[Feature] Considering adding pruning thinking token

[Steve37vn]

Problem:
When using reasoning models (e.g., Gemini 3 Pro/Flash, Claude Opus 4.5 with extended thinking), thinking tokens consume context window space but provide no utility once the final response is generated.

Proposed Solution:

Detect reasoning blocks across providers:

• OpenAI: reasoning field in response
• Anthropic: tags
• Other providers as needed

Dynamically prune thinking tokens from context before subsequent requests

[Tarquinen]

Hey, thanks for the suggestion. This is a good idea, but also a very complicated one. Pretty much all providers handle reasoning parts in their own unique way, so implementing this would be very fragile and require a lot of edge cases. For example, Anthropic extended thinking models require the last assistant turn to have reasoning at the start, and they also recommend passing back all thinking blocks as their API automatically filters unnecessary reasoning blocks (I think this automatic filtering is pretty common for non interleaved thinking models, mitigating the benefit of this feature request). They also require consecutive thinking blocks to match the original output, they can't be modified. See more here: https://platform.claude.com/docs/en/build-with-claude/extended-thinking

I'm pretty hesitant to do this because it would require a lot of research to get proper support for all providers, and would create significant maintenance overhead for any future provider changes. If someone wanted to PR this feature and commit to maintaining I would definitely be open to that.

Let me know what you think!

[huypham37]

Opencode has a mechanism to detect and render the thinking tokens of agents, I think we can use that. I will do research on this, can you assign this issue to me?

[Tarquinen]

sure

[BouquetAntoine]

Would also love to see that ! Anthropic have a beta header for that "context-management-2025-06-27" which then allows to use clear_tool_uses_20250919 or clear_thinking_20251015 (https://platform.claude.com/docs/en/build-with-claude/context-editing)
Never tried DCP + clear_thinking_20251015 together but a 100% compatible solution with all providers inside DCP would be fire !

[Tarquinen]

Think this is doable only for interleaved models, should have access to model reasoning capabilities now that the model object is passed into the transform hook. Not sure how to test negative performance impact due to this in interleaved models, or how many messages of reasoning blocks to keep...