Custom subagent role silently inherits parent model/effort on CLI 0.144.5 (Windows)
Related to #14671, which was closed, but reproducible on the current Windows CLI.
Environment
- Codex CLI 0.144.5
- Windows
- Parent session:
gpt-5.6-terra/medium multi_agent_v2enabled for this one test invocation
Reproduction
Created a valid custom role at $CODEX_HOME/agents/luna-canary.toml:
name = "luna_canary"
model = "gpt-5.6-luna"
model_reasoning_effort = "low"
description = "Canary profile for a narrow, read-only sub-agent model-routing test."
developer_instructions = "You are a canary sub-agent. Complete only the requested read-only task. Do not edit files or run shell commands."
Then ran a read-only canary requesting exactly one luna_canary child. The named custom subagent was accepted, spawned, and returned CANARY_OK.
Actual behavior
The persisted child-session turn_context recorded:
model = "gpt-5.6-terra"
effort = "medium"
multi_agent_version = "v2"
Those are the parent settings, not the role's explicit Luna/low settings. The role name and instructions were applied, but its model and reasoning configuration were silently ignored.
Expected behavior
The child should run as gpt-5.6-luna with low effort, or Codex should fail clearly before spawning. Silent substitution makes cost and compute governance impossible.
Please expose the effective child model/effort in the UI and logs, and ensure that valid role overrides either take effect or produce an explicit error.
This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