agent-pulse 1.4.2 → 1.4.4

package/README.md

@@ -450,6 +450,41 @@ const result = await agent.run("A futuristic city skyline in neon colors.");
 // "![Generated Image](data:image/png;base64,...)" or "![Generated Image](https://...)"
 ```
 
+#### Image Editing (Reference Image)
+
+Edit or remix an existing image by passing `reference_image` in config. This is **provider-agnostic** — the same code works with both Grok and Google providers.
+
+```typescript
+import { Agent, grok } from 'agent-pulse';
+
+const agent = new Agent({
+  name: 'editor',
+  provider: new grok('grok-imagine-image'),
+  config: {
+    aspect_ratio: '3:4',
+    response_format: 'b64_json',
+    reference_image: 'data:image/png;base64,iVBORw0KGgo...' // base64 data URI
+  }
+});
+
+const result = await agent.run("Change the background to a sunset beach.");
+// result.content contains the edited image as a markdown data URI
+```
+
+To swap providers, just change the provider line — `reference_image` works the same way:
+
+```typescript
+// Grok
+provider: new grok('grok-imagine-image', process.env.GROK_API_KEY)
+
+// Google Gemini
+provider: new google('gemini-2.5-flash-image', process.env.GOOGLE_API_KEY)
+```
+
+> **How it works under the hood:**
+> - **Grok**: Sends a JSON request to `/v1/images/edits` with `image_url` (since the OpenAI SDK's `images.edit()` uses multipart/form-data, which x.ai doesn't support).
+> - **Google**: Injects the image as `inlineData` alongside the text prompt in the multimodal content.
+
 #### Google (Gemini)
 Gemini models can generate images as part of their response.
 

package/dist/providers/google.js

@@ -188,6 +188,21 @@ class GoogleProvider {
                     }];
             }
         }
+        // Inject reference_image from config as inlineData (provider-agnostic image editing)
+        if (config?.reference_image) {
+            const dataUriMatch = String(config.reference_image).match(/^data:(image\/\w+);base64,(.+)$/);
+            if (dataUriMatch) {
+                const lastUserMsg = [...contents].reverse().find((c) => c.role === 'user');
+                if (lastUserMsg) {
+                    lastUserMsg.parts.push({
+                        inlineData: {
+                            mimeType: dataUriMatch[1],
+                            data: dataUriMatch[2]
+                        }
+                    });
+                }
+            }
+        }
         // 4. Call API (Streaming)
         const result = await this.client.models.generateContentStream({
             model: this.model,

package/dist/providers/grok.js

@@ -23,44 +23,18 @@ class GrokProvider {
             const promptText = Array.isArray(prompt)
                 ? prompt.filter(m => m.role === 'user').map(m => m.content).join('\n')
                 : String(prompt);
-            let images;
-            if (config?.image_url) {
-                // Image editing — direct JSON request to /v1/images/edits
-                // (OpenAI SDK's images.edit() uses multipart/form-data, but x.ai requires JSON)
-                const body = {
-                    model: this.model,
-                    prompt: promptText,
-                    image_url: config.image_url,
-                    n: config?.n || 1,
-                    response_format: config?.response_format || 'b64_json',
-                    ...(config?.aspect_ratio && { aspect_ratio: config.aspect_ratio }),
-                };
-                const res = await fetch('https://api.x.ai/v1/images/edits', {
-                    method: 'POST',
-                    headers: {
-                        'Authorization': `Bearer ${this.client.apiKey}`,
-                        'Content-Type': 'application/json',
-                    },
-                    body: JSON.stringify(body),
-                });
-                if (!res.ok) {
-                    const errText = await res.text();
-                    throw new Error(`Request failed with status ${res.status}: ${errText}`);
-                }
-                const json = await res.json();
-                images = json.data;
-            }
-            else {
-                // Standard image generation
-                const response = await this.client.images.generate({
-                    model: this.model,
-                    prompt: promptText,
-                    n: config?.n || 1,
-                    response_format: config?.response_format || 'b64_json',
-                    ...(config?.aspect_ratio && { aspect_ratio: config.aspect_ratio }),
-                });
-                images = response.data;
-            }
+            // Both generation and editing use /v1/images/generations
+            // For editing, just add image_url to the same request
+            const imageUrl = config?.reference_image || config?.image_url;
+            const response = await this.client.images.generate({
+                model: this.model,
+                prompt: promptText,
+                n: config?.n || 1,
+                response_format: config?.response_format || 'b64_json',
+                ...(config?.aspect_ratio && { aspect_ratio: config.aspect_ratio }),
+                ...(imageUrl && { image_url: imageUrl }),
+            });
+            const images = response.data;
             const markdownParts = images.map((img, i) => {
                 if (img.b64_json) {
                     return `![Generated Image${images.length > 1 ? ` ${i + 1}` : ''}](data:image/png;base64,${img.b64_json})`;

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "agent-pulse",
-    "version": "1.4.2",
+    "version": "1.4.4",
     "description": "A lightweight, agentic AI framework for JavaScript/TypeScript",
     "main": "dist/index.js",
     "types": "dist/index.d.ts",
@@ -44,4 +44,4 @@
         "typescript": "^5.3.3",
         "vitest": "^1.3.1"
     }
-}
+}