Fix McpExecutor kag_project_config initialization for MCP HTTP store_…

[Codingwqy]

Background

When configuring an external knowledge base with MCP support (e.g. Baidu Map MCP example), the solver uses mcp_executor to call an MCP tool via a Python script specified by store_path.

In this scenario, store_path can be an HTTP URL (e.g. a hosted map.py script). During runtime, the following error is raised when the MCP tool is invoked:

'McpExecutor' object has no attribute 'kag_project_config'

Root cause

In kag/solver/executor/mcp/mcp_executor.py, the initialization order in McpExecutor.__init__ is:

self.mcp_file_path = self.download_data(store_path)
...
task_id = kwargs.get(KAGConstants.KAG_QA_TASK_CONFIG_KEY, None)
kag_config = KAGConfigAccessor.get_config(task_id)
self.kag_project_config = kag_config.global_config

However, download_data() uses self.kag_project_config when store_path is an HTTP URL:

local_file_path = os.path.join(
    self.kag_project_config.ckpt_dir, "mcp_service"
)

So when store_path starts with http:// or https://, download_data() is called before self.kag_project_config is set, which leads to:

AttributeError: 'McpExecutor' object has no attribute 'kag_project_config'

Fix

Change the initialization order in McpExecutor.__init__:

1. First load the KAG project config via KAGConfigAccessor.get_config(...).
2. Set self.kag_project_config.
3. Then call self.download_data(store_path) so that download_data() can safely use self.kag_project_config.ckpt_dir.

In pseudocode:

task_id = kwargs.get(KAGConstants.KAG_QA_TASK_CONFIG_KEY, None)
kag_config = KAGConfigAccessor.get_config(task_id)
self.kag_project_config = kag_config.global_config

self.mcp_file_path = self.download_data(store_path)

This ensures that both local paths and HTTP URLs for store_path work correctly.

Impact

• Fixes MCP scenarios where mcp_executor.store_path is an HTTP URL.
• No behavior change for:
  • Local store_path values.
  • Other executors and non-MCP pipelines.

How I tested

1. Use the latest KAG docker image (2025-11-06).
2. Deploy on Ubuntu 22.04.
3. Build an external knowledge base using the examples/baidu example.
4. Apply for MCP credentials and configure mcp_executor with an HTTP store_path pointing to map.py.
5. Create an application and connect it to the external KB.
6. Ask a query such as “今天深圳的天气”.
7. Before this change: MCP call fails with AttributeError: 'McpExecutor' object has no attribute 'kag_project_config'.
8. After this change: MCP executor runs successfully and returns a normal answer.

[whfcarter]

I have reviewed the changes and I am fine with it.

[caszkgui]

LGTM