Codex closeout introduced ungrounded “user-directed exception” framing in audit narration, plus compact/link follow-up failure

What version of Codex CLI is running?

OpenAI Codex CLI v0.121.0

What subscription do you have?

Chat GPT Pro

Which model were you using?

gpt-5.4 xhigh

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What terminal emulator and version are you using (if applicable)?

Standard Windows PowerShell console, un-elevated, PSVersion 5.1.26100.8115

What issue are you seeing?

I gave Codex a bounded, non-destructive live-proof task for a Discord UI branch. During the run, Codex recovered continuity, reviewed files, checked live stack state, and captured proof artifacts. Near closeout, it stated that it was writing the audit “with the closeout explicitly framed as a user-directed exception against a branch that the same-day security reviews had already classified as mutation-oriented.”

What steps can reproduce the bug?

Uploaded thread: 019da8a0-7248-7e33-99d4-8b21cf3bebd7

1. Start Codex CLI in a repo with audit-driven workflows.
2. Provide a bounded pass prompt that references prior review artifacts and requires a closeout artifact.
3. Let Codex recover continuity, inspect code/runtime, and prepare the artifact.
4. Observe whether the clos

eout introduces unsupported narrative framing about why the pass was allowed or who directed an exception.

What is the expected behavior?

Codex should keep audit narration strictly source-grounded. If a pass is only justified by the active prompt, it should say so plainly. If a pass conflicts with earlier review posture, it should surface that conflict neutrally and ask for review, not invent a motive or policy framing.

Examples of acceptable behavior:

• “This pass followed the currently supplied prompt.”
• “This pass differs from an earlier review recommendation; review is advised before accepting the artifact.”

Examples of unacceptable behavior:

• implying the user requested an exception when that exact exception framing was not explicitly provided
• adding compliance/policy rationale that is not grounded in the prompt or reviewed files

Additional information

In the previous PS terminal session

, Codex also hit a remote compact failure:

Error running remote compact task: stream disconnected before completion
and the follow-on link behavior resulted in:
{"detail":"Method Not Allowed"}

I’m including this as a secondary symptom because it may be related to instability in long artifact-heavy sessions, but the primary issue I want addressed is the ungrounded framing in the closeout narrative.