llm

type AnthropicClient struct {
	// contains filtered or unexported fields
}

type AnthropicConfig struct {
	Enabled        bool
	APIKey         string
	Model          string
	BaseURL        string
	RequestTimeout time.Duration
}

type CacheTTL string

type Client interface {
	// CompleteStructured sends system+user prompts and a JSON schema the
	// response must conform to. Returns raw schema-constrained JSON.
	CompleteStructured(ctx context.Context, req StructuredRequest) (*StructuredResponse, error)
}

type FakeClient struct {
	Raw          json.RawMessage
	Model        string
	InputTokens  int
	OutputTokens int
	Err          error
	Requests     []StructuredRequest
	Responses    []*StructuredResponse
	Errors       []error
	// contains filtered or unexported fields
}

type OutputError struct {
	Reason       string
	StopReason   string
	OutputTokens int
	// Text is the raw provider text content (possibly truncated JSON or prose).
	Text string
}

type RateLimitError struct {
	RetryAfter time.Duration
	Err        error
}

type StructuredRequest struct {
	System    string
	Prompt    string
	Schema    map[string]any // JSON Schema for the output object
	MaxTokens int

	// Prompt caching (additive; zero values preserve the uncached behaviour).
	// SystemCacheTTL sets a cache breakpoint at the end of the system block.
	SystemCacheTTL CacheTTL
	// CachedUserPrefix, when non-empty, is emitted as a user text block before
	// Prompt with its own cache breakpoint. Use it for the large, stable portion
	// of the user content so the volatile tail in Prompt stays uncached.
	CachedUserPrefix    string
	CachedUserPrefixTTL CacheTTL
}

type StructuredResponse struct {
	Raw json.RawMessage
	// EmptyOutput is true when the provider returned no structured JSON output
	// (e.g. a plain-text "nothing applies" message) rather than a schema-shaped
	// payload. Raw is nil in that case. Callers should treat this as a valid
	// "no result" outcome, not a failure. It is distinct from a truncated
	// response, which surfaces as ErrInvalidOutput so it can be retried.
	EmptyOutput  bool
	Model        string
	InputTokens  int
	OutputTokens int
	// Prompt-cache accounting; cache reads do not count against ITPM rate limits
	// for most models, so these are useful for verifying caching is effective.
	CacheReadInputTokens     int
	CacheCreationInputTokens int
}