5.6 Custom agent profile + model selector not functioning

Codex
1

GPT-5.6 models appear unable to use custom subagent profiles, or otherwise select models and reasoning levels for spawned subagents. I frequently use custom subagent profiles with predefined models and reasoning levels to make my usage stretch further by routing easy tasks to cheaper models.

Today when trying out 5.6-Sol I asked it to spawn GPT-5.6-Terra models on medium effort for implementation tasks, and 5.6-Sol responded with:

“The current subagent API does not expose a model selector, so I can’t truthfully force Luna-low or Codex Spark from this task. I’ll use explorers only for tightly bounded, read-only audits; mechanical implementation will use narrowly scoped fast-worker work orders. I won’t represent an unspecified agent as a requested model.”

Then, I clarified and told it to use the two custom subagent profiles I had, and it acknowledged that it would use them and said:

“Understood. I’ll route mechanical leaf commits through implementation_worker_fast (Codex Spark) and reserve implementation_worker (Terra-medium) for shared-state integration, solver math, and conflict resolution. Explorers will remain Luna-low only.”

I came back a couple hours later to a notice that I had hit my 5 hour limit, and saw that my weekly limit had dropped from 100% all the way to 69%. I went to investigate and I found in my .codex/session/ logs that all subagents had been inheriting the model and effort levels from the main thread, being gpt-5.6-sol and reasoning effort on high:

image
image329×352 19.4 KB

I looked into the issue and I saw that an issue has been opened for it, being related to the multi agent version v2 not exposing a model selector or effort configuration for spawned subagents

Session ID: 019f4876-d5c8-7a22-a1e7-b0d74afd9794
Plan: prolite

2

Exactly the same issue here. Implemented custom subagent profiles and, by the time I found out it was failing, it had burned through the 5-hourly limit (within 30 minutes or so).

3

Please figure this out. I really want to be able to call luna, and terra from sol. I don’t want to have to manually switch between models. I hope this isn’t the intended behaviour.

4

Configuration and model routing are now enabled:

  • Multi-Agent: true
  • Multi-Agent V2: true
  • Sol, Terra, and Luna models: available

But this task still does not expose any native agents dispatch tools. There is no callable spawn/wait/close interface in my current tool inventory, so native delegation remains unavailable in this task runtime.

5

If you go to your .codex/config.toml file, and add this:

[features.multi_agent_v2]
hide_spawn_agent_metadata = false
tool_namespace = "agents"

It will allow 5.6 threads to spawn specific models with whatever reasoning effort