opencode_vision：Image Analysis for OpenCode Without a Native-Vision Main Model

English | 中文

Give text-only, cost-effective opencode models (e.g. deepseek-v4-flash) the ability to see images by delegating analysis to the mimo-v2.5 vision subagent — no extra plans needed, the built-in opencode-go plan suffices.

Highlights

• Cost-effective text models, upgraded with vision. Main models such as deepseek-v4-flash — pure-text, extremely cost-effective models from the opencode-go plan — cannot understand images on their own.
• Vision via subagent. This project uses opencode's subagent mechanism to give those text models "eyes": when the main model needs to analyze an image (screenshot, UI design, game frame, a pasted picture, etc.), it delegates the task to the vision subagent (vision model mimo-v2.5) via the task tool, which reads the image and returns a description.
• No extra plans required. Everything works with the opencode-go plan you already have; no additional subscription or model package is needed.

How It Works

1. The main model encounters an image (a file path or a pasted picture).
2. It calls the task tool with subagent_type="vision", passing the image path.
3. The vision subagent reads the image with the Read tool and returns a detailed description.
4. The main model integrates the analysis result into its answer.

File Overview

File	Purpose
opencode.json	Declares that mimo-v2.5 under opencode-go supports image input
vision.md	Definition of the vision subagent (read-only, image analysis only)
vision-image-handler.ts	Pasted-image plugin: saves pasted images into .opencode/tmp/vision/ and injects text guidance for the main model to delegate to the vision subagent

Installation

One-liner (recommended)

Ask opencode to install it in a single sentence, e.g.:

Install the vision subagent and the pasted-image plugin from this repository.

opencode will copy vision.md → ~/.config/opencode/agent/vision.md, vision-image-handler.ts → ~/.config/opencode/plugins/vision-image-handler.ts, and merge the provider.opencode-go block of opencode.json into ~/.config/opencode/opencode.json for you. Then quit and restart opencode.

Manual

1. Copy vision.md → ~/.config/opencode/agent/vision.md
2. Copy vision-image-handler.ts → ~/.config/opencode/plugins/vision-image-handler.ts
3. Merge the provider.opencode-go block from opencode.json into ~/.config/opencode/opencode.json (keep the existing $schema, provider.deepseek, mcp, and other fields)
4. If global plugins are loaded as npm-style plugins, make sure @opencode-ai/plugin is in the dependencies of ~/.config/opencode/package.json (skip if it already exists)
5. Exit and restart opencode for the config to take effect

Usage

• Image files: when the main model sees an image path, call the task tool with subagent_type="vision" and pass the full image path.
• Pasted images: after the user pastes an image, the plugin saves it to disk and injects delegation guidance, so the main model calls the vision subagent accordingly.
• It is recommended to add the following conventions to your project AGENTS.md:
  • When an image needs to be analyzed, do not read it directly with the Read tool (the raw bytes are not understandable);
  • Use the task tool to call the vision subagent with the full image file path.

The screenshots below show the flow: 1.png is an example of a pasted image, and 2.png shows the text-only main model (here DeepSeek V4 Flash) delegating to the Vision Task subagent and returning the description.

Notes

• The model ID opencode-go/mimo-v2.5 in vision.md must match the actual ID under opencode-go (run /models inside opencode to confirm; if it differs, update the model ID in vision.md).
• The image storage directory .opencode/tmp/vision/ is ignored by each project's .gitignore and never enters the repository.