sudoprivacy/image-analysis

Image Analysis Skill

Analyze images via LLM provider's chat completions API with vision.

When to Use

Trigger this skill when the user asks to:

• Describe or analyze an image ("describe this image", "what's in this picture")
• Recognize content in an image ("识别图片", "分析图片内容")
• Extract text from an image (OCR) ("read the text in this screenshot")
• Answer questions about an image ("how many people are in this photo", "what color is the car")
• Reference a previous image ("tell me about the image above", "what's in the previous picture")
• Explain visual content ("explain this diagram", "what does this chart show")
• Identify objects ("what kind of plant is this", "identify this")

Steps

Step 1: Identify the image

Get the image from the user's message or workspace context.

• If the user provides a local file path, use it directly.
• If the user provides an image URL (http/https), ask the user whether they'd like to download the image to a local file first, then download it (e.g. with curl -sL "<url>" -o /tmp/image_to_analyze.<ext>) before proceeding.
• If the user references a previous image without a path (e.g. "the above image", "the image I just sent"), search the workspace directory for recent image files:
  1. List files in the workspace matching common image extensions (*.png, *.jpg, *.jpeg, *.gif, *.webp, *.bmp)
  2. Sort by modification time (most recent first)
  3. Use the most recent image. If there are multiple candidates and it's ambiguous, ask the user which image they mean.

Step 2: Run the image analysis script

bash scripts/analyze_image.sh "<image_path>" "<question or prompt>"

The script reads API credentials and model from sudoclaw.json via the SUDOCLAW_CONFIG_PATH env var (auto-set at gateway startup). No manual env var configuration is needed.

The script prints the analysis result (LLM response) on success, or an error message on failure.

Step 3: Show the result

Display the analysis result to the user.