helix.mcp: Skill for Claude Code

MCP Exception Audit Skill

Purpose: Audit a set of MCP tool methods for exception handling hygiene and McpException wrapping compliance.

Scope: Any MCP server with tool methods (classes decorated with [McpServerToolType] or methods decorated with [McpServerTool]).

Effort: ~2 hours for a typical MCP tool set (20–30 tools)

Audience: Product Analysts performing code quality audits; Architects reviewing exception handling patterns; Implementation teams verifying compliance with established patterns (e.g., BinlogMcp, HelixTool).

---

Methodology

Phase 1: Enumeration (15 min)

1. Locate all tool classes:

   grep -r "\[McpServerToolType\]" src/
   

2. Count total tools:

   grep -rn "\[McpServerTool(" src/ | wc -l
   

3. Extract tool metadata (name, file, line):

   grep -rn "\[McpServerTool(" src/ | cut -d: -f1,2 | sort
   

Deliverable: List of all tool names, file paths, and line numbers.

Phase 2: Classification (1.5 hours)

For each tool method, walk the code and classify by exception posture:

Posture	Definition	Pattern	Fix Needed?
A	Pre-wrapped	Has try/catch that throws McpException	No
B	Raw throw	No try/catch; raw .NET exceptions propagate uncaught	Yes
C	Implicit	No explicit throw, but calls can fail uncaught	Yes
D	Swallowed	Catches and returns error object instead of throwing	Depends on architecture
E	Mixed	Multiple strategies on different code paths	Yes

Classification checklist per method:

□ Parameter validation (pre-call)
  - How are required parameters validated?
  - What exception type is thrown for invalid input?
  - Is it McpException or raw ArgumentException?

□ Service/I/O calls (try/catch)
  - Are service calls wrapped in try/catch?
  - What exception types are caught?
  - Are exceptions re-thrown as McpException?

□ Config/guard checks
  - Are configuration failures surfaced as McpException?
  - Or are they silently ignored?

□ Return paths
  - Does the method always throw on error?
  - Or does it return error objects instead?

□ Context-specific handling
  - Are semantic errors (e.g., "not found") detected and enhanced with hints?
  - Or is there a broad one-size-fits-all catch?