OpenAI dot paused and cannot resume: recovery steps and open source options

What the pause message says about your dot

A dot is an always-on agent inside ChatGPT. The banner "Your dot paused out of precaution" is its safety review holding the session. The explanation underneath is specific: "Auto-review did not approve some recent actions, so your dot was paused." The GitHub issue that quotes both lines reports the strange part. The dot kept working. The profile still showed "Active", its computer still showed "Connected", and typing was blocked on desktop and mobile.

The pause is a judgment about actions the dot already took. A review layer over the running session applies it. The session itself is often still alive. That is why the message can sit on a dot that looks busy everywhere else, and why a refresh changes nothing.

The recovery steps that exist, in order

Four moves appear in public reports. Each has a different job. None changes the review decision behind the pause.

1. 1.Reload the page. The reporter in the 409 thread tried this more than once, including a retry after a fresh load on 2026-10-02 (UTC). The same error returned.
2. 2.Open the dot at chatgpt.com/dots on desktop web. OpenAI Support asked that reporter to try it there on 2026-10-18, with the same account and workspace. The request asked whether the pause still appeared. The thread records no reply.
3. 3.Restart the dot. This is the one recovery a public report confirms. The GitHub reporter wrote that restarting the dot restored interaction. They added that the cause of the pause and of the failed Resume remained unconfirmed.
4. 4.Wait out an abuse-prevention pause. A second report describes a Portuguese notice. It says the account had reached an abuse-prevention limit and should try again shortly. The same reporter's earlier incident on October 11 was observed recovering about 1 hour 32 minutes after the first record. They did not treat that interval as a full outage measurement.

A person with one dot and no infrastructure can live inside this loop: retry, wait, restart, and keep the work small enough to redo. The loop stops being reasonable when a team depends on the dot's output every day.

When the Resume button itself fails

The banner's own button can fail with a generic "Something went wrong". The same reporter captured the request that button generates and the reply it receives. The thread keeps both:

POST /backend-api/tbo/<redacted>/runtime/resumeContent-Type: application/json {"clear_safety_flag":true} HTTP 409{"detail":"Your dot's safety failure has changed"}

The state beside that call is the tell. While the request failed, the dot profile read "status": "active", "isSafetyFlagged": true, "safety_flag": "guardian_circuit_breaker", "is_paused": false. The button was trying to clear a safety flag. The profile already called the dot active and not paused. Desktop attempts during the same recurrence also logged 409 replies reading "Your dot's runtime is not idle" and "Orbit safety state changed".

The reporter could not establish the backend cause. The public record does not supply one. What the capture does establish is narrower and more useful: the control surface and the stored state disagree. Pressing the same button again sends the same request. A retry cannot clear a flag whose state has moved underneath it.

What happens to work already in flight

The GitHub report is the closest thing to an answer. It is uncomfortable. The dot continued working while its owner could not send an instruction from either app. The reporter said plainly that they had not established which background operations ran, or what side effects those operations had.

The second reporter asks the operational version of the same question: which work is finished, which is pending, and what a pause does to scheduled tasks. Neither thread shows a way to move a dot's memory, its schedule or its task list into another system.

That is the migration question arriving early. A pause is also the moment a team finds out how much of the job lives in a console it rents. For a single task the cost is redoing the work. For a weekly report other people wait on, it is the thing to fix before the next pause.

Keep waiting or move the job

• —A one-off task, one person. Wait and retry. The restart restored one reporter's dot, and for a single task the delay costs less than moving the job.
• —A recurring job a team depends on. Move it. A pause stops the team that depends on the output, and the recovery loop has no owner you can page.
• —Regulated or client data. Move it, and put the rules where your reviewers already look. A gate you cannot read is a gate you cannot explain to an auditor.

The grant is where Kortix answers this question. Session work reaches main through a change request. Merge is default-deny for agents. Approval gates you set. Off until you set them. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free.

What to set up before the next pause

The always-on example is small enough to read in one sitting. It is the shape to copy.

The agent file carries its own permission block. In agents/feedback.md, bash is ask by default. git status, git diff and git log are allowed, and rm -rf * is denied. The manifest adds the grants: the agent may read its own session and submit work for review, and project.gitops.merge is left out so a person merges every change request.

Two files, one review step, and a gate a person owns. When the agent proposes something wrong, you read the diff. When the gate trips, you read the file that set it, in the same repository, at the commit that changed it. The gate exists on both platforms. The difference is whether you can diff it.

On a first setup, the self-hosted walkthrough covers the install. The comparison puts the two ownership models side by side. If the job recurs, start with a schedule that fires once a week. A weekly run with a review step is easier to trust than a process nobody watches.