Record model thinking/reasoning

[dannykopping]

OpenAI and Anthropic's models may respond with "thinking" or "reasoning" blocks before a tool call is determined. Capturing these "thoughts" may provide auditors with key information about why a particular tool was called.

Here is some research about how these providers support this feature.

This will introduce a new recording primitive:

type ModelThoughtRecord struct {
	InterceptionID string
	ToolUsageID    string // This is an optional relation to a `ToolUsageRecord`.
	Content        string
	Metadata       Metadata
	CreatedAt      time.Time
}

We only care about these thoughts when a tool call is to be invoked. We don't capture regular inference output, and these thoughts are just a subset of that.

Acceptance criteria:

• Model thoughts/reasoning is recorded by all interceptor implementations
  • Messages
    • blocking
    • streaming
  • Chat Completions
    • blocking
    • streaming
  • Responses
    • blocking
    • streaming
• Recorder types added
• Tests added