Eval bug: integer params issue when using zod4

[serjrd]

Name and Version

root@llm:/app# /app/llama-server --version
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
Device 0: NVIDIA GeForce RTX 5090, compute capability 12.0, VMM: yes
load_backend: loaded CUDA backend from /app/libggml-cuda.so
load_backend: loaded CPU backend from /app/libggml-cpu-icelake.so
version: 6730 (e60f01d)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu

Operating systems

Linux

GGML backends

CUDA

Hardware

Nvidia 5090, Nvidia 4060

Models

openai/gpt-oss-20b, mistralai/magistral-small-2509

Problem description & steps to reproduce

Created an issue at modelcontextprotocol/typescript-sdk originally, but @ochafik figured it's more of a llama.cpp thing.

Here's the original comment:

I was using zod4 with "@modelcontextprotocol/sdk": "npm:@socotra/modelcontextprotocol-sdk" and everything was fine during the development process when I served my model with LMStudio.

But as soon as I tried serving the model with llama.cpp server I got integer parameters truncation to a single digit. Tried different models - same result.

server.registerTool('test',
  {
    title: 'Get test record by ID',
    description: 'Get test record by ID',
    inputSchema: { testId: z.number().int().positive().describe('The ID of the test record') }
  },
  ({ testId }) => {
    return {
      content: [{ type: 'text', text: 'You requested test record ID: ' + testId }]
    }
  }
)

Leads to:

{
  type: 'reasoning',
  text: "The user wants to call the function test with testId: 654321. We'll need to use the function."
}
{
  type: 'tool-call',
  toolCallId: 'ADR3LRcFhzirsDqSvlfKPxlYmYLh3fKB',
  toolName: 'test',
  input: { testId: 6 },
  providerExecuted: undefined,
  providerMetadata: undefined,
  dynamic: true
}

Again: same code, same model, but served with LMStudio - no issues. The tool gets called with proper argument 654321.

If I declare 'testId' to be a string - both LMStudio and llama.cpp work fine.

When I switch to "zod": "3" and "@modelcontextprotocol/sdk": "^1.20.0", llama.cpp server works fine.

Here's the the simplest code to reproduce the issue:

package.json
test_llm.js
test_mcp.js

Hope, this helps!

First Bad Commit

No response

Relevant log output

{
  type: 'reasoning',
  text: "The user wants to call the function test with testId: 654321. We'll need to use the function."
}
{
  type: 'tool-call',
  toolCallId: 'ADR3LRcFhzirsDqSvlfKPxlYmYLh3fKB',
  toolName: 'test',
  input: { testId: 6 },
  providerExecuted: undefined,
  providerMetadata: undefined,
  dynamic: true
}