Anthropic API DEBUG & WORKAROUND

A simple project to debug and a workaround for the GLM Anthropic API endpoint issue zai-org/GLM-4.5#57

Updated(11/3/2025): Add workaround for Minimax at ./minimax-fix.js. To run it:

npm i
npm run minimax

Workaround

The workaround proxy server is implemented at ./server-fix.js. It transforms all the problematic tool_result messages to the well-supported format.

How to use:

1. Clone this repository
2. Run npm install
3. Run the fix server and your CC:

npm run fix
ANTHROPIC_BASE_URL=http://localhost:3000 claude

Steps to reproduce and debug:

1. Run the server

npm install
npm start

2. Run the Claude Code

ANTHROPIC_BASE_URL=http://localhost:3000 claude

3. Play with the CC:

• user: write a simple text to a.txt
• cc: <output and ask for permission>
• user: <reject>
• user: write it to b.txt instead
• cc: <output and ask for permission>

4. Check the recorded requests under requests/

You should see the agent still wants to write to a.txt NOT b.txt by checking the last request.md file.

Compare the results of official API endpoint and Anthropic API endpoint

Copy the relative path of the step 4. request.md file, update the api.js

const ccRequestPath = "requests/req-xxx.md";

Run it via node api.js

My Conclusion:

The official Anthropic API implementation seems to ignore the following user message content pieces after the tool result content:

{
  "role": "user",
  "content": [
    {
      "type": "tool_result",
      "content": "The user doesn't want to proceed with this tool use. The tool use was rejected (eg. if it was a file edit, the new_string was NOT written to the file). STOP what you are doing and wait for the user to tell you how to proceed.",
      "is_error": true,
      "tool_use_id": "call_421hc6d81hl"
    },
    {
      "type": "text",
      "text": "[Request interrupted by user for tool use]"
    },
    {
      "type": "text",
      "text": "changed my mind, let's write it to b.txt",
      "cache_control": {
        "type": "ephemeral"
      }
    }
  ]
}

The tool_result message contains multiple content pieces, but after running the api.js test for multiple times, the API has two behavior：

• the agent still tries to write to a.txt
• the agent realizes the user rejected it, but the agent is not able to pick up the b.txt at all.

Based on the above, it's highly likely that the Anthropic to official API transformation is not correctly implemented content with multiple pieces (at least for the tool_result)