OpenAI merged send_user_message_async into the public Codex repository, letting the coding agent send a question or status update to a developer and keep working instead of blocking for a reply, unlike the existing request_user_input tool. An experimental feature flag requirement was removed within a day, leaving model support as the only gate on availability. The change does not solve timing conflicts: there is no checkpoint stopping Codex from acting on a decision before the developer answers, no expiration on pending questions, and no automatic rollback if Codex's work conflicts with the eventual reply. Subagents cannot use the new tool, so only a root agent communicates with the developer. OpenAI has not announced a release date, target model, or which Codex surface (CLI, desktop, IDE extension, ChatGPT Work) will get it first.
Table of contents
Async messaging replaces blockingFlag removed within hoursNo checkpoint, no rollbackSteer or restart, not reconcileQuestions this post answers
What does the send_user_message_async tool do in OpenAI Codex?
It lets Codex send a question or status update to the developer without blocking, unlike the existing request_user_input tool which waits for a reply before returning control to the model. Once sent, Codex receives an accepted confirmation and can keep calling tools, inspecting files, or generating its answer. Any developer reply arrives later as a new user message rather than a structured response tied to that question. Track how OpenAI evolves Codex's agent-developer messaging model on daily.dev.
Does Codex roll back work if it acts on the wrong assumption while waiting for a developer's answer using send_user_message_async?
No, there is no checkpoint or automatic rollback. Codex can continue working past a pending decision, such as starting to implement SQLite while waiting to hear whether the project should use PostgreSQL, and nothing forces it to stop or undo conflicting code once the developer's actual answer arrives. Developers weighing agent autonomy risks can follow Codex reconciliation gaps on daily.dev.
Can Codex subagents use the new async messaging tool?
No, according to the registration logic merged into the public Codex repository, subagents do not receive the send_user_message_async tool. Only a root agent is responsible for communicating with the developer, even when a Codex task is split across several subagents. Anyone architecting multi-agent Codex workflows can track these capability limits on daily.dev.