Question: recommended global DCP thresholds for GPT-5.4 / GPT / DeepSeek / GLM?

[smithyyang]

Question

I'm trying to find practical recommended global settings for DCP minContextLimit / maxContextLimit, especially when mixing several models.

Right now I mainly use:

• GPT-5.4
• other GPT-family models
• DeepSeek models
• GLM models

Problem

With GPT-5.4 in particular, DCP feels too aggressive for me with default behavior. It starts pushing compression earlier and more often than I would like.

I understand that:

• percentages are supported
• per-model overrides exist
• very high percentages on 1M-class models may be too late

But I still haven't found a clear answer to this question:

What global thresholds do you actually recommend as a good starting point for users who mix GPT / DeepSeek / GLM models?

What I'm looking for

I would really appreciate guidance like:

• a recommended global range, for example whether something like 25% / 35%, 30% / 40%, or another pair is a better default starting point
• whether GPT-family models usually need more relaxed settings than DeepSeek / GLM
• whether you recommend global percentage settings first, then adding per-model overrides only if needed
• whether there are any known good defaults specifically for GPT-5.4 if it feels compression-happy

My environment

• OpenCode: 1.14.33
• DCP: @tarquinen/opencode-dcp@latest (currently resolves to 3.1.9 in my environment)
• Main model often used: gpt-5.4

Why I'm asking

There are many discussions about DCP being too aggressive or about large-context percentages being tricky, but I haven't found a concrete, maintained recommendation for:

1. global settings that work reasonably well across multiple model families, and
2. what people should try first when GPT-5.4 feels too eager to compress.

Even a short maintainer recommendation like "start with X/Y globally, then override GPT-5.4 to A/B if needed" would be very helpful.

[Tarquinen]

This really depends a lot on your workflow, generally I like to keep my context under 200k on any model as I think almost all models get significantly worse beyond that. You may have use cases for requiring more, so I can't give a solid recommendation that can apply to anyone.

Generally a good rule of thumb I think is that you want your context all relevent to your current task and not to exceed something like 200k or quality drops off and it gets expensive.

[Tarquinen]

Feel free to message here or create another issue if you have anything else