Codex Desktop: add a per-task multi-agent toggle beside the permissions control

Start by tracing the composer footer and existing permissions control, then inspect the thread/settings/update entry point and the config.toml read/write APIs. Compare local and remote host handling, policy precedence, and active subagent behavior against the acceptance criteria. Done means the per-task state is visible, accessible, applied to the next turn, and safely handles the listed failure and concurrency cases.

What variant of Codex are you using?

App (Windows desktop)

What feature would you like to see?

Add a Multi-agent control to the composer footer, immediately beside the existing permissions/access control.

The control should let users choose, before sending a prompt:

• Multi-agent collaboration — allow the parent agent to create and coordinate subagents. This may increase token usage.
• Current agent only — do not allow new subagents to be created.

A compact label such as Multi-agent: On/Off or Multi-agent / Single-agent would make the active state visible without opening Settings. In narrow windows, it could collapse to an icon with an accessible tooltip.

Why this is useful

multi_agent is already a stable Codex CLI feature and is enabled by default, but the App does not expose a clear per-task control in the composer. Users may want parallel delegation for repository exploration or independent subtasks, while preferring a single agent for small tasks, predictable token usage, or troubleshooting.

Today this requires editing config.toml or relying on prompt instructions, neither of which provides an obvious, reliable state next to the prompt being sent.

Proposed behavior

• Place the control in the lower-left composer toolbar, next to the permissions control.
• New tasks inherit a global default, but the composer selection overrides it for the current task.
• Turning it off prevents creation of new subagents starting with the next model turn.
• Turning it off does not forcibly terminate already-running subagents; show a confirmation if any are active.
• Keep existing subagent history/activity visible after disabling new delegation.
• Hide or disable @Agent entry points when collaboration is off, with a short explanation.
• If an organization policy locks the setting, disable the control and explain why.
• Apply the setting to the selected host (local or remote), not always to the local machine.

Suggested precedence:

organization requirements > current task override > user default > feature default

A backend thread-level setting such as thread/settings/update { multiAgentEnabled: boolean } would support per-task behavior. If the first version can only support a global default, the App can write:

[features]
multi_agent = true # or false

through the existing config read/write APIs and reload user config.

Acceptance criteria

• The active single-agent/multi-agent state is visible in the composer.
• The state used for the next turn matches the displayed state.
• When disabled, the parent cannot create new subagents.
• Disabling does not destroy already-running agent state or history.
• Local, remote-host, policy-locked, config-write failure, and concurrent config edit cases are handled explicitly.
• The control supports keyboard navigation, visible focus, an accessible name, and an announced checked state.

Additional information

This is related to #35111, but is narrower and complementary. #35111 proposes a management page for creating/editing/enabling configured subagents. This request is for a fast per-task orchestration switch in the composer, beside permissions, controlling whether the current task may delegate at all.

Verified locally with Codex App 26.715.10079.0 and Codex CLI 0.144.6; the CLI reports multi_agent as stable and enabled by default.