Summary

When using Claude Opus 5.5 with extended thinking on the official Factory subscription, any single
response whose thinking phase runs longer than 600 s is cut off by the host with:

The response exceeded the 600s host limit and was cut off. Ask for a shorter response
or split the task into smaller steps, then try again.

The stream ends without a message_stop and all thinking done so far is discarded. Droid then
retries the same request once, which starts thinking from scratch, hits the same 600 s limit
again, and finally stops with the error above.

So each attempt costs ~20 minutes and twice the usage, yet produces no output. Retrying the
task by hand repeats the same cycle and keeps burning usage.

Environment

• Factory Desktop: 0.202.0
• Platform: Windows 10.0.26200 x64
• Model: Claude Opus 5.5 with extended thinking
• Reasoning effort: medium/high
• Plan: official Factory subscription (no BYOK, no custom proxy)

Steps to reproduce

1. Select Opus with high reasoning effort.
2. Give a large single-shot generation task, e.g. "Create a single self-contained HTML file
   with a detailed Three.js rocket launch scene" (with a long list of requirements).
3. The model plans extensively in one thinking block before producing any output.
4. At ~600 s the response is cut off.
5. Droid automatically retries once; the retry starts over and is cut off again at ~600 s.
6. After ~20 minutes in total, the error above is shown and the turn stops.

Expected behavior

Long-thinking models should be able to finish a response, or at least not lose all progress
and usage.

Actual behavior

• Hard cut at 600 s, even though thinking tokens are still being generated.
• All thinking progress is lost.
• The automatic retry repeats the identical request, so it almost always hits the same limit,
  doubling the time and usage spent.

Suggestions

Any of the following would help:

1. Raise the host limit for models with extended thinking (e.g. 30 min), or make it
   configurable.
2. Make the limit an idle timeout rather than a total-duration cap, as long as thinking
   deltas or pings are still streaming.
3. Do not automatically retry the identical request after this specific error; it is
   almost guaranteed to fail again. Show the error right away instead.
4. When a response approaches the limit, have the agent prompt the model to work
   incrementally (emit tool calls or partial output) instead of failing the whole turn.
5. Document the 600 s limit so users can adjust reasoning effort and task size.