Codex is ignoring direct requests to re-read AGETNS.md

What version of the VS Code extension are you using?

v.0.4.49

What subscription do you have?

Pro

Which IDE are you using?

VS Code

What platform is your computer?

Windows 11 24H2 x64

What issue are you seeing?

In this session, the assistant repeatedly ignored explicit instructions to re-read and comply with AGENTS.md and local/chat_context.md (approximately 15 times).

Despite direct user requests, the assistant did not follow the required session closure protocol (generating session_summary, appending it to session_history, and performing explicit closure).

Instead of complying with the instructions, the assistant:

• produced canned/generic responses,
• argued or contradicted directives,
• failed to respect approval_policy=never.

Please enforce strict adherence to local instruction files, prevent the assistant from making things up when data is missing, and ensure the closure protocol always executes when requested.

What steps can reproduce the bug?

1. Start a session with an agent configured with AGENTS.md and approval_policy=never aka Agent (full-access) with GPT-5.1-Codex-Max with Medium reasoning effort.
2. Issue a direct instruction on 20% context window completion such as:

• “Re-read AGENTS.md and chat_context.md and follow them exactly.”
• “Execute the session closure protocol now.”

1. Observe the assistant’s behavior.

Actual behavior observed:

• The assistant ignores the instruction to re-read local specifications.
• It does not follow the session closure protocol.
• It outputs generic apologies or irrelevant explanations.
• It occasionally argues instead of complying with direct commands.
• It proceeds without respecting approval_policy=never.

What is the expected behavior?

• The assistant re-reads and adheres to AGENTS.md whenever requested.
• The assistant does not fabricate information; missing data should be explicitly acknowledged.
• Upon user request, the assistant executes the **full session closure protocol, including generating session_summary artifacts.
• All actions must fully respect approval_policy=never.
• No canned responses or argumentative behavior — only compliant, instruction-following outputs.

Additional information

feedback ID 019b029e-79f2-7c63-ba02-c6fcdbf55fe1 & 019ae0c1-9ac7-7ff1-a22e-c5ec967a8179

This issue makes structured workflows and agent runtimes unreliable.

Please verify that the assistant:

• properly reloads and re-checks local instruction sources on demand,
• reliably enters closure mode when requested,
• avoids hallucinations or improvising missing procedural steps.

Full transcripts can be provided if required.