Summary

Adds MCP tool annotations (readOnlyHint, destructiveHint, title) to all tools to help LLMs better understand tool behavior and make safer decisions about tool execution.

Changes

Enhanced Annotation Generation (src/serena/mcp.py)

• Added destructiveHint: true to tools that can modify data (those inheriting ToolMarkerCanEdit)
• Added title annotations with human-readable display names (e.g., find_symbol → "Find Symbol")
• Set readOnlyHint: true for read-only tools, destructiveHint: true for edit tools

Fixed Memory Tools (src/serena/tools/memory_tools.py)

• WriteMemoryTool, DeleteMemoryTool, and EditMemoryTool now correctly inherit ToolMarkerCanEdit
• This ensures these tools properly receive destructiveHint: true instead of incorrectly being marked as read-only

Why This Matters

• Safety: LLMs can distinguish read-only from destructive operations before execution
• Better Tool Selection: Helps LLMs prioritize read-only tools for information gathering
• Human Readability: Title annotations provide clear display names for tool interfaces

Before/After

Before:

annotations = ToolAnnotations(readOnlyHint=not tool.can_edit())
# WriteMemoryTool incorrectly marked as readOnlyHint=True

After:

annotations = ToolAnnotations(
    title=tool_title,
    readOnlyHint=not can_edit,
    destructiveHint=can_edit,
)
# WriteMemoryTool correctly marked as destructiveHint=True

Testing

• Server builds successfully (uv sync)
• Type checker passes (pyright)
• Linter passes (ruff check)
• Memory tool inheritance verified:
  • WriteMemoryTool.can_edit() → True
  • DeleteMemoryTool.can_edit() → True
  • EditMemoryTool.can_edit() → True
  • ReadMemoryTool.can_edit() → False
  • ListMemoriesTool.can_edit() → False

🤖 Generated with Claude Code