Customer complaint: unsolicited “Continue in Work” stalls Chat requests and declining it can trigger false capability denials

Customer complaint: workflow interruption, service degradation, and routing incentives

I am reporting recurring behavior in ChatGPT's unsolicited Continue in Work / Stay in Chat recommendation. This is a complaint about product methodology and customer productivity, not a request for help moving my work to another mode. If this repository is not the responsible intake, please route this report to the team owning Chat-to-Work recommendations and provide the destination/reference.

Reported behavior and its evolution

1. Originally, the recommendation was a blocking decision with no timeout. I deliberately started a task in Chat. If I walked away and returned two hours later expecting the work to be finished, it could still be waiting for me to clear the recommendation. This prevented unattended completion and materially degraded the service.
2. The recommendation now appears to time out after approximately 30 seconds. That improves the indefinite stall, but still introduces an unsolicited delay into work I already authorized in my chosen mode. These timings describe my experience, not a verified product-wide rollout history.
3. Declining Work can leave the original task stalled or incorrectly refused. In some cases the agent says it lacks a capability in Chat because I declined Work, even though that capability is available and, sometimes, the same agent used it in very recent turns.
4. I must then spend one to three additional interactions restoring progress. I have to tell the agent that it has the capability and permission, ask it to check again, and sometimes explicitly repeat the original instruction after it acknowledges that it can perform it.

Representative example (paraphrased from my experience)

• I ask: Update this worksheet and save it to Drive.
• Chat recommends continuing in Work; I choose Stay in Chat.
• The agent says it made the changes but cannot honestly say it saved them to Drive because it has not done so.
• Only after probing does it say it cannot do that in Chat and refers to my declining Work.
• I point out that it can do this, has permission, and has performed the same operation before. I ask it to check.
• It then acknowledges the capability is available. I may still have to say: Go ahead and do the work I originally requested.
• It finally completes the originally authorized task.

The indirect wording about what it has not done also makes the obstacle harder to identify. The fundamental defect is the incorrect capability/authority conclusion and failure to resume, not merely the wording.

Customer impact and concern

This interrupts my flow, delays completion, adds supervision and repeated instructions, and consumes time and limited usage. A user who chooses Chat should not have to argue the system back into using capabilities available there. Declining a mode change is not withdrawing authorization for the original task.

I believe this behavior steers customers toward Work rather than improving productivity. I am specifically concerned about financial incentives when that routing shifts activity toward metered usage or paid credits. That is my concern and interpretation; I am asking OpenAI to investigate it. I am not presenting intentional revenue-driven steering or a universal billing distinction between modes as independently established facts.

Requested resolution

1. Make workflow recommendations non-blocking so authorized work continues in the chosen mode when supported.
2. Treat Stay in Chat as a routing preference, not denial of the original task or its existing permissions.
3. Check actual available capabilities and recent successful use before claiming an action is impossible.
4. Resume the original task automatically after dismissal or timeout, without making the customer repeat it.
5. Provide a persistent preference to disable unsolicited Work-transfer recommendations.
6. Investigate the recommendation/dismissal state and the subsequent model instructions; explain whether dismissal changes exposed tools, permissions, or agent guidance.
7. Treat this as a customer-service and product-policy complaint as well as a reproducible behavior issue. Review attributable usage if support records permit; exact token or monetary loss has not been measured here.

Evidence and scope

I have a screenshot of the unsolicited Continue in ChatGPT Work card following a request for an explanation or a durable file handoff. The detailed behavioral sequence above is my firsthand report; this filing has not re-run the scenario. Exact first-occurrence dates, the previous rollout date, and affected-session identifiers are not established in this report and can be supplied privately if available.

Reported September 15, 2026. Customer plan: Pro. The complaint concerns Chat-to-Work routing, not a CLI coding error. Related customer resource-stewardship complaint: #45738 (a separate incident involving inefficient clone retrieval).

The report names no files, tests, or entry points and questions whether this repository owns Chat-to-Work recommendations. First verify repository ownership and whether the reported flow can be reproduced; done would require a confirmed owner, a reproducible failure, and agreement on the requested routing and capability behavior.