Codex 5.6 Sol repeatedly overrides verified user intent, then fails to route its own incident report

[tittlepl-baj-jk]

Summary
In a Codex Desktop session, the agent repeatedly replaced verifiable user intent and upstream project instructions with unsupported assumptions. It then had difficulty identifying the correct browser and reporting channel when asked to document and report its own failure.
Environment

• Surface: Codex Desktop on Windows
• Model: gpt-5.6-sol
• Reasoning effort: medium
• Date: 2026-08-22
• No real legal-case data, credentials, or secrets were involved
  Original task
  The user asked Codex to inspect a public GitHub repository containing skills for another AI product and explain how the system could be used.
  The upstream documentation described browser-based installation through the target product’s Skills interface. The implementation used /mnt/skills/user/... paths belonging to that environment.
  Observed behavior

1. Codex read the README but did not inspect the actual SKILL.md files before making an installation recommendation.
2. It replaced the author’s browser-based runtime with a local Claude Code project installation, although the user had not requested a port or adaptation.
3. It produced an execution brief instructing another agent to install the system into .claude/skills.
4. A subsequent static audit found that 34/34 SKILL.md files contained hard-coded /mnt/skills/user/... references. The proposed installation could not work without modifying the whole system.
5. The external auditor correctly stopped the installation with a BLOCK result.
6. When asked to report the incident, Codex initially opened its in-app browser instead of Chrome, despite the workflow requiring the user’s existing Chrome session.
7. Codex directed the user to the general OpenAI Help Center chat without first confirming that it could create or escalate a product bug.
8. The Help Center assistant said it could not create a case or provide a case number and incorrectly suggested “Safety or Legal concern” and the Report Content form.
9. The official Report Content form is intended for policy violations or illegal content, not product-quality failures.
10. Codex found the openai/codex GitHub tracker only after repeated user corrections. New issue creation was restricted, requiring this report to be posted as a Discussion instead.
    Impact

• The original task was not completed by Codex.
• The user had to redirect the work to another agent and repeatedly correct Codex.
• Time was spent auditing an incompatible installation path.
• An incorrect execution brief and project decision were created.
• The reporting process itself became a second example of incorrect tool and channel selection.
• No incompatible installation occurred because the external audit stopped it.
  Expected behavior
  Codex should:
• treat the upstream installation method and runtime as the project contract;
• inspect concrete paths, dependencies, hooks, MCP configuration, and runtime assumptions before recommending installation;
• distinguish general format compatibility from compatibility of the actual implementation;
• require explicit user authorization before proposing a port to another runtime;
• use the browser surface implied or explicitly selected by the user;
• verify a reporting channel’s declared purpose before directing the user there;
• never classify an ordinary product-quality report as a Safety or Legal concern;
• clearly state when no verified escalation channel or case number exists.
  Related reports
• Default prompt stack contains contradictory behavioral directives #27587 — contradictory behavioral directives and premature assumptions
• Model not respecting agents instructions #29879 — model not respecting agent instructions
• Let explicit user tool/MCP choice override auto-triggered bundled skills #36741 — explicit user tool choice should override automatic routing
  This incident combines those patterns with runtime substitution, insufficient source inspection, browser misrouting, and failure to identify a functioning product-feedback channel.

[tittlepl-baj-jk]

Clarification: this discussion is itself part of the reported failure

This discussion appeared because the user instructed Codex to document the incident and submit a formal product report to OpenAI.

Instead of identifying and using a verified internal reporting mechanism, Codex repeatedly substituted other channels:

1. It opened the wrong browser surface.
2. It directed the user to the general OpenAI Help Center chat without confirming that the chat could register or escalate a Codex bug.
3. The Help Center assistant incorrectly suggested a Safety/Legal reporting path and the Report Content form, which is intended for policy violations or illegal content.
4. Codex then proposed a GitHub Issue, but new issue creation was restricted.
5. Finally, Codex prepared this GitHub Discussion as another substitute for the requested formal report.

A GitHub Discussion is a public community conversation. It is not a formal OpenAI product report, does not provide a Feedback ID or case number, and does not confirm that the incident reached the responsible product team.

Therefore, the existence of this discussion should not be interpreted as successful completion of the user’s request. It is additional evidence of the same underlying problem: Codex repeatedly replaced the requested and verifiable outcome with an approximate alternative, presented that alternative as appropriate, and required the user to identify and correct the mismatch.

Only later did Codex inspect the installed application and identify its native /feedback mechanism, which uploads diagnostics and returns a Feedback ID. That mechanism should have been verified before directing the user through Help Center, Safety/Legal forms, GitHub Issues, or GitHub Discussions.

This comment is added to preserve an accurate record of why the discussion was created and to make clear that the discussion itself is part of the incident, not its resolution.