Use the SFT + RL CAPI proxy model to power the Search Subagent

[24anisha]

Dogfooding the fine-tuned Qwen3-4b model for the search subagent

[Copilot]

Pull request overview

This PR updates the Search Subagent feature to use a fine-tuned Qwen3-4b model accessed via a CAPI proxy endpoint. The changes replace the dynamic endpoint selection logic with a hardcoded endpoint for the specialized agentic-search-v1 model.

Changes:

• Created a new ProxyAgenticSearchEndpoint class that connects to the fine-tuned model via CAPI proxy
• Simplified SearchSubagentToolCallingLoop to always use the specialized endpoint instead of dynamically selecting one
• Enabled the Search Subagent feature by default (changed configuration default from false to true)

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
src/platform/endpoint/node/proxyAgenticSearchEndpoint.ts	New endpoint implementation for the agentic-search-v1 model with tool calling capabilities
src/extension/prompt/node/searchSubagentToolCallingLoop.ts	Removed dynamic endpoint selection logic, now always uses ProxyAgenticSearchEndpoint; removed unused endpointProvider parameter
src/platform/configuration/common/configurationService.ts	Changed default value for SearchSubagentToolEnabled from false to true
package.json	Updated github.copilot.chat.searchSubagent.enabled configuration default to true

[Copilot]

The endpointProvider parameter is no longer used in this class after the refactoring to always use ProxyAgenticSearchEndpoint. This parameter should be removed from the constructor entirely, along with the corresponding parameter in the super() call on line 49. The base class ToolCallingLoop still requires it, but this derived class doesn't need to accept it as a constructor parameter.

[Copilot]

The tokenizer type O200K is typically associated with OpenAI's GPT-4 and GPT-4o models. According to the PR description, this endpoint is for a fine-tuned Qwen3-4b model. Verify that O200K is the correct tokenizer for this model, as Qwen models typically use their own tokenizers. If the model was fine-tuned to be compatible with O200K tokenization, this is acceptable, but it should be documented why this choice was made.

Suggested change

	tokenizer: TokenizerType.O200K,
	tokenizer: TokenizerType.Unknown,