feat: add Anthropic extended thinking support

[jordibeen]

Summary

This PR implements Anthropic's Extended Thinking feature in compliance with their official documentation. The feature enables Claude models (Sonnet 4, Opus 4.1, Haiku 3.5) to display their reasoning process wrapped in <think> tags before providing the final response.

Changes

• Full support for Anthropic's thinking feature with configurable token budgets
• Proper <think> tag wrapping for clear visual separation
• Tracks processing state between thinking and regular content blocks
• Prevents thinking blocks in chat topic generation
• Removed legacy line.delta.text fallback (superseded by proper text_delta handling)
• Removed legacy anthropic-beta: messages-2023-12-15 header
• Isolated to Anthropic models without affecting other providers

Technical Implementation

1. Payload Preparation (dispatcher.lua)

• Added thinking_budget parameter support for Anthropic models
• When thinking_budget is present in model configuration, adds thinking object to API payload:

  payload.thinking = {
    type = "enabled",
    budget_tokens = model.thinking_budget
  }

2. Response Stream Parsing (dispatcher.lua)

Enhanced Anthropic response parser to handle new thinking block events:

Event Type	Action
content_block_start with type: "thinking"	Outputs <think> tag
content_block_delta with type: "thinking_delta"	Outputs thinking content
content_block_start with type: "text"	Closes thinking block with </think>
content_block_delta with type: "text_delta"	Outputs regular response content

Usage Example

Example configuration of using Sonnet 4 with thinking budget tokens:

{
	provider = "anthropic",
	name = "ChatClaude-Sonnet-4-Thinking",
	chat = true,
	command = false,
	model = {model = "claude-sonnet-4-20250514", thinking_budget = 1024 },
	system_prompt = require("gp.defaults").chat_system_prompt,
}