Bug description

The kanban plugin leaks file descriptors. After ~4 days of gateway uptime, the gateway process hits the system FD limit (1024) and crashes with [Errno 24] Too many open files.

Root cause

Every kanban_db.connect() call opens kanban.db but callers in plugins/kanban/dashboard/plugin_api.py never close the connection. Several handlers are missing conn.close() in a finally block:

• get_home_channels() — no close ← leak
• decompose_task_endpoint() — no close ← leak
• Same pattern likely in kanban CLI tools (kanban_decompose, kanban_swarm, etc.)

File descriptor evidence

After 4 days uptime:

$ ls -la /proc/<gateway_pid>/fd/ | grep kanban
→ 900+ open FDs to kanban.db and kanban.db-wal
$ lsof -p <gateway_pid> | grep kanban
→ all point to (deleted) kanban.db

Environment

• OS: Ubuntu Linux 6.17
• Hermes: 0.14.x
• Python sqlite3, gateway runs as long-lived daemon

Suggested fix

Audit all plugin_api.py handlers — wrap every _conn() call in try/finally with conn.close(). Consider a context manager:

@contextlib.contextmanager
def _conn_cm(board=None):
    conn = _conn(board=board)
    try:
        yield conn
    finally:
        conn.close()

Same fix needed in hermes_cli/kanban_decompose.py, kanban_swarm.py, etc.