Files
goclaw/internal/tools/create_image.go
T
viettranx 96845d1e44 fix(media): set result.Media on create_image and add MEDIA: fallback in subagent exec
Root cause: create_image tool only set ForLLM:"MEDIA:path" but never
populated result.Media. The main agent loop parses MEDIA: prefix via
parseMediaResult(), but the subagent exec loop only checked result.Media
— so media paths were silently lost for all subagent/spawn workflows.

This caused the entire downstream pipeline (task.Media → AnnounceQueue →
PublishInbound → ContentSuffix → session) to receive empty media, making
images invisible in WS chat despite Telegram working fine.

Fixes:
- create_image.go: set result.Media = []string{imagePath}
- subagent_exec.go: add MEDIA: prefix fallback parsing as safety net
2026-03-08 00:06:13 +07:00

494 lines
15 KiB
Go

package tools
import (
"bytes"
"context"
"encoding/base64"
"encoding/json"
"fmt"
"io"
"log/slog"
"net/http"
"os"
"path/filepath"
"strings"
"time"
"github.com/nextlevelbuilder/goclaw/internal/providers"
)
// credentialProvider is a narrow interface for providers that expose API credentials.
type credentialProvider interface {
APIKey() string
APIBase() string
}
// imageGenProviderPriority is the default order for image generation providers.
var imageGenProviderPriority = []string{"openrouter", "gemini", "openai"}
// imageGenModelDefaults maps provider names to default image generation models.
var imageGenModelDefaults = map[string]string{
"openrouter": "google/gemini-2.5-flash-image",
"openai": "dall-e-3",
"gemini": "gemini-2.5-flash-image",
}
// CreateImageTool generates images using an image generation API.
// Uses OpenRouter (Gemini image model) or OpenAI (DALL-E) via per-agent ImageGenConfig.
type CreateImageTool struct {
registry *providers.Registry
}
func NewCreateImageTool(registry *providers.Registry) *CreateImageTool {
return &CreateImageTool{registry: registry}
}
func (t *CreateImageTool) Name() string { return "create_image" }
func (t *CreateImageTool) Description() string {
return "Generate an image from a text description using an image generation model. Returns a MEDIA: path to the generated image file."
}
func (t *CreateImageTool) Parameters() map[string]interface{} {
return map[string]interface{}{
"type": "object",
"properties": map[string]interface{}{
"prompt": map[string]interface{}{
"type": "string",
"description": "Text description of the image to generate.",
},
"aspect_ratio": map[string]interface{}{
"type": "string",
"description": "Aspect ratio: '1:1' (default), '3:4', '4:3', '9:16', '16:9'.",
},
},
"required": []string{"prompt"},
}
}
func (t *CreateImageTool) Execute(ctx context.Context, args map[string]interface{}) *Result {
prompt, _ := args["prompt"].(string)
if prompt == "" {
return ErrorResult("prompt is required")
}
aspectRatio, _ := args["aspect_ratio"].(string)
if aspectRatio == "" {
aspectRatio = "1:1"
}
// Resolve provider from per-agent config or defaults
providerName, model := t.resolveConfig(ctx)
p, err := t.registry.Get(providerName)
if err != nil {
return ErrorResult(fmt.Sprintf("image generation provider %q not available", providerName))
}
cp, ok := p.(credentialProvider)
if !ok {
return ErrorResult(fmt.Sprintf("provider %q does not expose API credentials for image generation", providerName))
}
slog.Info("create_image: calling image generation API",
"provider", providerName, "model", model, "aspect_ratio", aspectRatio)
// Route to the correct image generation endpoint per provider:
// - gemini: native Gemini generateContent API (responseModalities)
// - openrouter: OpenAI-compat /chat/completions with modalities
// - others (openai, etc.): /images/generations
var imageBytes []byte
var usage *providers.Usage
switch providerName {
case "gemini":
var genErr error
imageBytes, usage, genErr = t.callGeminiNativeImageGen(ctx, cp.APIKey(), cp.APIBase(), model, prompt)
err = genErr
case "openrouter":
var genErr error
imageBytes, usage, genErr = t.callImageGenAPI(ctx, cp.APIKey(), cp.APIBase(), model, prompt, aspectRatio)
err = genErr
default:
var genErr error
imageBytes, usage, genErr = t.callStandardImageGenAPI(ctx, cp.APIKey(), cp.APIBase(), model, prompt)
err = genErr
}
if err != nil {
return ErrorResult(fmt.Sprintf("image generation failed: %v", err))
}
// Save to workspace under date-based folder (e.g. generated/2026-03-02/)
workspace := ToolWorkspaceFromCtx(ctx)
if workspace == "" {
workspace = os.TempDir()
}
dateDir := filepath.Join(workspace, "generated", time.Now().Format("2006-01-02"))
if err := os.MkdirAll(dateDir, 0755); err != nil {
return ErrorResult(fmt.Sprintf("failed to create output directory: %v", err))
}
imagePath := filepath.Join(dateDir, fmt.Sprintf("goclaw_gen_%d.png", time.Now().UnixNano()))
if err := os.WriteFile(imagePath, imageBytes, 0644); err != nil {
return ErrorResult(fmt.Sprintf("failed to save generated image: %v", err))
}
result := &Result{ForLLM: fmt.Sprintf("MEDIA:%s", imagePath)}
result.Media = []string{imagePath}
result.Deliverable = fmt.Sprintf("[Generated image: %s]\nPrompt: %s", filepath.Base(imagePath), prompt)
result.Provider = providerName
result.Model = model
if usage != nil {
result.Usage = usage
}
return result
}
// resolveConfig returns the provider name and model to use for image generation.
