feat: add support for reasoning models (K2-Think, DeepSeek R1, QwQ)

Add support for reasoning models that return responses in the
reasoning_content field instead of the standard content field.

Changes:
- Add reasoning_content field to response parsing
- Fallback to reasoning_content if content is empty
- Add reasoning_effort parameter for K2-Think models
- Strip special tokens (<|im_end|>, <|endoftext|>, etc.) from output

This enables compatibility with:
- LLM360/K2-Think-V2 (with vLLM)
- DeepSeek R1
- QwQ and other reasoning models

Tested with K2-Think-V2 on vLLM with --tool-call-parser hermes.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

pkg/providers/http_provider.go

@@ -198,6 +198,15 @@ func (p *HTTPProvider) parseResponse(body []byte) (*LLMResponse, error) {
 		content = choice.Message.ReasoningContent
 	}
 
+	// Strip special tokens that some models include in their output
+	content = strings.TrimSpace(content)
+	specialTokens := []string{"<|im_end|>", "<|endoftext|>", "</think_fast>", "<|end|>"}
+	for _, token := range specialTokens {
+		content = strings.TrimSuffix(content, token)
+		content = strings.TrimPrefix(content, token)
+		content = strings.TrimSpace(content)
+	}
+
 	return &LLMResponse{
 		Content:      content,
 		ToolCalls:    toolCalls,