agent-pulse 1.4.3 → 1.4.5

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

@@ -72,12 +72,17 @@ class GoogleProvider {
                 }
                 if (m.tool_calls) {
                     for (const tc of m.tool_calls) {
-                        parts.push({
+                        const fcPart = {
                             functionCall: {
                                 name: tc.name,
                                 args: tc.arguments
                             }
-                        });
+                        };
+                        // Preserve thought_signature for Gemini 3+ models
+                        if (tc.thought_signature) {
+                            fcPart.thoughtSignature = tc.thought_signature;
+                        }
+                        parts.push(fcPart);
                     }
                 }
                 if (m.content) {
@@ -225,20 +230,18 @@ class GoogleProvider {
             // Tool Calls
             // New SDK usually aggregates or provides them in chunks.
             // We need to check `candidates[0].content.parts` for `functionCall`.
-            // Streaming tool calls might be tricky as they come in parts? 
-            // Google usually sends the full function call in one chunk or at the end? 
+            // Streaming tool calls might be tricky as they come in parts?
+            // Google usually sends the full function call in one chunk or at the end?
             // Let's inspect the chunk structure via candidates.
             const parts = chunk.candidates?.[0]?.content?.parts;
             if (parts) {
                 for (const part of parts) {
                     if (part.functionCall) {
-                        // It seems Google sends full function call or we need to accumulate?
-                        // Usually it's complete in the response object if stream is done, but in stream?
-                        // Let's collect them.
                         toolCalls.push({
                             name: part.functionCall.name,
-                            arguments: part.functionCall.args, // Already parsed JSON usually
-                            id: 'call_' + Date.now() // Google doesn't always send ID, need to check
+                            arguments: part.functionCall.args,
+                            id: 'call_' + Date.now(),
+                            thought_signature: part.thoughtSignature || undefined
                         });
                     }
                 }

package/dist/providers/grok.js

@@ -23,45 +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?.reference_image || 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 imageUrl = config.reference_image || config.image_url;
-                const body = {
-                    model: this.model,
-                    prompt: promptText,
-                    image_url: imageUrl,
-                    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": "1.4.3",
+    "version": "1.4.5",
     "description": "A lightweight, agentic AI framework for JavaScript/TypeScript",
     "main": "dist/index.js",
     "types": "dist/index.d.ts",