open-agents-ai 0.184.74 → 0.184.75

package/dist/index.js

@@ -64391,6 +64391,32 @@ ${lines.join("\n")}
   if (flowEnabled) {
     dynamicContext += "\n\n" + FLOWSTATE_PROMPT;
   }
+  if (modelCaps?.vision) {
+    dynamicContext += `
+
+<vision-capabilities>
+You have vision capabilities available. You can analyze images using these tools:
+
+1. **image_read** \u2014 Read/view an image file directly. Use this for screenshots, photos,
+   diagrams, or any image the user provides. Pass the file path.
+   Example: image_read(image="/path/to/image.png")
+
+2. **vision** \u2014 Analyze images with Moondream vision model. Supports:
+   - caption: Describe what's in an image
+   - query: Ask a question about an image (visual QA)
+   - detect: Find all instances of an object (bounding boxes)
+   - point: Find the center location of an object (for click targeting)
+   Example: vision(image="/tmp/screenshot.png", action="caption")
+   Example: vision(image="photo.jpg", action="query", prompt="What color is the car?")
+
+3. **screenshot** \u2014 Capture the current screen (desktop automation).
+   Returns a screenshot path you can then analyze with vision or image_read.
+
+When the user asks you to look at, describe, or analyze an image or camera feed,
+use these tools proactively. Do not say you cannot see images \u2014 you can.
+For video streams or camera feeds, capture frames and analyze them sequentially.
+</vision-capabilities>`;
+  }
   let localFirstOverride = false;
   if (config.backendType === "nexus") {
     try {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "open-agents-ai",
-  "version": "0.184.74",
+  "version": "0.184.75",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",