[Bug] MCP tool argument mismatch: analyze_image rejects snake_case param "image_source" with "imageSource must be provided" when called through ZCode

Summary

When calling the analyze_image tool of the @z_ai/mcp-server MCP server through ZCode (as mcp__4_5v_mcp__analyze_image), the call fails with:

MCP error -400: imageSource must be provided

The same tool, called directly over stdio with the exact same arguments, works correctly.

Environment

• ZCode client (Windows 10 x64), session date 2026-07-12
• MCP server: @z_ai/mcp-server@0.1.4 (also reproduced with 0.1.2), launched via npx -y @z_ai/mcp-server@latest
• Provider: Z.AI Coding Plan (GLM-5.2)
• Server registered in ~/.zcode/cli/config.json as:

"4_5v_mcp": {
  "command": "npx",
  "args": ["-y", "@z_ai/mcp-server@latest"],
  "env": {
    "Z_AI_API_KEY": "***",
    "Z_AI_MODE": "ZAI"
  }
}

Steps to reproduce

1. Register @z_ai/mcp-server under name 4_5v_mcp in cli/config.json (config above).
2. Restart ZCode so the MCP server is indexed.
3. In a session, call the tool with arguments matching the server's JSON schema:
   • image_source: a valid local file path or URL
   • prompt: any non-empty string
4. Observe the error.

Expected behavior

The tool accepts image_source (snake_case) — this is the parameter name declared in the server's inputSchema (confirmed by tools/list below) — and returns a text analysis of the image.

Actual behavior

The call returns MCP error -400: imageSource must be provided.

Evidence that the MCP server itself is fine

(A) tools/list schema declares snake_case:

{
  "name": "analyze_image",
  "inputSchema": {
    "type": "object",
    "properties": {
      "image_source": { "type": "string", "description": "Local file path or remote URL to the image" },
      "prompt": { "type": "string", "description": "..." }
    },
    "required": ["image_source", "prompt"]
  }
}

All 8 tools of @z_ai/mcp-server use snake_case in their schema (image_source, video_source, expected_image_source, actual_image_source, output_type, programming_language, diagram_type, analysis_focus).

(B) Direct stdio call with snake_case works:
Sending this over stdio to npx -y @z_ai/mcp-server@latest returns a correct analysis:

{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{
  "name":"analyze_image",
  "arguments":{"image_source":"<path>","prompt":"What website is this? One sentence."}
}}

Server log:

INFO: Starting general image analysis [{"imageSource":"<path>","prompt":"..."}]
INFO: Request ZAI chat completions API for vision analysis [{"model":"glm-4.6v","messageCount":2}]
INFO: analyze-image analysis completed successfully

Response: {"content":[{"type":"text","text":"This is the Hacker News website."}]}

(C) Direct stdio call with camelCase (imageSource) fails with -32602:

{"jsonrpc":"2.0","id":4,...,"arguments":{"imageSource":"<path>","prompt":"..."}}
→ {"error":{"code":-32602,"message":"Invalid arguments: image_source is required"}}

Hypothesis (please verify on your side)

The MCP client layer in ZCode appears to convert snake_case argument names to camelCase before forwarding to the server (or the server's internal validator expects camelCase from the ZCode path but snake_case from raw stdio). The error message wording (imageSource must be provided) does not originate from @z_ai/mcp-server's zod schema (which would say image_source is required, as shown in evidence C) — it looks like an intermediate layer's validation message.

A second possibility is a name collision with a built-in vision tool, but I have no way to confirm this without ZCode debug logs.

Impact

All 8 tools of @z_ai/mcp-server are unusable through ZCode even though the server and the underlying GLM-4.6V API are fully functional. Users have to fall back to manual stdio calls, which defeats the purpose of registering it as an MCP server.

What would help me help you

• Confirmation of whether ZCode's MCP client rewrites argument names.
• If there's a debug-logging flag for MCP traffic in ZCode, I'm happy to capture and attach the actual wire payload.

Thank you.

Start with the MCP registration in cli/config.json and reproduce the analyze_image call through ZCode, then compare its forwarded arguments with the direct stdio request and the tools/list schema shown here. Inspect the MCP client-layer payload or debug output; done means the argument path is identified and all declared snake_case tool arguments work through ZCode.