unsupportedParts() "Inform the user" instruction overrides system prompt, preventing fallback to vision MCP for non-vision models

[yang-neu]

Description

When using a non-vision model (e.g. deepseek-v4-pro), pasting an image causes unsupportedParts() in packages/opencode/src/provider/transform.ts to replace the image with:

ERROR: Cannot read clipboard (this model does not support image input). Inform the user.

The trailing "Inform the user." is treated by the LLM as a high-priority instruction. Even when the system prompt (via prompt_append or AGENTS.md) explicitly instructs the model to handle images via a workaround (pngpaste → vision MCP), the model always responds with "I can't see images" instead.

What we tried to override this (all failed):

• prompt_append in mode.build
• AGENTS.md (project-level instructions)
• oh-my-openagent agent prompt
• opencode plugin with experimental.chat.messages.transform

Plugins

@zilliz/memsearch-opencode oh-my-openagent@latest

OpenCode version

1.18.12

Steps to reproduce

1. Configure a vision MCP server (e.g. kimi-vision-mcp)
2. Add to prompt_append or AGENTS.md: "When you see 'ERROR: Cannot read clipboard', run pngpaste + vision_analyze_image instead of telling the user you can't see it"
3. Use a non-vision model (e.g. deepseek-v4-pro)
4. Paste a clipboard image via Cmd+V in opencode
5. Observe: model ignores system prompt and outputs "I can't view images directly" or similar

Screenshot and/or share link

Operating System

macOS 15 (Apple Silicon)

Terminal

iTerm2

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Non-vision models blocked from passing images to vision-capable MCP tools #29216: Same root cause — unsupportedParts() in transform.ts replaces images with the ERROR: ... Inform the user. text when using a non-vision model, and the LLM cannot then delegate to a vision-capable MCP tool. That issue also proposes passing image metadata as text so MCP tools remain reachable.
• [FEATURE]: Allow plugins to customize behavior when a model doesn't support a media modality (image/audio/video) #36006: Feature request specifically about allowing plugins to customize the unsupportedParts() replacement behavior so non-vision models can be guided to call plugin-provided vision tools — essentially the same plugin-hook gap you encountered.