feat(vision-analysis): mmx-cli primary, MCP fallback, clipboard handling (v1.4)

[divitkashyap]

Summary

vision-analysis skill v1.4 — mmx-cli as primary tool, MCP as fallback, plus clipboard handling and OpenCode workarounds.

Changes

mmx-cli as primary tool (v1.4)

• mmx vision describe is now the recommended primary tool — direct REST call to MiniMax VLM, no MCP transport issues, handles URLs and local files automatically
• Uses direct HTTP to /v1/coding_plan/vlm endpoint — sidesteps all OpenCode MCP stdio bugs entirely
• MCP (auto-skill-loader_minimax_understand_image) is now the fallback when mmx is not installed
• npm install -g mmx-cli — works in any host (Claude Code, OpenCode, terminal)

Clipboard fix

• Added scripts/clipboard_image.py — extracts clipboard images to temp files for analysis
• Supports macOS (osascript), Linux (xclip/wl-paste), Windows (PowerShell)
• Handles clipboard screenshot paths that don't exist on disk

OpenCode stdio bug workaround

• OpenCode direct minimax-coding-plan-mcp has broken stdio transport causing "login fail" errors
• Skill recommends auto-skill-loader proxy: https://github.com/divitkashyap/auto-skill-loader
• Proxy tool bypasses broken stdio transport

Metadata bug fixes

• Removed requires_mcp field — caused plugin to filter skill from get_available_skills
• Removed sources list — failed z.record(z.string(), z.string()) schema validation

Trigger tightening

• Tightened description to only fire when actual image is explicitly shared/referenced
• Avoids false triggers on generic text like project advice, code reviews

Tool Priority

1. mmx vision describe (recommended): npm install -g mmx-cli, set MINIMAX_API_KEY
2. auto-skill-loader MCP fallback: set MINIMAX_TOKEN_PLAN_KEY, enable in opencode.json

Testing

• mmx vision describe: works in Claude Code and terminal, URLs handled natively
• Clipboard extraction: works on macOS
• Image analysis via auto-skill-loader proxy: works
• OpenCode direct minimax-coding-plan-mcp: broken stdio transport (use proxy or mmx instead)
• Skill auto-triggers on image path/URL text: confirmed

Closes

Closes #35

---

Submitted by: divitkashyap

[divitkashyap]

Hey @zest01998 @liyuan97 — this PR is ready for review. Key changes in v1.4: mmx-cli is now the primary vision tool (direct REST, no MCP transport issues), with the auto-skill-loader MCP as fallback. Also includes clipboard handling fixes and metadata bug fixes. Let me know if you have any questions!

[divitkashyap]

Hey @zest0198 @liyuan97 @AkairoDev — this PR is ready for review. Major changes since last update (v1.4):

• mmx-cli is now the primary vision tool (direct REST, no MCP transport issues)
• MCP is now the fallback when mmx is not installed
• Clipboard handling, metadata fixes, and trigger tightening also included

The skill is installable via npx skills add MiniMax-AI/cli once merged. Happy to address any feedback!