GPT WORLD / INDEPENDENT CODEX FIELD MANUAL Source review · 2026-07-26

Field guide / 10 min

Design Realtime as a live system

Transport, turn detection, interruption, and tools shape the experience.

Track
OpenAI Platform / API
Level
Platform architecture
Maturity
Stable

Realtime sessions are designed for low-latency live interactions. A production voice experience is more than streaming audio: it needs a transport chosen for the client, explicit session state, interruption behavior, tool latency budgets, and recovery from partial turns.

A

Working principles

01

Outcome selects architecture

Use a live session for interactive low-latency work; use request-based audio for files and bounded generation.

02

Interruption is normal

Users speak over systems, networks jitter, and tools finish late. Treat partial audio and cancellation as core states.

03

Keep credentials server-side

Clients should receive narrowly scoped session credentials or use a server-mediated path; never ship a long-lived platform key.

B

Field procedure

  1. 01

    Define the conversation job

    State whether the system responds, translates, or transcribes. These outcomes use different session lifecycles.

  2. 02

    Choose the transport

    Prefer browser-oriented realtime transport for direct media clients and server-oriented streaming for telephony or media pipelines.

  3. 03

    Model turn states

    Track listening, committing, thinking, speaking, interrupted, tool-waiting, reconnecting, and ended.

  4. 04

    Budget tool latency

    Set deadlines, acknowledgement cues, and fallback language for tools that cannot finish inside a conversational pause.

  5. 05

    Test real conditions

    Evaluate noisy rooms, accents, slow networks, device changes, double-talk, reconnects, and long sessions.

PASS / FAIL

Acceptance checklist

  • Long-lived API keys never reach clients.
  • The UI exposes listening and speaking state.
  • Interruption cancels stale output and tool work safely.
  • Session logs connect transport, turn, tool, and outcome events.

WATCH / REJECT

Failure patterns

  • Treating a demo microphone loop as production architecture.
  • Ignoring echo, reconnect, or overlapping speech.
  • Allowing late tool results to mutate a newer conversational turn.