Safety block persists for explanation-only follow-ups in a Paseo-managed session

Problem

A persisted Codex session used through Paseo repeatedly fails with:

This request was blocked by our safety systems. Reason: Potentially unintended activity.

After the initial block, follow-up requests asking only for an explanation or recovery guidance also returned the same system error. The user had explicitly authorized the intended local operation, but we cannot determine whether the block was a false positive or what specific action triggered it.

Observed diagnostics

• Codex session ID: 01a07f48-dac1-7100-8acb-eb3aebfa5e80
• Most recently observed failure: 2026-09-09 01:06:50 UTC
• Provider/model reported by Paseo: codex / gpt-6-astra
• Reasoning: high
• Permission mode: auto-review
• Agent lifecycle: error; active turn: null; pending permissions: []
• Local installed CLI reports: codex-cli 0.153.4 (the exact binary used by the affected provider process has not been independently verified)
• Client: Paseo Desktop on macOS; local Paseo CLI reports 0.7.2
• Repeated follow-ups included "retry" and requests for instructions on what to do next.

Expected behavior / request

Please review the possible false positive and advise on supported recovery. Ideally, a blocked operation should still permit an explanation and a safe recovery/help response, or provide an actionable diagnostic/reporting path. We are not requesting a bypass of safety checks.

Privacy

This report is submitted at the user's explicit request, with explicit approval for this public destination and these minimal diagnostic fields. No original transcript, screenshots, sensitive business data, personal records, credentials, or raw logs are attached. Only minimal session/error diagnostics are included.

No repository file, test, or entry point is identified in the report. Start by determining whether the behavior is implemented in this repository or supplied by the provider, then reproduce the blocked-session follow-up behavior using the listed diagnostics; done means identifying the responsible component and documenting a supported recovery or reporting path.