Image Analysis and Vision

Triggerfish supports image input across all interfaces. You can paste images from your clipboard in the CLI or browser, and the agent can analyze image files on disk. When your primary model does not support vision, a separate vision model can automatically describe images before they reach the primary model.

Image Input

CLI: Clipboard Paste (Ctrl+V)

Press Ctrl+V in the CLI chat to paste an image from your system clipboard. The image is read from the OS clipboard, base64-encoded, and sent to the agent as a multimodal content block alongside your text message.

Clipboard reading supports:

• Linux -- xclip or xsel
• macOS -- pbpaste / osascript
• Windows -- PowerShell clipboard access

Tidepool: Browser Paste

In the Tidepool web interface, paste images directly into the chat input using your browser's native paste functionality (Ctrl+V / Cmd+V). The image is read as a data URL and sent as a base64-encoded content block.

image_analyze Tool

The agent can analyze image files on disk using the image_analyze tool.

Parameter	Type	Required	Description
path	string	yes	Absolute path to the image file
prompt	string	no	Question or prompt about the image (default: "Describe this image in detail")

Supported formats: PNG, JPEG, GIF, WebP, BMP, SVG

The tool reads the file, base64-encodes it, and sends it to a vision-capable LLM provider for analysis.

Vision Model Fallback

When your primary model does not support vision (e.g., Z.AI glm-5), you can configure a separate vision model to automatically describe images before they reach the primary model.

How It Works

1. You paste an image (Ctrl+V) or send multimodal content
2. The orchestrator detects image content blocks in the message
3. The vision model describes each image (you see an "Analyzing image..." spinner)
4. Image blocks are replaced with text descriptions: [The user shared an image. A vision model described it as follows: ...]
5. The primary model receives a text-only message with the descriptions
6. A system prompt hint tells the primary model to treat the descriptions as if it can see the images

This is completely transparent -- you paste an image and get a response, regardless of whether the primary model supports vision.

Configuration

Add a vision field to your models config:

models:
  primary:
    provider: zai
    model: glm-5 # Non-vision primary model
  vision: glm-4.5v # Vision model for image description
  providers:
    zai:
      model: glm-5

The vision model reuses credentials from the primary provider's keychain entry. In this example, the primary provider is zai, so glm-4.5v uses the same API key stored in the OS keychain for the zai provider.

Key	Type	Description
models.vision	string	Optional vision model name for automatic image description

When Vision Fallback Activates

• Only when models.vision is configured
• Only when the message contains image content blocks
• String-only messages and text-only content blocks skip the fallback entirely
• If the vision provider fails, the error is handled gracefully and the agent continues

Events

The orchestrator emits two events during vision processing:

Event	Description
vision_start	Image description begins (includes imageCount)
vision_complete	All images described

These events drive the "Analyzing image..." spinner in the CLI and Tidepool interfaces.

If your primary model already supports vision (e.g., Anthropic Claude,

OpenAI GPT-4o, Google Gemini), you do not need to configure models.vision. Images will be sent directly to the primary model as multimodal content. :::