func (t *CreateImageTool) resolveConfig(ctx context.Context) (providerName, model string) {
// 1. Check per-agent ImageGenConfig from context (highest priority)
if cfg := ImageGenConfigFromCtx(ctx); cfg != nil {
if cfg.Provider != "" {
providerName = cfg.Provider
}
if cfg.Model != "" {
model = cfg.Model
}
}
// 2. Check global builtin_tools.settings (DB defaults)
if providerName == "" || model == "" {
if settings := BuiltinToolSettingsFromCtx(ctx); settings != nil {
if raw, ok := settings["create_image"]; ok && len(raw) > 0 {
var cfg struct {
Provider string `json:"provider"`
Model string `json:"model"`
}
if json.Unmarshal(raw, &cfg) == nil && cfg.Provider != "" {
// DB settings are a provider+model pair — only use if provider is available
if _, err := t.registry.Get(cfg.Provider); err == nil {
if providerName == "" {
providerName = cfg.Provider
}
if model == "" && cfg.Model != "" {
model = cfg.Model
}
}
}
}
}
}
// 3. If provider not set, find first available from priority list
if providerName == "" {
for _, name := range imageGenProviderPriority {
if _, err := t.registry.Get(name); err == nil {
providerName = name
break
}
}
}
if providerName == "" {
providerName = "openrouter" // fallback even if unavailable (error handled later)
}
// 4. If model not set, use default for this provider
if model == "" {
if m, ok := imageGenModelDefaults[providerName]; ok {
model = m
}
}
return providerName, model
}
// callImageGenAPI calls the OpenAI-compatible image generation endpoint.
// Works with OpenRouter (modalities: ["image","text"]) and OpenAI (/images/generations).
func (t *CreateImageTool) callImageGenAPI(ctx context.Context, apiKey, apiBase, model, prompt, aspectRatio string) ([]byte, *providers.Usage, error) {
// OpenRouter / OpenAI-compat: use chat completions with modalities
body := map[string]interface{}{
"model": model,
"messages": []map[string]interface{}{
{"role": "user", "content": prompt},
},
"modalities": []string{"image", "text"},
}
if aspectRatio != "" && aspectRatio != "1:1" {
body["image_config"] = map[string]interface{}{
"aspect_ratio": aspectRatio,
}
}
jsonBody, err := json.Marshal(body)
if err != nil {
return nil, nil, fmt.Errorf("marshal request: %w", err)
}
url := strings.TrimRight(apiBase, "/") + "/chat/completions"
req, err := http.NewRequestWithContext(ctx, "POST", url, bytes.NewReader(jsonBody))
if err != nil {
return nil, nil, fmt.Errorf("create request: %w", err)
}
req.Header.Set("Content-Type", "application/json")
req.Header.Set("Authorization", "Bearer "+apiKey)
client := &http.Client{Timeout: 120 * time.Second}
resp, err := client.Do(req)
if err != nil {
return nil, nil, fmt.Errorf("http request: %w", err)
}
defer resp.Body.Close()
respBody, err := io.ReadAll(resp.Body)
if err != nil {
return nil, nil, fmt.Errorf("read response: %w", err)
}
if resp.StatusCode != http.StatusOK {
return nil, nil, fmt.Errorf("API error %d: %s", resp.StatusCode, truncateBytes(respBody, 500))
}
return t.parseImageResponse(respBody)
}
// callStandardImageGenAPI uses the /images/generations endpoint (Gemini, OpenAI, and compatible providers).
