Bug fixes:
- Remove 100-paragraph cap that prevented extracting content past ~page 4
Now calculates limit based on number of pages requested (300 paras/page)
- Add fallback page estimation when docs lack explicit page breaks
Uses ~25 paragraphs per page for navigation in non-paginated docs
- Fix _get_available_headings to scan full document (was only first 100 elements)
Headings like Chapter 10 at element 1524 were invisible
- Fix MCP connection by disabling FastMCP banner (show_banner=False)
ASCII art banner was corrupting stdout JSON-RPC protocol
Changes:
- Default image_mode changed from 'base64' to 'files' to avoid huge responses
- Add proper .mcp.json config with command/args format
- Add test document to .gitignore for privacy
- Create @resolve_field_defaults decorator to handle Pydantic FieldInfo
objects when tools are called directly (outside MCP framework)
- Create @handle_office_errors decorator for consistent error wrapping
- Apply decorators to Excel and Word mixins, removing ~100 lines of
boilerplate code
- Fix Excel formula extraction performance: load workbooks once before
loop instead of per-cell (100x faster with calculated values)
- Update test suite to use correct mock patch paths (patch where names
are looked up, not where defined)
- Add torture_test.py for real document validation
- Use app.run_stdio_async() instead of deprecated stdio_server import
- Aligns with FastMCP 2.11.3 API
- Server now starts correctly with uv run mcp-office-tools
- Maintains all MCPMixin functionality and tool registration
