Add a bounded `/orchestrator <task>` command using native multi-agent tools

What variant of Codex are you using?

CLI / TUI

What feature would you like to see?

Add a single-task /orchestrator <task> command that coordinates Codex's existing native multi-agent tools.

The current agent should remain the coordinator: plan the work, delegate read-only investigation to explorer agents and implementation to worker agents, verify their evidence, and synthesize the final response. The coordinator should use the thread's current model. Workers should prefer gpt-5.6-terra with medium reasoning when available and compatible, otherwise inherit the current model settings.

The initial contract should be bounded to eight subtasks and four concurrent workers. Independent tasks or disjoint file ownership may run in parallel, while overlapping writes should be serialized through worker follow-ups.

Bare /orchestrator should display concise usage and policy help. The command should preserve the original user-facing transcript, command history, mentions, structured text elements, and image attachments. If native multi-agent support is disabled, it should show an actionable error without submitting a turn.

Additional information

This is inspired by ozgureyilmaz/pi-orchestrator, adapted to reuse Codex's native multi-agent implementation.

A focused local TUI proof of concept has been implemented from current upstream/main, including preservation, queuing, disabled-feature, and snapshot coverage. The initial version uses instruction-enforced coordinator isolation. Persistent mode and hard coordinator tool filtering would remain separate follow-up proposals.

I understand external contributions are invitation-only and will not open a pull request unless a Codex maintainer explicitly invites one.