Codex conflates calling a model API with delegating work to a full coding CLI agent

What issue are you seeing?

The user asked Codex to act as a primary reviewer/orchestrator and delegate implementation or review work to other already-installed coding agents.

Codex treated this as equivalent to sending a large prompt directly to the underlying model API. That substitution removed the worker's coding-agent runtime: repository access, tool use, terminal state, file edits, diff production, completion lifecycle, and recoverable session state.

As a result:

• large review packets were serialized into single API requests;
• timeouts and partial reasoning were misclassified as worker outcomes;
• there was no reliable worker start/wait/result lifecycle;
• Codex claimed or implied that a model response represented a completed coding-agent review;
• repeated retries consumed resources without producing a repository artifact.

Steps to reproduce

1. Ask Codex to delegate a repository task to an installed third-party coding CLI.
2. Make clear that the target is the full coding agent, not only its underlying model.
3. In affected runs, Codex constructs a raw model API call instead of launching or clearly declining the requested CLI delegation.

Expected behavior

Codex should distinguish:

• using a model/provider for inference;
• launching a complete coding CLI as an external worker.

If heterogeneous CLI dispatch is not natively available, Codex should say so plainly and require an explicit process-control integration. It should never present a raw API completion as equivalent to a coding-agent task with repository tools, state, diffs, and completion evidence.

Environment:

• OpenAI Codex Desktop 26.727.6591.0 (x64 MSIX)
• Windows 10.0.26200.8875
• Model observed: GPT-5.6 Sol

No private workspace files or transcripts are attached.