spawn_agent model override to gpt-5.4-mini is ignored; spawned subagent runs as gpt-5.4

Resolved 💬 2 comments Opened Mar 23, 2026 by piloudu Closed Mar 23, 2026

spawn_agent model override to gpt-5.4-mini is ignored; spawned subagent runs as gpt-5.4

Summary

When using spawn_agent with an explicit model override of gpt-5.4-mini, the spawned subagent thread appears to run as gpt-5.4 instead.

This was observed in Codex Desktop.

Expected behavior

If spawn_agent is called with:

{
  "model": "gpt-5.4-mini"
}

then the spawned child/subagent thread should run on gpt-5.4-mini.

Actual behavior

The parent thread requested gpt-5.4-mini, but:

  • the child thread runtime metadata reported model: "gpt-5.4"
  • the Codex Desktop UI also showed the subagent as using gpt-5.4

So it looks like the override was requested but not honored.

Repro shape

Parent thread:

  • model: gpt-5.4
  • thread/session ID: 019d1a89-9b1c-7c20-a95f-b163d2aa9b7d

Subagent spawn:

  • agent_type: "explorer"
  • fork_context: true
  • model: "gpt-5.4-mini"
  • reasoning_effort: "medium"

Observed child thread:

  • model shown in UI: gpt-5.4
  • model shown in child thread turn context: gpt-5.4
  • child thread/session ID: 019d1a8a-1d07-7e11-97f8-31200d81bbd9

Evidence

In the parent/related session log, the spawn_agent calls explicitly requested gpt-5.4-mini:

  • spawn_agent(... "model":"gpt-5.4-mini", "reasoning_effort":"medium", ...)

In the spawned child thread log, the child turn_context reported:

  • "model":"gpt-5.4"

So there is a mismatch between the requested subagent model and the actual/runtime reported subagent model.

Relevant IDs for correlation:

  • parent thread/session ID: 019d1a89-9b1c-7c20-a95f-b163d2aa9b7d
  • child thread/session ID: 019d1a8a-1d07-7e11-97f8-31200d81bbd9

Environment

  • Codex Desktop
  • version: 0.116.0-alpha.10

Notes

This may be one of:

  1. the spawn_agent.model override is ignored
  2. the runtime silently falls back to the parent model
  3. the UI/logging reports inherited/default model instead of the actual execution model

I found issue #2840 potentially adjacent in spirit because it also involves model override/config not being honored, but this issue is specifically about subagent model overrides in Codex Desktop.

Authored with Codex.

View original on GitHub ↗

This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