[Bug]: v0.1.1 sends same-model Codex subagents through delegate_task, using about 2.8× the allowance #23525

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/web

Steps to reproduce

1. Start a Codex thread on gpt-6.1-sol and set reasoning to Extra High.
2. Run a long workflow that encourages subagent use without specifying subagent settings.
3. orchestrator_capabilities reports non-empty inheritedOptions, because a reasoning level is selected.
4. The agent calls delegate_task instead of native subagent tools (spawn_agent in Codex). T3's injected instructions require this: "When model options are requested or the child keeps the parent's provider instance and model with nonempty inheritedOptions […], use delegate_task even for native-supported models."

Expected behavior

Runtime and allowance use match v0.1.0, for example by running same-model subagents as native subagents.

Actual behavior

I ran one workflow with the same prompt, settings and data:

• v0.1.0 used 19 native subagents and finished in 1h 13m. It used 13% of my weekly Codex usage allowance (123M tokens).
• v0.1.1 used 56 subagents through delegate_task. It took 2h 48m and used 37% of the weekly allowance (432M tokens).
• A second v0.1.1 run hit Codex capacity errors. The agent then sent 6 of its 29 subagent tasks to Claude Opus 5.5 without being asked. They hit my 5-hour Claude limit, and I stopped the run after 1h 17m.
• Codex CLI still finishes the workflow normally. It used 18 native subagents, took 1h 17m and used 12% of the weekly allowance (104M tokens).

Likely cause:
The rule assumes native subagents lose the parent's reasoning effort. In Codex they keep it. build_agent_spawn_config copies the parent's model and reasoning effort into the child. Only three things change the child's effort:

1. A reasoning_effort value in the spawn_agent call.
2. A default subagent effort in the Codex config.
3. An agent role that sets model_reasoning_effort. Codex's built-in explorer role sets none.

The rule came from a CodeRabbit review comment on #16595. The PR notes it didn't test how agents follow the new instructions.

The same rule has other costs:

• Lean context for `delegate_task` children: let a worker skip the orchestration instructions and orchestration tools #21783: each child's first request carries about 3.1k tokens of T3 instructions and 1.0k tokens of T3 tool names.
• [Bug]: T3 Code makes Claude Code subagents cost about 15% more #22749: for one user, delegated Claude children cost $110.52 instead of $95.86 (15% more). They get a 1-hour cache, while native subagents get a 5-minute cache.

Suggested fix: apply the rule only to settings native subagents can't carry, such as Claude fast mode. Or make it a setting.

Impact

Major degradation or frequent failure

Version or commit

T3 Code v0.1.1. Commit 9a05676 (#16595) added the rule. v0.1.0 doesn't have it.

Environment

Windows 11 Pro, Codex CLI 0.163.0, gpt-6.1-sol at Extra High

Logs or stack traces

Screenshots, recordings, or supporting files

No response

Workaround

Downgrade to v0.1.0, or run the workflow in Codex CLI via Powershell. Asking for native subagents in AGENTS.md probably doesn't help: #23258 reports that Codex follows T3's injected instructions over repo instructions.