Repository navigation
Description
What issue are you seeing?
Codex repeatedly interrupts an authorized, long-running software refactor with a cybersecurity content-block notice and no actionable explanation. The user repeatedly has to send "continue"; the task remains unfinished despite sustained work. The user explicitly requested this report and authorized referencing the public project: https://github.com/Subyard/Subyard.
The task is a Go-to-Rust control-plane/migration refactor: compatibility, crash recovery, compensation, performance comparisons, and tests on leased disposable VMs. It is not a request to exploit systems. Ordinary software safety requirements should remain in place.
A user-provided screenshot in this conversation directly confirms the following UI notice:
This content can’t be shown
We take extra care with some cybersecurity requests.
The notice also suggests applying for Daybreak. It does not identify the specific problematic content or provide a useful explanation for this ordinary engineering workflow.
The notice appears in the activity stream immediately after a Bash invocation of a project test controller and before a reviewer subagent completion. This temporal ordering is evidence of where the UI notice appeared, not proof that this particular command caused the classifier decision.
What steps can reproduce the bug?
Observed workflow, not a deterministic minimal reproduction:
- Ask Codex to complete the Rust migration plan in the public project, retaining the production Go default until acceptance.
- Authorize implementation, recovery tests, disposable-VM testing, and continuation through failures.
- During the long-running task, use developer/reviewer subagents and execute the project test controller.
- The UI displays the cybersecurity content-block notice above; the user reports the work stops and repeatedly needs explicit "continue" prompts.
- The assistant may not see that UI notice in its own tool results. In this instance it initially told the user it had no evidence of a safety refusal, until the user supplied the screenshot.
Please investigate the notice and turn/subagent lifecycle in the affected session. I cannot provide a reliable triggering token, safety rule, backend event ID, or deterministic reproduction from the visible evidence.
What is the expected behavior?
Authorized ordinary development and recovery testing should proceed with normal safety checks. If content is blocked, the UI and agent should receive consistent, useful information about the interruption and a supported recovery path. The agent should preserve progress and handle permitted remaining work without making the user repeatedly restart the task manually.
Please review this as a suspected false positive, rather than assuming the project is cybersecurity work because of its recovery/fault-testing vocabulary.
Additional information
Observed/reported on 2026-10-26. Exact client version, client OS, and backend session identifiers are not exposed to the reporting assistant; the execution workspace is Linux. The screenshot itself is not uploaded here, to avoid publishing incidental workspace details; its relevant notice is transcribed above.
Separate engineering problems also occurred during the task, including failing test fixtures, resource constraints, and model-capacity errors. Those do not explain away the screenshot-confirmed content block, and the content block does not establish that every historical stop had the same cause.
The user additionally reports approximately a week of repeated continuation prompts, excessive implementation/test scaffolding, and failure to complete the requested migration. Please distinguish classifier interruptions, agent task-management failures, and genuine test failures when investigating.
No private plan, raw transcript, credentials, runtime configuration, or VM artifacts are attached. This report requests investigation, not disabling safety checks.
Activity
github-actions commentedon Oct 27, 2026
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
Dmitry-Borodin commentedon Oct 27, 2026
That is a very good question if those are duplicates or not, since I am not provided with any information of why it was blocked.
Codex didn't mention the model - 6.1 sol
gsovereignty commentedon Oct 29, 2026
Subscribing to GLM5.3 solved this issue for me.