rony-llm-agent/pkg/llm/providers/openai/client.go

package openai

import (
	"bufio"
	"bytes"
	"context"
	"encoding/json"
	"fmt"
	"io"
	"iter"
	"net/http"
	"strings"

	"github.com/VictorVargas/rony-llm-agent/pkg/llm"
)

// Config holds the settings needed to create an OpenAI client.
type Config struct {
	APIKey  string
	Model   string
	BaseURL string // defaults to https://api.openai.com/v1
}

// Client implements llm.LLMClient for OpenAI.
type Client struct {
	apiKey  string
	model   string
	baseURL string
	http    *http.Client
}

// New returns a new OpenAI client.
func New(cfg Config) (*Client, error) {
	if cfg.APIKey == "" {
		return nil, fmt.Errorf("openai: API key is required")
	}

	baseURL := cfg.BaseURL
	if baseURL == "" {
		baseURL = "https://api.openai.com/v1"
	}

	return &Client{
		apiKey:  cfg.APIKey,
		model:   cfg.Model,
		baseURL: baseURL,
		http:    http.DefaultClient,
	}, nil
}

func (c *Client) Generate(ctx context.Context, req llm.CompletionRequest) (llm.CompletionResponse, error) {
	endpoint := c.baseURL + "/chat/completions"

	payload, err := c.buildRequest(req, false)
	if err != nil {
		return llm.CompletionResponse{}, fmt.Errorf("building request: %w", err)
	}

	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, payload)
	if err != nil {
		return llm.CompletionResponse{}, fmt.Errorf("creating request: %w", err)
	}
	httpReq.Header.Set("Authorization", "Bearer "+c.apiKey)
	httpReq.Header.Set("Content-Type", "application/json")

	resp, err := c.http.Do(httpReq)
	if err != nil {
		return llm.CompletionResponse{}, fmt.Errorf("request failed: %w", err)
	}
	defer resp.Body.Close()

	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
		body, _ := io.ReadAll(resp.Body)
		return llm.CompletionResponse{}, fmt.Errorf("API error %d: %s", resp.StatusCode, string(body))
	}

	var apiResp openaiChatResponse
	if err := json.NewDecoder(resp.Body).Decode(&apiResp); err != nil {
		return llm.CompletionResponse{}, fmt.Errorf("decoding response: %w", err)
	}

	return c.toResponse(apiResp)
}

func (c *Client) Stream(ctx context.Context, req llm.CompletionRequest) iter.Seq2[llm.StreamChunk, error] {
	return func(yield func(llm.StreamChunk, error) bool) {
		endpoint := c.baseURL + "/chat/completions"

		payload, err := c.buildRequest(req, true)
		if err != nil {
			yield(llm.StreamChunk{}, fmt.Errorf("building request: %w", err))
			return
		}

		httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, payload)
		if err != nil {
			yield(llm.StreamChunk{}, fmt.Errorf("creating request: %w", err))
			return
		}
		httpReq.Header.Set("Authorization", "Bearer "+c.apiKey)
		httpReq.Header.Set("Content-Type", "application/json")
		httpReq.Header.Set("Accept", "text/event-stream")

		resp, err := c.http.Do(httpReq)
		if err != nil {
			yield(llm.StreamChunk{}, fmt.Errorf("request failed: %w", err))
			return
		}
		defer resp.Body.Close()

		if resp.StatusCode < 200 || resp.StatusCode >= 300 {
			body, _ := io.ReadAll(resp.Body)
			yield(llm.StreamChunk{}, fmt.Errorf("API error %d: %s", resp.StatusCode, string(body)))
			return
		}

