Problem

When gbrain embed is running (which can take minutes with 700+ pages) and another process tries to connect to the same PGLite database (e.g. gbrain query), PGLite throws Aborted() because embedded Postgres (WASM) only supports one connection at a time. This is a hard crash — no recovery, no helpful error.

Root Cause

PGLite uses a single WASM Postgres instance bound to a filesystem directory. Concurrent access from multiple processes to the same data directory causes an assertion failure in the Postgres WAL layer, producing Aborted() with no useful context.

Fix

Added a file-based advisory lock (src/core/pglite-lock.ts) that:

1. Acquires lock before PGLite connect — uses atomic mkdir which is POSIX-atomic, so no race conditions between processes
2. Stores PID + timestamp + command — in .gbrain-lock/lock inside the data directory, so you can see exactly what process holds the lock
3. Auto-cleans stale locks — if the PID is dead or the lock is >5 min old (embed jobs can be long), it's removed and acquired
4. Skips locking for in-memory PGLite — undefined dataDir means process-scoped, so no concurrent access is possible
5. Clear error on timeout — after 30s default, shows which process holds the lock, its PID, and how to recover (rm -rf .gbrain-lock)
6. Releases on disconnect — PGLiteEngine.disconnect() calls releaseLock() in a finally-like pattern

Testing

• 6 new tests in test/pglite-lock.test.ts: acquire/release, concurrent prevention, stale detection, in-memory skip, PID tracking, cleanup on disconnect
• All 348 existing tests pass — no regressions
• Manually verified — the Aborted() crash no longer occurs; concurrent processes get a clear error message instead

Files Changed