MCP server hangs indefinitely on Windows + OneDrive paths during indexing

[aditya-nawal]

Environment:

• OS: Windows 11
• Shell: bash (Git Bash / Claude Code CLI)
• Project path: OneDrive-synced folder (C:\Users...\xyz)
• token-savior version: (installed via venv at ~/.local/token-savior-venv/)
• Client: claude-code

---

What happens:

set_project_root and startup indexing via WORKSPACE_ROOTS hang indefinitely — no response, no error, no timeout. The process stays alive (shows in tasklist) but uses barely any RAM (~11MB) and never
completes.

---

What we tried:

1. Full project root in WORKSPACE_ROOTS → hung for 10+ minutes, process showed 11MB RAM (not a memory issue)
2. Added EXTRA_EXCLUDE_PATTERNS (wrong env var — we later discovered the correct name is EXCLUDE_EXTRA) → exclusions were silently ignored, still hung
3. Scoped WORKSPACE_ROOTS to only experiment2/ (895 files) → still hung 15+ min
4. Added EXCLUDE_EXTRA with /outputs/ to remove JSONL data files → still hung
5. Added INCLUDE_PATTERNS=**/*.py to limit to 53 Python files, single glob traversal → still hung

In all cases: process starts, uses ~11MB RAM, and never returns a response within 15+ minutes.

---

Root cause hypothesis:

The project lives on a Windows OneDrive-synced folder. Even when OneDrive sync is paused, Windows keeps an NTFS filesystem filter driver active on OneDrive paths. This intercepts every file open/stat call to
check sync status, making directory traversal and file reads extremely slow.

Specifically:

• pathlib.Path.glob() traverses the full directory tree once per include pattern (10 traversals by default for **/.py, **/.ts, **/*.json, etc.)
• Each file stat/open on OneDrive goes through the filter driver
• Result: what takes <1 second on a local path takes 15+ minutes on an OneDrive path

Confirming evidence: get_project_summary (which doesn't trigger indexing) returns instantly. Only operations that trigger _build_slot → slot.indexer.index() hang.

---