		// toolCallAccum buffers one tool call's fragments as they stream in:
		// the SSE format sends the id/name in the first delta for a given
		// tool-call index and the (potentially large) arguments JSON in
		// pieces across many subsequent deltas, so it can't be handed to a
		// tool handler until it's fully assembled. Same accumulation the
		// llamacpp client does — without it, tool calls made over a stream
		// were silently dropped and the agent loop never executed them.
		type toolCallAccum struct {
			id   string
			name string
			args strings.Builder
		}
		toolCallFrags := map[int]*toolCallAccum{}
		var toolCallOrder []int

		flushToolCalls := func() []llm.ToolCall {
			if len(toolCallOrder) == 0 {
				return nil
			}
			calls := make([]llm.ToolCall, 0, len(toolCallOrder))
			for _, idx := range toolCallOrder {
				frag := toolCallFrags[idx]
				calls = append(calls, llm.ToolCall{
					ID:        frag.id,
					Name:      frag.name,
					Arguments: json.RawMessage(frag.args.String()),
				})
			}
			toolCallFrags = map[int]*toolCallAccum{}
			toolCallOrder = nil
			return calls
		}

		scanner := bufio.NewScanner(resp.Body)
		// A single SSE line can exceed bufio.Scanner's 64KB default cap
		// (e.g. a large tool-call arguments delta), which would kill the
		// stream with "token too long" — same headroom the anthropic
		// client already reserves.
		scanner.Buffer(make([]byte, 0, 64*1024), 4*1024*1024)
		for scanner.Scan() {
			line := scanner.Text()
			if !strings.HasPrefix(line, "data: ") {
				continue
			}
			data := strings.TrimPrefix(line, "data: ")
			if data == "[DONE]" {
				return
			}

			var event openaiStreamEvent
			if err := json.Unmarshal([]byte(data), &event); err != nil {
				yield(llm.StreamChunk{}, fmt.Errorf("decoding event: %w", err))
				return
			}

			var usage llm.TokenUsage
			if event.Usage != nil {
				usage = llm.TokenUsage{
					InputTokens:  event.Usage.PromptTokens,
					OutputTokens: event.Usage.CompletionTokens,
					TotalTokens:  event.Usage.TotalTokens,
				}
			}

			if len(event.Choices) == 0 {
				// The usage-only event (per stream_options.include_usage)
				// carries no choices, so it needs its own chunk.
				if event.Usage != nil {
					if !yield(llm.StreamChunk{Usage: usage}, nil) {
						return
					}
				}
				continue
			}

			for _, choice := range event.Choices {
				hasFragment := len(choice.Delta.ToolCalls) > 0
				for _, tc := range choice.Delta.ToolCalls {
					frag, ok := toolCallFrags[tc.Index]
					if !ok {
						frag = &toolCallAccum{}
						toolCallFrags[tc.Index] = frag
						toolCallOrder = append(toolCallOrder, tc.Index)
					}
					if tc.ID != "" {
						frag.id = tc.ID
					}
					if tc.Function.Name != "" {
						frag.name = tc.Function.Name
					}
					frag.args.WriteString(tc.Function.Arguments)
				}

				chunk := llm.StreamChunk{
					Delta: choice.Delta.Content,
					Usage: usage,
				}
				if choice.FinishReason != "" {
					chunk.FinishReason = choice.FinishReason
					chunk.ToolCalls = flushToolCalls()
				}

				// A fragment-only event (a piece of a tool call's streamed
				// arguments, with nothing else in this delta) has nothing
				// yet for the agent loop to act on: it was buffered above,
				// so skip yielding an empty chunk for it.
				if hasFragment && chunk.Delta == "" && chunk.FinishReason == "" {
					continue
				}

				if !yield(chunk, nil) {
					return
				}
			}
		}
		if err := scanner.Err(); err != nil {
			yield(llm.StreamChunk{}, fmt.Errorf("stream error: %w", err))
		}
	}
}

func (c *Client) Name() string {
	return "openai"
}

func (c *Client) Capabilities() llm.ProviderCapabilities {
	return llm.ProviderCapabilities{
		SupportsTools:    true,
		SupportsVision:   true,
		SupportsVideo:    false,
		SupportsJSON:     true,
		MaxContextWindow: 128000,
	}
}

