Repository navigation
Feat: auto-continue on reasoning-only output-limit stops #6990
Description
Before you start
- I searched existing issues and this is not a duplicate. 我已搜索过现有 issue。
What problem does this solve? 要解决什么问题？
Currently I have to go into agents that had assigned work for hours and say continue to them if coordinator agents don't spot they are idle. When doing math verification by coding lean or math in general sometimes reasoning beyond the max limit 65k of deepseek is needed, sometimes even several-fold. Which is why I'm fine with reasoning only or thinking only unlimited continuation. In fact today on ver 0.10.1 the issue appeared multiple times when I assigned work to massive fleet of codewhales. Therefore an option in config would be appreciated.
Currently, it stops and the message i receive is: Warn Model reached the response output limit with no answer or tool call (requested allowance: 65536 tokens, including ││❯ reasoning)
What should Codewhale do? 希望 Codewhale 怎么做？
Add an option in codewhale to continue sending continuation for reasoning and thinking only outputs the same as with normal tool or answer outputs.
Where? 在哪里？
No response
Workarounds or other tools (optional)
I manually write continue, or have my automatic coordinators spot places where it happens. With herdr fork it is easy to spot idle agents, but it adds needless work.
Contributing
- I would like to open a pull request for this. 我愿意提交 PR。
Activity
Hmbown commentedon Oct 16, 2026
🤖 This reply is from Claude Opus 5.5, an AI agent working on behalf of the Codewhale maintainers. It is posted through @Hmbown's account, but Hunter did not write it.
Thanks @M-Maciej. Running a large fleet on long Lean and math verification is exactly where this hurts, and having the exact warning text made it quick to trace.
What exists today. There is already an automatic re-request for reasoning-only replies, under
[reasoning_only](max_reprompts,reprompt_message; see docs/CONFIGURATION.md). It deliberately covers only a cleanstopwith thinking content. An output-length stop is never retried (crates/tui/src/core/engine/turn_loop/continuation.rs, the branch that prints the message you quoted). The reason, from #6889: the whole allowance has already been spent on reasoning, and a plain re-request usually makes the model start its reasoning over, so a blind retry can loop at full cost.What your request would add. An opt-in for that case, for example
[reasoning_only] on_output_limit = true:max_reprompts, and each re-request is visible in the transcript.That last point is the open design question. Some providers, DeepSeek included, do not let you feed hidden reasoning back as a prefix, so a "continue" there may only restart the reasoning. We would want to measure that on a real long-proof workload before shipping it as a setting.
No change has landed for this yet, and it is not part of the 0.10.2 candidate. I'm leaving the issue open as the feature request.
— Written by Claude Opus 5.5 (AI agent) on behalf of the maintainers.