multi_agent_v1 tools are not injected for GPT-5.6 Codex Desktop tasks, but work on GPT-5.5

Start in the Codex Desktop task tool-injection path and compare the reported gpt-5.6-sol and gpt-5.5 behavior when multi-agent is enabled. Reproduce the missing multi_agent_v1 inventory and ensure the five named tools are exposed, or that the unsupported model limitation is clearly shown.

What version of the Codex App are you using (From “About Codex” dialog)?

26.707.91948

What subscription do you have?

ChatGPT Pro

What platform is your computer?

Darwin 27.0.0 arm64 arm

What issue are you seeing?

Summary

In Codex Desktop, multi_agent_v1 tools are not exposed when the task is using gpt-5.6-sol, even though multi-agent is enabled. Switching the same workflow back to gpt-5.5 makes the tools available.

Expected behavior

For a user-owned top-level Codex Desktop task with multi-agent enabled, the following tools should be injected and discoverable:

• multi_agent_v1.spawn_agent
• multi_agent_v1.send_input
• multi_agent_v1.wait_agent
• multi_agent_v1.close_agent
• multi_agent_v1.resume_agent

If gpt-5.6-sol does not support this tool surface, Codex Desktop should clearly show that limitation instead of silently omitting the tools.

Actual behavior

With gpt-5.6-sol, the task tool inventory does not include multi_agent_v1. The assistant therefore cannot create subagents and gets stuck in a recovery loop asking for a new task or tool exposure.

After switching the task/model back to gpt-5.5, multi_agent_v1 tools become available and are discoverable/callable.

Reproduction steps

1. Open Codex Desktop.
2. Use a top-level user-created task.
3. Ensure multi-agent is enabled in local config/features.
4. Select gpt-5.6-sol.
5. Ask the assistant to create subagents or inspect available multi-agent tools.
6. Observe that multi_agent_v1 tools are missing.
7. Switch the task back to gpt-5.5.
8. Repeat the same check.
9. Observe that multi_agent_v1.spawn_agent, send_input, wait_agent, close_agent, and resume_agent are available.

What steps can reproduce the bug?

Open Codex Desktop.
2. Use a top-level user-created task.
3. Ensure multi-agent is enabled.
4. Select gpt-5.6-sol.
5. Ask the assistant to create subagents or inspect multi-agent tools.
6. Observe that multi_agent_v1 tools are missing.
7. Switch the task/model to gpt-5.5.
8. Repeat the same check.
9. Observe that multi_agent_v1.spawn_agent, send_input, wait_agent, close_agent, and resume_agent are available.

What is the expected behavior?

No response

Additional information

No response