Agent Refusal: Model creates multi-stage plan but refuses to execute sequential stages due to "total scope" limits

What version of Codex is running?

codex-cli 0.63.0

What subscription do you have?

Pro subscr.

Which model were you using?

gpt-5.1-codex-max

What platform is your computer?

Linux 5.14.0-611.9.1.el9_7.x86_64 x86_64 x86_64

What issue are you seeing?

I am building a project called "tensorguard/gpubudget". I asked the agent to make a plan to build the system in full. The agent proposed a detailed 11-stage plan (Stage 0 through Stage 11).

What steps can reproduce the bug?

Uploaded thread: 019ac941-fc7d-7780-8597-add955384e47

I replied to the plan with: "cool, proceed" and subsequently "you made a plan, follow as it is stage by stage."

What is the expected behavior?

The agent successfully ran go get to install SDK dependencies (Stage 1 prep).

The agent then immediately stopped and refused to write the actual code.

The refusal message stated: "I can’t safely implement the entire remaining backend slice... in this single session."

The agent failed to recognize that "follow stage by stage" meant it should execute only Stage 1 and then stop to await further user input. Instead, it hallucinated that it was being asked to generate Stage 1 through Stage 11 in a single inference turn, triggered a safety/length guardrail, and refused to proceed.

Expected Behavior: When a staged plan is agreed upon, the agent should recognize the "stage by stage" instruction as a directive to execute Stage 1 only, verify the build, and then pause. It should not refuse the task based on the complexity of future stages (Stages 2-11) that it has not yet been asked to execute.

Additional information

This is a recurrent issue for case like this and inability to debug. I have to re-enable my old claude-cli subscription just to get the situation debugged.