Context
Outbound prompt scanning via the prompt hook works well. The Stop hook fires at the end of a conversation, letting us scan the final inference for PII/harmful content after the model has finished "thinking."
Problem
The Stop hook is detection-only. It does not prevent the response from rendering in the Codex GUI — by the time we scan, the content is already on screen. There is no API to rewrite or redact a response; we can only surface a message to the user and influence subsequent chats in that thread.
Impact
Disallowed content (PII, harmful output) is shown to the user regardless of the scan result. We get an alert, not a block.
Asks
- Confirm the current Codex hook surface and document what Stop can and cannot do.
- Track an upstream feature request for response blocking/rewrite at the GUI.
- Codex hooks are new and incompletely implemented — capture which specific gaps block us.
Raised by a collaborator during Codex hook integration review.
Context
Outbound prompt scanning via the prompt hook works well. The Stop hook fires at the end of a conversation, letting us scan the final inference for PII/harmful content after the model has finished "thinking."
Problem
The Stop hook is detection-only. It does not prevent the response from rendering in the Codex GUI — by the time we scan, the content is already on screen. There is no API to rewrite or redact a response; we can only surface a message to the user and influence subsequent chats in that thread.
Impact
Disallowed content (PII, harmful output) is shown to the user regardless of the scan result. We get an alert, not a block.
Asks
Raised by a collaborator during Codex hook integration review.