Description

The llm-task plugin tool has two parameters (input and schema) whose JSON Schema definitions contain only a description field without a type field. This is technically valid JSON Schema (implies "any type"), but llama.cpp's OpenAI-compatible endpoint rejects it with a 400 error during JSON schema-to-grammar conversion.

This breaks any agent using local LLM backends (llama.cpp / llama-server) when tools are enabled and llm-task is in the tool set.

Affected parameters

{
  "input": {"description": "Optional input payload for the task."},
  "schema": {"description": "Optional JSON Schema to validate the returned JSON."}
}

Both are missing "type". Compare with other parameters in the same tool that work fine:

{
  "prompt": {"description": "Task instruction for the LLM.", "type": "string"}
}

Reproduction

Send any chat completion request through an agent with llm-task enabled, using a llama.cpp-based backend:

curl -s http://localhost:8011/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"any","messages":[{"role":"user","content":"hi"}],"max_tokens":10,
    "tools":[{"type":"function","function":{"name":"llm_task","description":"test",
      "parameters":{"type":"object","properties":{
        "prompt":{"type":"string","description":"ok"},
        "input":{"description":"no type field"},
        "schema":{"description":"no type field"}
      },"required":["prompt"]}}}]}'

Response:

{"error":{"code":400,"message":"JSON schema conversion failed:\nUnrecognized schema: {\"description\":\"no type field\"}\nUnrecognized schema: {\"description\":\"no type field\"}","type":"invalid_request_error"}}

Adding "type": "object" to both parameters resolves the error.

Impact

• Any agent with llm-task enabled that routes through local/* TensorZero providers (or direct llama.cpp endpoints) gets a 400 on every request
• The error surfaces as HTTP 502 through TensorZero since it retries then fails
• Streaming requests return data: [DONE] immediately with no content (silent failure from the caller's perspective)

Workaround

Set tools.profile: "minimal" or tools.deny: ["llm-task"] on the affected agent to exclude the tool.

Suggested fix

Add "type": "object" (or "type": ["object", "string", "number", "array", "boolean", "null"]) to the input and schema parameter definitions in the llm-task plugin tool registration.

Environment

• OpenClaw: 2026.3.2
• llama.cpp: latest (llama-server with OpenAI-compatible endpoint)
• Model: Qwen3.5-35B-A3B (also confirmed with Qwen3.5-9B)
• OS: macOS (arm64)