Summary

The MCP server's factor_analysis tool cannot succeed for any input: the
@mcp.tool wrapper in agent/mcp_server.py forwards codes/factor_name
arguments, but the only registered implementation
(agent/src/tools/factor_analysis_tool.py) requires pre-built CSV artifact
paths (factor_csv, return_csv, output_dir). Every MCP call therefore
dies on KeyError: 'factor_csv' before any analysis runs.

Reproduction

Connect any MCP client to vibe-trading-mcp (tested over Streamable HTTP)
and call:

{
  "tool": "factor_analysis",
  "arguments": {
    "codes": ["SBIN.NS", "HDFCBANK.NS", "ICICIBANK.NS"],
    "factor_name": "pb",
    "start_date": "2025-07-16",
    "end_date": "2026-07-16",
    "source": "yfinance"
  }
}

Result (same for any market/symbols, e.g. US tickers, and for source: "auto" / "tushare"):

factor_analysis error: 'factor_csv'

Root cause

agent/mcp_server.py — the wrapper accepts a codes-based signature and
forwards exactly those keys:

@mcp.tool
def factor_analysis(
    codes: list[str],
    factor_name: str,
    start_date: str,
    end_date: str,
    source: str = "auto",
    top_n: int = 10,
    bottom_n: int = 10,
) -> str:
    ...
    return registry.execute(
        "factor_analysis",
        {
            "codes": codes,
            "factor_name": factor_name,
            "start_date": start_date,
            "end_date": end_date,
            "source": source,
            "top_n": top_n,
            "bottom_n": bottom_n,
        },
    )

agent/src/tools/factor_analysis_tool.py — the registered tool is
artifact-based and reads different keys:

"required": ["factor_csv", "return_csv", "output_dir"],
...
factor_csv=kwargs["factor_csv"],   # <- KeyError: 'factor_csv'

FactorAnalysisTool performs no data fetching (it only pd.read_csvs the
two matrices), and I could not find any other code path that builds a factor
matrix from codes + factor_name (the wrapper's docstring mentions
daily_basic fields like pe_ttm/pb, which suggests an intended
tushare-backed builder that isn't wired up).

Verified at commit 86f6012 and, by reading the current sources, still
present on main as of 2026-07-16.

Environment

• vibe-trading-ai installed from git @ 86f6012 (pip, Python 3.12, macOS arm64)
• vibe-trading-mcp --transport http (Streamable HTTP), stock configuration
• No broker connections, no mandate; pure research usage

Expected behavior (either would resolve it)

1. The MCP wrapper matches the registered tool's contract — i.e. it accepts
   factor_csv/return_csv/output_dir so agents can feed artifacts they
   built with get_market_data + write_file (the flow the
   factor-research skill documents); or
2. A codes→factor-matrix builder is added behind the current wrapper
   signature (e.g. tushare daily_basic for CN, with an honest error for
   markets lacking a fundamentals history source).

Happy to provide more detail or test a fix. Thanks for the project — the
MCP server design (research-only surface, no order tools) is excellent.