Codex CLI 0.107.0: Thread Forking, Voice Devices

TL;DR

• Fork threads into sub-agents without leaving the conversation
• Realtime voice now lets you pick microphone and speaker devices
• Custom tools can return multimodal output (images, structured content)
• Memories are now configurable with a new debug clear command

New

• Thread forking — Branch work into sub-agents while staying in the current conversation, making parallel exploration easier.
• Realtime voice device control — Pick microphone and speaker devices, persist your choices, and send audio in transcription-aligned format.
• Multimodal custom tool outputs — Tools can now return images and structured content instead of plain text only.
• Model availability metadata — App server exposes richer model info with plan-gated tooltips in the TUI to explain limited-run models.
• Configurable memories — Control memory behavior and use codex debug clear-memories to fully reset saved state when needed.

Fixed

• Thread resume sync — Reconnecting with thread/resume now restores pending approvals and input requests instead of leaving clients out of sync.
• Thread start blocking — thread/start no longer blocks unrelated app-server requests, reducing stalls during slow startup paths like MCP auth checks.
• Duplicate assistant output — Interactive terminal sessions no longer print the final assistant response twice.
• Large paste placeholders — Fixed regression from 0.106.0 where large pasted-content placeholders didn't survive file completion.
• ChatGPT account handling — Accounts arriving without plan info now read correctly instead of triggering repeated login issues.
• Diff rendering — TUI diffs now respect theme colors and display cleanly in Windows Terminal and low-color environments.
• MCP OAuth flows — OAuth login now correctly forwards oauth_resource values for servers requiring a resource parameter.

Security & Sandbox

• Sandbox filesystem hardening — Improved restricted read-only handling on Linux and blocked access to sensitive directories like ~/.ssh on Windows.
• Escalation sandbox config — Shell commands now keep their sandbox configuration when rerun, preventing approvals from losing intended restrictions.

Update with: npm install -g @openai/codex@latest or check the full release on GitHub.

Source: Codex