fix: exclude k2p5 from reasoning variants

[neavo]

Summary

• Add k2p5 to the exclusion list in variants() function to prevent Kimi K2.5 from incorrectly getting high/max thinking variants

Details

The Kimi For Coding provider uses k2p5 as the model ID for Kimi K2.5, which does not contain the string "kimi". This causes it to bypass the exclusion logic and incorrectly receive thinking variants support.

Confirmed with Moonshot official staff: Kimi-K2.5 does NOT support thinking budget configuration.

This is a temporary workaround. The correct solution is to fix the model ID in models.dev data to use kimi-k2.5 instead of k2p5.

Fixes #11917

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

No duplicate PRs found

[rekram1-node]

wait so kimi for coding has 2 model ids that map to same thing?

[neavo]

wait so kimi for coding has 2 model ids that map to same thing?

In fact, the model ID obtained via the https://api.kimi.com/coding/models endpoint is kimi-for-coding
which maps to the actual model displayed in the user console (currently K2.5)
I am not quite sure why the model ID provided in models.json is k2p5

[rekram1-node]

game to update it, I don't have a kimi for coding sub so I can't test unless I grab one

[luisf371]

Please revert this change or modify to allow thinking. Kimi Code documentation is limited but the below screenshots show non-thinking (normal) and thinking (high or max) based on the variant. I conducted several runs and normal variant consistently was faster making me believe it not a matter of reasoning tokens being hidden in the background.

[neavo]

I directly asked Kimi's official staff, and the conclusion I got was: K2.5 does not support reasoning configurations.

[rekram1-node]

@luisf371 they should not make any difference, it was a visual thing only their api doesnt support the variants

[luisf371]

That's fair but i think we are getting confused on the intent. Yes i understand that the Kimi Code API does not support control of the thinking token efforts (think alot, think little, etc) which is what "variants" typically control.
However, k2p5 support both thinking and non-thinking, and if its anything like the official API, its controlled thru ""thinking": {"type": "disabled"}" > https://platform.moonshot.ai/docs/guide/kimi-k2-5-quickstart#disable-thinking-capability-example
For whatever reason before the change on 1.1.50 - having the variant set to "none" was thinking disabled, and high/max was thinking enabled as showed in my previous post.

I even installed their official kimi client and verified that you can select between thinking and non-thinking.

Thank you for what you guys contribute!

[rekram1-node]

Yes that's true, we can add a non thinking variant

[AngDrew]

and if its anything like the official API, its controlled thru

so how can we control the thinking on / off in the new opencode version?

[rekram1-node]

#12090 (comment)