Headers altering / modifying

[SantiagoDePolonia]

@Neiko2002 said: "What I actually need is real-time observability and the ability to modify headers based on request headers, specifically to control and correct the behavior of existing agents. This kind of interaction is much more common in LLM proxy projects like UniClaudeProxy, CLIProxyAPI, cliproxyapi-plusplus, or llm-interactive-proxy"

[Neiko2002]

Thank you for opening this ticket. I would like to add some more information. One use case would be 'Thinking Budget Control' for Qwen, DeepSeek, and Nemotron models. The idea is that a model in thinking mode gets a 'thinking_token_budget'. If it exceeds this budget, it receives a 'reasoning_end_str' message from vLLM to wrap things up:

https://docs.vllm.ai/en/latest/features/reasoning_outputs/#online-serving

This would allow the creation of low, mid, and high-thinking routes for models that do not support it natively.