Bug: Response cut off on Typingmind despite the parameter max_completion_tokens was sent

[rustic pelican]

I’m experiencing a serious issue (or potentially multiple issues) that has rendered Typingmind almost unusable. Could you please investigate this as soon as possible and advise on any workarounds?

Details:

• Response Cut Off: Responses are limited to 5000 tokens (see attached 1) despite the max token setting being 50000.
• Timing: The issue began about one day after I changed the max token and default model settings. Is it related to prompt cache? If so, I’d like to know how to force update cache.
  I’ve attached a HAR file for reference. Please let me know if you need more information.

---

Response from Typingmind:

Regarding the cut off issue, I can see that TypingMind has correctly sent the parameter max_completion_tokens=50000 to OpenRouter. My best guest is that Open Router doesn't respect this parameter. You have two options in this case:

• Option 1: Modify your custom model to add a custom body parameter using max_tokens=50000(my understanding is that OpenRouter is using the deprecated parameter by OpenAI, but I could be wrong, only OpenRouter team can answer this)
• Option 2: contact OpenRouter team to make their API compatible with OpenAI API and accepting the max_completion_tokens parameter.

[gaunt bough]

OR doesn't set a max token limit by default. o3 naitvely has a max output of 100K tokens https://openrouter.ai/openai/o3

[rustic pelican]

So, you meant, somehow at some point, the max token of o3 was set to 5000, or else we'd never observe the cut-off?

[uneven jewel]

Have you tried not setting max tokens at all?

[rustic pelican]

"Not setting" is not very useful for me, because I need large max token.

[uneven jewel]

Then have you tried typingmind's suggestion? (Add custom body parameter?)

[jolly pulsar]

as far as i can tell, openrouter doesnt seem to document max_completion_tokens, but does support max_tokens