).

Why no separate class / no LangChain wrapper

The openai-python README shows realtime works on the standard client — no subclass needed — so it's folded into the existing clients. LangChain has no realtime chat-model abstraction (langchain#28086 is still open; ChatOpenAI has no .realtime), so there's nothing to wrap there — realtime is used from a LangChain project by dropping down to the core client.

Packages affected

Both (version together):

• core → 1.15.0: realtime clients in uipath.llm_client.clients.openai; openai extra now installs openai[realtime] (pulls in websockets); helper build_realtime_ws_base_url.
• langchain → 1.15.0: openai extra now also pulls uipath-llm-client[openai] so realtime works out of the box from a uipath-langchain-client[openai] install. No code change.

Tests / checks

• New unit tests in tests/core/clients/openai/test_realtime_unit.py (URL helper, lazy build, websocket_base_url + token + header wiring, default-model resolution, token refresh). Realtime is a WebSocket so it can't be VCR-recorded; end-to-end was verified manually against alpha (gpt-realtime).
• ruff check, ruff format --check, pyright all clean; pytest tests → 1968 passed, 0 failures.

🤖 Generated with Claude Code