// This is the standard OpenAI-compatible image generation endpoint that returns b64_json data.
func (t *CreateImageTool) callStandardImageGenAPI(ctx context.Context, apiKey, apiBase, model, prompt string) ([]byte, *providers.Usage, error) {
body := map[string]interface{}{
"model": model,
"prompt": prompt,
"n": 1,
"response_format": "b64_json",
}
jsonBody, err := json.Marshal(body)
if err != nil {
return nil, nil, fmt.Errorf("marshal request: %w", err)
}
url := strings.TrimRight(apiBase, "/") + "/images/generations"
req, err := http.NewRequestWithContext(ctx, "POST", url, bytes.NewReader(jsonBody))
if err != nil {
return nil, nil, fmt.Errorf("create request: %w", err)
}
req.Header.Set("Content-Type", "application/json")
req.Header.Set("Authorization", "Bearer "+apiKey)
client := &http.Client{Timeout: 120 * time.Second}
resp, err := client.Do(req)
if err != nil {
return nil, nil, fmt.Errorf("http request: %w", err)
}
defer resp.Body.Close()
respBody, err := io.ReadAll(resp.Body)
if err != nil {
return nil, nil, fmt.Errorf("read response: %w", err)
}
if resp.StatusCode != http.StatusOK {
return nil, nil, fmt.Errorf("API error %d: %s", resp.StatusCode, truncateBytes(respBody, 500))
}
// Parse OpenAI-compat images/generations response: {data: [{b64_json: "..."}]}
var imgResp struct {
Data []struct {
B64JSON string `json:"b64_json"`
} `json:"data"`
}
if err := json.Unmarshal(respBody, &imgResp); err != nil {
return nil, nil, fmt.Errorf("parse response: %w", err)
}
if len(imgResp.Data) == 0 || imgResp.Data[0].B64JSON == "" {
return nil, nil, fmt.Errorf("no image data in response")
}
imageBytes, err := base64.StdEncoding.DecodeString(imgResp.Data[0].B64JSON)
if err != nil {
return nil, nil, fmt.Errorf("decode base64: %w", err)
}
return imageBytes, nil, nil
}
// callGeminiNativeImageGen uses the native Gemini generateContent API with responseModalities.
// Gemini image models (gemini-2.5-flash-image, gemini-3.1-flash-image-preview) require this
// endpoint — they don't support the OpenAI-compat /images/generations or /chat/completions.
func (t *CreateImageTool) callGeminiNativeImageGen(ctx context.Context, apiKey, apiBase, model, prompt string) ([]byte, *providers.Usage, error) {
// Derive native Gemini base from OpenAI-compat base (strip /openai suffix)
nativeBase := strings.TrimRight(apiBase, "/")
nativeBase = strings.TrimSuffix(nativeBase, "/openai")
url := fmt.Sprintf("%s/models/%s:generateContent?key=%s", nativeBase, model, apiKey)
body := map[string]interface{}{
"contents": []map[string]interface{}{
{"parts": []map[string]interface{}{{"text": prompt}}},
},
"generationConfig": map[string]interface{}{
"responseModalities": []string{"TEXT", "IMAGE"},
},
}
jsonBody, err := json.Marshal(body)
if err != nil {
return nil, nil, fmt.Errorf("marshal request: %w", err)
}
req, err := http.NewRequestWithContext(ctx, "POST", url, bytes.NewReader(jsonBody))
if err != nil {
return nil, nil, fmt.Errorf("create request: %w", err)
}
req.Header.Set("Content-Type", "application/json")
client := &http.Client{Timeout: 120 * time.Second}
resp, err := client.Do(req)
if err != nil {
return nil, nil, fmt.Errorf("http request: %w", err)
}
defer resp.Body.Close()
respBody, err := io.ReadAll(resp.Body)
if err != nil {
return nil, nil, fmt.Errorf("read response: %w", err)
}
if resp.StatusCode != http.StatusOK {
return nil, nil, fmt.Errorf("API error %d: %s", resp.StatusCode, truncateBytes(respBody, 500))
}
// Parse native Gemini response: {candidates: [{content: {parts: [{inlineData: {mimeType, data}}]}}]}
var gemResp struct {
Candidates []struct {
Content struct {
Parts []struct {
