Enhance support for non-Qwen models and providers

What happened?

I saw a PR addressing an issue with kimi-k2 compatibility via Moonshot's API. This is great for users like me who love Qwen Code but need to use other models. If we are considering improving support for other models, I have a few more issues to raise:

1. The auto-compress action in qwen-code seems to send the value of max_tokens to be something larger than 8k and the DeepSeek server rejects any requests where max_tokens value is set larger than 8k. Maybe allowing building the summary over multiple requests should be fine.
2. Auto-compression virtually fails for GLM models via z.ai server. The returned summary is usually just around 300 tokens in size, clearly far too less and all context is indeed lost afterwards. Not sure what causes this but maybe max_tokens is the culprit here as well.

What did you expect to happen?

Non-qwen models to work very nicely with qwen-code and being able to use qwen-code as my default go-to driver for everything instead of having to use other tools for non-Qwen models.

Client information

0.1.12

Login information

No response

Anything else we need to know?

No response

Start by reproducing auto-compression with DeepSeek and GLM through the mentioned providers, using client version 0.1.12, and inspect the max_tokens values and returned summary sizes. Done means non-Qwen models can complete compression without rejected requests and retain an adequate summary across these providers.