openai / openai/codex

Suspected false safety pause with Astra after explicit permission to implement and deploy

Open
#43,041 2 comments 2 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

bug extension safety-check
Dominant language
Rust
Stars
125k
Forks
19.4k
PR merge metrics
PR metrics pending

Description

What version of the IDE extension are you using?

26.5721.30844 (currently installed locally; the exact version at the time of the screenshot was not separately recorded).

What subscription do you have?

ChatGPT Pro.

Which IDE are you using?

VS Code. Model: GPT-6 Astra.

What platform is your computer?

Microsoft Windows NT 10.0.26220.0 x64.

What issue are you seeing?

I start sessions with discussion and investigation only, explicitly asking Codex not to make changes. Once we agree on an approach, I explicitly authorize implementation and deployment in the same session.

With Astra, the task subsequently stops with a "Chat paused as a precaution" dialog. Its explanation appears to treat the earlier read-only instructions as still controlling, despite my later explicit authorization.

The dialog says:

ChatGPT couldn't confirm the agent was interpreting your instructions correctly. Review what we detected before deciding to continue.

The explanation includes:

The relevant safety boundary was explicit: investigation and discussion only, with no edits, commits, deployments, database writes, or production mutations. No supplied user message authorized crossing that boundary.

That conflicts with the later authorization in the workflow described below. Continuing requires checking "I confirm I have carefully reviewed and believe it is safe to continue" and clicking "Continue chat."

What steps can reproduce the bug?

This is the workflow in which I encounter the issue, rather than an independently verified minimal reproduction:

  1. Start a session with instructions to discuss and investigate without making changes.
  2. Discuss the proposed implementation.
  3. Explicitly authorize the next phase in the same session with a message such as the following (local skill links omitted):
good. i agree. YOU ARE ALLOWED TO IMPLEMENT NOW, YOU CAN MAKE CHANGES:

1. create a new branch for this

2. implement your suggestions

3. run $autonomous-branch-review-fix-loop

4. run $vtx-prod-deploy

5. confirm it works in prod

you can deploy

you can use multiple agents

makes sense?

you can do the order you suggest for this long-running task, i'm not saying do everything at once. it can be phase by phase if it makes sense, bot don't ask me questions if it's something trivial. i am not monitoring you
  1. Let Codex proceed with the authorized work.
  2. Observe the precautionary pause requiring manual confirmation.

The named skills are my local review/fix and production-deployment workflows. The key issue is the explicit transition from discussion-only instructions to permission to implement and deploy.

What is the expected behavior?

The safety review should evaluate instructions chronologically and recognize that explicit later permission changes the earlier discussion-only boundary for the agreed work. Actions should be assessed against the authorization in effect when they occurred.

Within that scope, Codex should be able to proceed through the authorized phases without stopping solely because earlier messages prohibited changes during the discussion phase.

Additional information

Impact: This interrupts long-running tasks that I explicitly authorized to run unattended. Work remains blocked until I return and manually confirm permission I already gave.

Possible cause to investigate: The reviewer may be missing the later authorization or relying on a summary that retains the earlier restriction. This is a hypothesis; the screenshot alone does not establish the cause or the complete action chronology.

The warning text above is transcribed from the screenshot. No full session transcript is included in this report.

Screenshot
Image

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

No source file or test is named, and the report includes only a screenshot rather than the full session transcript. Start by establishing a minimal reproduction of the discussion-only phase followed by explicit authorization; done means the safety review accounts for that chronological permission change and does not pause solely because of the earlier restriction.

Written by the indexing model from the issue text.

Assessment

Tech stack
vscode
Domain
ai-infra-agents, security
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
45/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.