// buildRequest converts an llm.CompletionRequest to the OpenAI API format.
func (c *Client) buildRequest(req llm.CompletionRequest, stream bool) (io.Reader, error) {
	// Convert messages to OpenAI format
	messages := make([]openaiMessage, len(req.Messages))
	for i, m := range req.Messages {
		content, err := buildContentValue(m)
		if err != nil {
			return nil, err
		}
		messages[i] = openaiMessage{
			Role:       string(m.Role),
			Content:    content,
			ToolCallID: m.ToolCallID,
			Name:       m.Name,
		}
		if len(m.ToolCalls) > 0 {
			calls := make([]openaiToolCall, len(m.ToolCalls))
			for j, tc := range m.ToolCalls {
				calls[j] = openaiToolCall{
					ID:   tc.ID,
					Type: "function",
					Function: openaiFunction{
						Name:      tc.Name,
						Arguments: string(tc.Arguments),
					},
				}
			}
			messages[i].ToolCalls = calls
		}
	}

	tools := make([]openaiTool, len(req.Tools))
	for i, t := range req.Tools {
		var tool openaiTool
		if err := json.Unmarshal(t, &tool); err != nil {
			return nil, fmt.Errorf("parsing tool %d: %w", i, err)
		}
		tools[i] = tool
	}

	// The per-request model wins when set; otherwise fall back to the
	// client's configured one (Config.Model used to be discarded entirely,
	// so every request went out with an empty model — a hard API error on
	// OpenAI, and the agent loop never sets req.Model).
	model := req.Model
	if model == "" {
		model = c.model
	}

	openaiReq := openaiChatRequest{
		Model:    model,
		Messages: messages,
		Stream:   stream,
	}
	if stream {
		// Ask for a final SSE event carrying token usage (OpenAI-style
		// streaming omits it otherwise), so callers can track real token
		// counts per turn instead of always seeing zero.
		openaiReq.StreamOptions = &openaiStreamOptions{IncludeUsage: true}
	}
	if len(tools) > 0 {
		openaiReq.Tools = tools
	}
	if req.ToolChoice != nil {
		openaiReq.ToolChoice = req.ToolChoice
	}
	if req.Temperature != nil {
		tmp := *req.Temperature
		openaiReq.Temperature = &tmp
	}
	if req.MaxTokens != nil {
		max := *req.MaxTokens
		openaiReq.MaxTokens = &max
	}
	if len(req.Stop) > 0 {
		openaiReq.Stop = req.Stop
	}

	data, err := json.Marshal(openaiReq)
	if err != nil {
		return nil, fmt.Errorf("marshaling request: %w", err)
	}
	return bytes.NewReader(data), nil
}

// toResponse converts an OpenAI API response to our CompletionResponse.
func (c *Client) toResponse(resp openaiChatResponse) (llm.CompletionResponse, error) {
	if len(resp.Choices) == 0 {
		return llm.CompletionResponse{}, fmt.Errorf("openai: response contained no choices")
	}
	choice := resp.Choices[0]
	result := llm.CompletionResponse{
		ID:         resp.ID,
		Model:      resp.Model,
		Content:    choice.Message.Content,
		StopReason: choice.FinishReason,
	}

	for _, tc := range choice.Message.ToolCalls {
		result.ToolCalls = append(result.ToolCalls, llm.ToolCall{
			ID:        tc.ID,
			Name:      tc.Function.Name,
			Arguments: json.RawMessage(tc.Function.Arguments),
		})
	}

	result.Usage = llm.TokenUsage{
		InputTokens:  resp.Usage.PromptTokens,
		OutputTokens: resp.Usage.CompletionTokens,
		TotalTokens:  resp.Usage.TotalTokens,
	}

	return result, nil
}

// OpenAI API types

type openaiChatRequest struct {
	Model         string               `json:"model"`
	Messages      []openaiMessage      `json:"messages"`
	Tools         []openaiTool         `json:"tools,omitempty"`
	ToolChoice    interface{}          `json:"tool_choice,omitempty"`
	Temperature   *float32             `json:"temperature,omitempty"`
	MaxTokens     *int                 `json:"max_tokens,omitempty"`
	Stop          []string             `json:"stop,omitempty"`
	Stream        bool                 `json:"stream"`
	StreamOptions *openaiStreamOptions `json:"stream_options,omitempty"`
}

type openaiStreamOptions struct {
	IncludeUsage bool `json:"include_usage"`
}

type openaiMessage struct {
	Role string `json:"role"`
	// Content is either a plain string (the common case) or a
	// []openaiContentPart when the source llm.Message carried Parts - see
	// buildContentValue.
	Content    interface{}      `json:"content"`
	ToolCallID string           `json:"tool_call_id,omitempty"`
	Name       string           `json:"name,omitempty"`
	ToolCalls  []openaiToolCall `json:"tool_calls,omitempty"`
}

