Codex Voice cuts tool latency by streaming context to the harness before delegation
SIGKITTEN · x · 2026-11-05
OpenAI Codex engineer daidaijiayu shared how Codex Voice reduces backend tool latency: instead of opening a fresh turn only after the audio model delegates, all audio, text, and delegation notifications are streamed continuously to the harness, which aggregates context throughout the conversation. When delegation happens, context is already built, so tools execute immediately.
An accompanying diagram contrasts the two timings: the naive path is delegate → build context → execute tools; the streaming path builds context while audio streams, shrinking the wait after delegation.
The implementation lives in codex-rs/core/src/realtimeconversation.rs (2,900 lines) in the openai/codex repo.
More from coding & agent
- Revolut self-hosts GLM on 56 B200s, cutting coding token cost to ~$0.30/M — alex_verem · 2026-11-05
- "Hand-writing code will sound like a fever dream to our children," dev predicts — tlakomy · 2026-11-05
- Hybrid search plus reranker cuts retrieval failures 67%: Obsidian AI search guide — dSebastien · 2026-11-05
- Agent harnesses aren't dying: harness distillation changes the job, says O'Reilly — bibryam · 2026-11-05
- Anthropic shares Claude Code trick: run Sonnet, Haiku and Opus as builder, scout and advisor — Arindam_1729 · 2026-11-05
- Using JBang to Add Deterministic Java Guardrails to Your Coding Agent via Lifecycle Hooks — myfear3 · 2026-11-05