My MCP server had 32 green tests. Not one of them had ever called a tool.

Short summary

Safari MCP had 32 passing tests but none actually tested tool behavior—they validated structure (schemas, tool counts) but not function (whether tools worked). The security boundary (tab ownership) had zero behavioral coverage, creating false confidence. Core lesson: green CI doesn't prove code works, only that it's shaped correctly; audit tests by asking what's the worst thing that could break.

• •32 passing tests never verified tools actually worked—only checked structure and schemas
• •Critical security boundary (tab ownership) had zero behavioral test coverage despite highest risk
• •Green CI gives false confidence; audit your test suite against worst-case failures

Generated with AI, which can make mistakes.