LongCat 2.0 API bug breaks Hermes when reasoning effort is set wrong
LongCat-2.0 is a low-cost API model with caching that keeps usage costs very cheap, plus a plan, similar to Deepseek's offerings, making it a regularly used option. Its capability sits somewhere between Deepseek-4-Flash and Deepseek-4-Pro, priced accordingly.
However, setting the parameter to anything other than "high", "med", or "low" causes the LongCat API to send malformed responses. This results in an empty response stream with no finish_reason, producing a cryptic error in agent.log such as "Provider returned an empty stream with no finish_reason." When this happens, the detects the failure and automatically falls back to a different provider.
Key points
- Setting to a value other than high/med/low breaks the LongCat-2.0 API
- Symptom: empty response stream with no finish_reason, triggering Hermes's automatic switch
- LongCat-2.0 offers cheap via caching and a plan, with capability between Deepseek-4-Flash and Deepseek-4-Pro
- Fix: explicitly set to one of high, med, or low