Highlights

MIE now supports multiple concurrent MCP clients (Claude, Cursor, etc.) without "database locked" errors. A single daemon process holds the exclusive RocksDB lock and multiplexes access over a Unix domain socket.

How it works

mie --mcp  →  auto-starts daemon if not running  →  connects via socket

No configuration needed — it just works. You can also manage the daemon manually:

mie daemon start              # foreground
mie daemon start --background # detached
mie daemon stop
mie daemon status

What's New

Added

• Daemon server serving CozoDB over Unix domain socket
• SocketBackend client forwarding queries to daemon via newline-delimited JSON
• mie daemon start|stop|status CLI commands
• Auto-start daemon from mie --mcp with retry/backoff
• MetaBackend interface for transparent local/remote storage
• NewClientWithBackend constructor for daemon-connected clients
• Ping-based liveness detection for stale socket cleanup
• PID file locking with flock() to prevent concurrent starts
• Socket permissions restricted to owner-only (0600)
• Embedding dimension validation (daemon ↔ client)
• 15 new tests (13 socket/daemon + 2 dimension validation)

Fixed

• Deadlock in SocketBackend.Close() — mutex released before closing connection
• Silent fallback to embedded mode — now fails with clear error instead
• PID file race condition — atomic flock() instead of write-then-check
• Stale socket from crashed daemon — ping + cleanup on connect
• Scanner errors in daemon now logged
• Context propagation in daemon dispatch
• Protocol response ID validation
• Clean disconnect via MethodClose on close

Changed

• mie --mcp requires daemon (auto-started or manual) instead of opening CozoDB directly
• Graceful shutdown waits up to 5s for active handlers

Full Changelog: v1.2.0...v1.3.0

Full Changelog: v1.2.0...v1.3.0