Summary

The MCP stdio server in src/mcp/server.ts doesn't exit when its client disconnects. The bun process keeps running, holding open the PGLite data directory.

Symptom that surfaced this

Over days of normal use, 20+ orphaned gbrain serve processes accumulated. Because PGLite is single-writer, they contended for the write lock. Email sync hit its 15s per-put timeout on every page: 114 puts × 15s = 28.5min runs that wrote 0 emails. This was the root cause behind a "no gbrain run in 50h" alert in my setup.

Fix

After server.connect(transport), listen for the standard MCP stdio shutdown signals:

• process.stdin end and close (client disconnect → EOF on stdin)
• transport.onclose (MCP SDK's own signal)
• SIGTERM / SIGINT / SIGHUP (orchestrator shutdown)

Each path calls engine.disconnect?.() and process.exit(0). A shuttingDown flag prevents double-execution if multiple signals fire.

Test plan

• Reproduced original behavior: gbrain serve left running after Claude Code disconnected; orphans stacked up across days
• With fix: stdin EOF → [gbrain-serve] shutdown: stdin end\n on stderr → process exits cleanly
• Verified engine.disconnect() is awaited before exit so PGLite lock releases
• No change to dispatch path or tool call behavior

Files

• src/mcp/server.ts — +20 lines, no removals

🤖 Generated with Claude Code

^{Need help on this PR? Tag @codesmith with what you need.}

• Let Codesmith autofix CI failures and bot reviews