Deep Research repeatedly fails mid-report with “Resume stream unavailable” and cannot recover final output

What issue are you seeing?

Deep Research tasks repeatedly fail while generating the final report.

This is not a one-off. I have encountered the same failure pattern multiple times across different Deep Research runs.

Typical behavior:

1. Deep Research completes a substantial portion of the research.
2. It begins generating the final report normally.
3. The output stream stops partway through.
4. The UI may show “Resume stream unavailable”, indicate that the response is waiting for the rest of the message, or simply stop producing output.
5. Reopening the conversation or attempting to resume does not recover the missing remainder.
6. The completed research is therefore left permanently incomplete even though significant time/usage was already spent.

Most recent example:

• A multi-source Deep Research task comparing doctors/clinics, qualifications, reviews, safety information, prices, and value for money.
• The report successfully generated several sections and then stopped abruptly with “Resume stream unavailable.”
• The remaining report never became available.

I have seen this same class of failure on previous Deep Research tasks as well.

What steps can reproduce the bug?

1. Start a reasonably large Deep Research task that requires multiple web sources and a structured final report.
2. Let the research complete and wait for the final report generation.
3. In some runs, the report stream stops partway through.
4. The UI displays “Resume stream unavailable” or otherwise fails to continue the response.
5. Reopen or retry the conversation.
6. Observe that the missing remainder of the report cannot be recovered.

The issue is intermittent rather than guaranteed on every run, but it has reproduced multiple times across separate Deep Research tasks.

What is the expected behavior?

If the client loses the output stream during final report generation:

• the generated/in-progress report should remain persisted server-side;
• reopening the conversation should reload the already-generated report;
• the system should be able to resume from the last persisted boundary instead of losing the remainder.

A transient UI/stream failure should not make an expensive Deep Research run permanently incomplete.

Actual behavior

The report stream becomes unrecoverable and the final response remains truncated.

The research itself appears to have progressed substantially, but there is no way to retrieve the missing remainder after the stream fails.

Additional information

The most recent visible error was:

Resume stream unavailable

This has happened multiple times, not only on the latest task.

The workload was within the normal intended scope of Deep Research: multi-source research, cross-checking, comparisons, and a structured report. Splitting the prompt into smaller research tasks may reduce the chance of failure, but that is only a workaround; long multi-source research is a core Deep Research use case.

A useful reliability fallback would be to persist report-generation checkpoints or the completed report server-side so a stream/UI disconnect can be recovered without losing the output.

No files, tests, or entry points are named. Start by tracing Deep Research final-report stream handling and the server-side persistence and resume behavior described in the issue; done means an interrupted stream can reload or resume the generated report without leaving the final response truncated.