// openaiContentPart is one block of a multipart "content" array, following
// the same shape OpenAI's vision-capable chat completions endpoint expects.
type openaiContentPart struct {
	Type     string          `json:"type"`
	Text     string          `json:"text,omitempty"`
	ImageURL *openaiMediaURL `json:"image_url,omitempty"`
}

type openaiMediaURL struct {
	URL string `json:"url"`
}

// buildContentValue converts an llm.Message's Parts into the OpenAI
// multipart content shape, or falls back to the plain Content string when
// there are no Parts - existing callers building a plain-text Message are
// completely unaffected. A video part is rejected outright: OpenAI's chat
// completions API has no video content type, so sending one would just
// produce a confusing API error instead of this clear one.
func buildContentValue(m llm.Message) (interface{}, error) {
	if len(m.Parts) == 0 {
		return m.Content, nil
	}
	parts := make([]openaiContentPart, 0, len(m.Parts))
	for _, p := range m.Parts {
		switch p.Type {
		case "text":
			parts = append(parts, openaiContentPart{Type: "text", Text: p.Text})
		case "image":
			parts = append(parts, openaiContentPart{Type: "image_url", ImageURL: &openaiMediaURL{URL: p.MediaURL}})
		case "video":
			return nil, fmt.Errorf("openai: video attachments are not supported by the chat completions API")
		default:
			return nil, fmt.Errorf("openai: unknown content part type %q", p.Type)
		}
	}
	return parts, nil
}

type openaiTool struct {
	Type     string          `json:"type"`
	Function json.RawMessage `json:"function"`
}

type openaiChatResponse struct {
	ID      string         `json:"id"`
	Model   string         `json:"model"`
	Choices []openaiChoice `json:"choices"`
	Usage   openaiUsage    `json:"usage"`
}

type openaiChoice struct {
	Index        int                 `json:"index"`
	Message      openaiMessageResult `json:"message"`
	FinishReason string              `json:"finish_reason"`
}

type openaiMessageResult struct {
	Role      string           `json:"role"`
	Content   string           `json:"content"`
	ToolCalls []openaiToolCall `json:"tool_calls"`
}

type openaiToolCall struct {
	ID       string         `json:"id"`
	Type     string         `json:"type"`
	Function openaiFunction `json:"function"`
}

type openaiFunction struct {
	Name      string `json:"name"`
	Arguments string `json:"arguments"`
}

type openaiUsage struct {
	PromptTokens     int `json:"prompt_tokens"`
	CompletionTokens int `json:"completion_tokens"`
	TotalTokens      int `json:"total_tokens"`
}

// Stream event types

type openaiStreamEvent struct {
	ID      string               `json:"id"`
	Choices []openaiStreamChoice `json:"choices"`
	Usage   *openaiUsage         `json:"usage"`
}

type openaiStreamChoice struct {
	Index        int               `json:"index"`
	Delta        openaiStreamDelta `json:"delta"`
	FinishReason string            `json:"finish_reason"`
}

type openaiStreamDelta struct {
	Content   string                 `json:"content"`
	Role      string                 `json:"role"`
	ToolCalls []openaiStreamToolCall `json:"tool_calls"`
}

type openaiStreamToolCall struct {
	Index    int                  `json:"index"`
	ID       string               `json:"id"`
	Type     string               `json:"type"`
	Function openaiStreamFunction `json:"function"`
}

type openaiStreamFunction struct {
	Name      string `json:"name"`
	Arguments string `json:"arguments"`
}