[BUG] Native MCP tool calls fail when OpenAI-compatible providers return string arguments

[vamsi-01]

Bug Description

Lightspeed Stack Version

The issue has been reproduced in library mode with the following versions:

1. Lightspeed Stack 0.6.2

• Image: quay.io/lightspeed-core/lightspeed-stack@sha256:1317d65e10c34183de36144a8c9effc81a28aacae62f5ba6be448b60

2. Latest development image

• Source commit: 546d0ef70913751b46b5a3f83a637abbab0e19c1
• OGX version: 1.2.5
• Image digest: sha256:384cfbabc43dbcb7d0335cd32a1e8a37cbb978d022a3a999c758c33856947147

Providers / Models

1. remote::vllm

   • Model: ibm-granite/granite-4.0-h-small

2. remote::openai

   • Model: qwen2.5:7b served through local Ollama

MCP Server

• FastMCP: 4.0.1
• Transport: Streamable HTTP
• Endpoint: http://mcp-server:9000/mcp

---

Describe the Bug

LSCore fails while processing a native MCP tool call from an OpenAI-compatible model.

The model correctly selects an MCP tool and returns its arguments using the standard OpenAI function-calling format, where arguments is represented as a JSON string.

LSCore passes this raw string directly to ToolCallSummary.args, which expects a dictionary. This results in a Pydantic validation error.

The request fails before LSCore invokes the MCP tool, receives a tool result, or generates a final response.

This issue reproduces in LSCore library mode, where OGX runs embedded inside the LSCore image. Server mode was explored separately but was not used for the final reproduction.

---

To Reproduce

1. Start LSCore in library mode with an embedded OGX configuration.

2. Configure either of the following providers:

   provider_type: remote::vllm
   provider_id: granite-h-small

   or:

   provider_type: remote::openai
   provider_id: ollama

3. Configure a FastMCP Streamable HTTP server:

   mcp_servers:
     - name: test-mcp
       url: http://mcp-server:9000/mcp

4. Expose a simple MCP tool, for example:

   @mcp.tool()
   def get_subscriptions(offset: int = 0, limit: int = 10) -> dict:
       return {"data": [], "offset": offset, "limit": limit}

5. Submit an agent request:

   curl -N -sS -X POST http://127.0.0.1:8080/v1/streaming_query \
     -H 'Content-Type: application/json' \
     -d '{"query":"List my subscriptions."}'

6. Observe successful MCP tool discovery followed by a ToolCallSummary validation error.

---

Expected Behavior

LSCore should:

1. Discover the MCP tool.
2. Parse the model-provided JSON tool arguments.
3. Invoke get_subscriptions on the MCP server.
4. Return the MCP result to the model.
5. Generate and stream the final agent response.

---

Actual Behavior / Error Output

The request fails with:

pydantic_core._pydantic_core.ValidationError:
1 validation error for ToolCallSummary
args
  Input should be a valid dictionary
  input_value='{\n  "limit": 20\n}'
  input_type=str

Relevant code path

lightspeed_stack/utils/agents/tool_processor.py
└── summarize_native_tool_call()

The failing path passes:

args.get("tool_args")

directly to:

ToolCallSummary.args

However, tool_args is a JSON string for OpenAI-compatible function calls rather than a dictionary.

---

Additional Context

• The issue occurs after model inference and MCP tool discovery succeed.
• It reproduces with both:
  • Granite 4.0 H Small through remote::vllm
  • Qwen 2.5 7B through remote::openai / Ollama
• It affects both:
  • /v1/query
  • /v1/streaming_query
• The model/provider behavior is OpenAI-compatible: function-call arguments are returned as JSON strings.
• The MCP server is successfully discovered and the tool is available before the failure occurs.