InlineData *struct {
MimeType string `json:"mimeType"`
Data string `json:"data"`
} `json:"inlineData"`
Text string `json:"text"`
} `json:"parts"`
} `json:"content"`
} `json:"candidates"`
UsageMetadata *struct {
PromptTokenCount int `json:"promptTokenCount"`
CandidatesTokenCount int `json:"candidatesTokenCount"`
TotalTokenCount int `json:"totalTokenCount"`
} `json:"usageMetadata"`
}
if err := json.Unmarshal(respBody, &gemResp); err != nil {
return nil, nil, fmt.Errorf("parse response: %w", err)
}
// Extract first image from parts
for _, cand := range gemResp.Candidates {
for _, part := range cand.Content.Parts {
if part.InlineData != nil && strings.HasPrefix(part.InlineData.MimeType, "image/") {
imageBytes, err := base64.StdEncoding.DecodeString(part.InlineData.Data)
if err != nil {
return nil, nil, fmt.Errorf("decode base64: %w", err)
}
var usage *providers.Usage
if gemResp.UsageMetadata != nil {
usage = &providers.Usage{
PromptTokens: gemResp.UsageMetadata.PromptTokenCount,
CompletionTokens: gemResp.UsageMetadata.CandidatesTokenCount,
TotalTokens: gemResp.UsageMetadata.TotalTokenCount,
}
}
return imageBytes, usage, nil
}
}
}
return nil, nil, fmt.Errorf("no image data in Gemini response")
}
// parseImageResponse extracts base64 image data from the OpenAI-compat chat response.
// Looks for images in choices[0].message.content (multipart) or choices[0].message.images.
func (t *CreateImageTool) parseImageResponse(respBody []byte) ([]byte, *providers.Usage, error) {
var resp struct {
Choices []struct {
Message struct {
Content interface{} `json:"content"`
Images []struct {
ImageURL struct {
URL string `json:"url"`
} `json:"image_url"`
} `json:"images"`
} `json:"message"`
} `json:"choices"`
Usage *struct {
PromptTokens int `json:"prompt_tokens"`
CompletionTokens int `json:"completion_tokens"`
TotalTokens int `json:"total_tokens"`
} `json:"usage"`
}
if err := json.Unmarshal(respBody, &resp); err != nil {
return nil, nil, fmt.Errorf("parse response: %w", err)
}
if len(resp.Choices) == 0 {
return nil, nil, fmt.Errorf("no choices in response")
}
msg := resp.Choices[0].Message
// Try images array first (OpenRouter format)
for _, img := range msg.Images {
if imageBytes, err := decodeDataURL(img.ImageURL.URL); err == nil {
return imageBytes, convertUsage(resp.Usage), nil
}
}
// Try multipart content array (some providers return content as array of parts)
if parts, ok := msg.Content.([]interface{}); ok {
for _, part := range parts {
if m, ok := part.(map[string]interface{}); ok {
if m["type"] == "image_url" {
if imgURL, ok := m["image_url"].(map[string]interface{}); ok {
if url, ok := imgURL["url"].(string); ok {
if imageBytes, err := decodeDataURL(url); err == nil {
return imageBytes, convertUsage(resp.Usage), nil
}
}
}
}
}
}
}
return nil, nil, fmt.Errorf("no image data found in response")
}
// decodeDataURL decodes a data:image/...;base64,... URL into raw bytes.
func decodeDataURL(dataURL string) ([]byte, error) {
// Format: data:image/png;base64,iVBORw0KGgo...
idx := strings.Index(dataURL, ";base64,")
if idx < 0 {
return nil, fmt.Errorf("not a base64 data URL")
}
b64 := dataURL[idx+8:]
return base64.StdEncoding.DecodeString(b64)
}
func convertUsage(u *struct {
PromptTokens int `json:"prompt_tokens"`
CompletionTokens int `json:"completion_tokens"`
TotalTokens int `json:"total_tokens"`
}) *providers.Usage {
if u == nil {
return nil
}
return &providers.Usage{
PromptTokens: u.PromptTokens,
CompletionTokens: u.CompletionTokens,
TotalTokens: u.TotalTokens,
}
}
func truncateBytes(b []byte, max int) string {
if len(b) <= max {
return string(b)
}
return string(b[:max]) + "..."
}