Allow direct conversation with a subagent thread without interrupting the parent thread

What variant of Codex are you using?

Codex Desktop App with subagent / multi-agent workflows.

What feature would you like to see?

Please allow users to open an individual subagent thread and converse with that subagent directly, while the parent thread continues its current work unchanged.

The subagent inspector currently behaves primarily as a read-only activity view. For interactive multi-agent work, it should also provide a composer whose messages are routed exclusively to the selected subagent.

Expected interaction model

• Opening a subagent thread provides an interactive composer, not only a read-only activity view.
• Messages entered there are delivered only to that selected subagent.
• The parent thread continues its current turn and orchestration without being interrupted, steered, restarted, or receiving the same message.
• The UI clearly shows the active message target, for example Parent or Subagent: <name>.
• Users can send multiple follow-up questions, corrections, or verification requests to the same subagent while it is working.
• Queue versus interrupt behavior is scoped independently to the selected subagent.
• Subagent replies remain visible in the subagent thread and can optionally be summarized or forwarded to the parent.
• Navigating between the parent and subagent threads does not lose drafts, queued messages, or execution state.

Why this matters

For complex research and engineering tasks, subagents are not only background workers. They are temporary collaborators with isolated context.

A common workflow is:

1. The parent delegates several tasks in parallel.
2. The user inspects one subagent and notices that it misunderstood an equation, is validating the wrong implementation, or needs an additional test.
3. The user sends a correction directly to that subagent.
4. The parent and the other subagents continue without disruption.

Today, the user must return to the parent thread and ask it to relay the correction. This adds latency, consumes parent context, and can distort or delay the instruction. It also makes the supposedly parallel workflow feel centralized and serial.

Message-routing requirements

The routing contract should be explicit and testable:

• A message sent from a subagent thread must not appear in the parent thread as user input.
• It must not trigger turn/steer, interruption, or a new turn on the parent.
• It must not be broadcast to sibling subagents.
• If the selected subagent is busy, the UI should let the user choose between queueing the message and interrupting that subagent only.
• The parent should receive a subagent update only through the normal result/handoff path, not because the user directly messaged the child.

Suggested UX

A target-aware composer could show:

Send to: Subagent — mesh-validation
[ Queue ] [ Interrupt subagent ]

A plain message in the parent view would continue to target the parent, while a message in the subagent view would target only that child. An @subagent-name routing option could also be supported from the parent view, but it should not be required for the basic interaction.

Regression context

In earlier Codex App behavior, subagent threads appeared more directly interactive. The newer read-only inspector model feels like a regression in human–multi-agent collaboration, even if the underlying orchestration implementation has changed.

Related issues

• #31178 — parent-thread follow-ups are queued and subagent workflows feel serial
• #12047 — async multi-agent UX and targeted @agent messaging
• #33330 — request to restore user-controlled subagent orchestration capabilities

This request is narrower than those issues: it specifically asks for direct user-to-subagent conversation with strict isolation from the parent thread.