Description

plugins/memory/honcho/client.py exposes get_honcho_client() — a lazy-init accessor for the process-wide _honcho_client singleton (line 665–776). The check-before-init pattern is:

_honcho_client: Honcho | None = None

def get_honcho_client(config=None) -> Honcho:
    global _honcho_client
    if _honcho_client is not None:
        return _honcho_client
    # ... full config resolution + lazy SDK install ...
    _honcho_client = Honcho(**kwargs)
    return _honcho_client

There is no lock around the check-then-initialize sequence. When two threads call get_honcho_client() concurrently before the singleton is set, both pass the is not None guard, both run the expensive initialization (config loading, optional lazy SDK install, Honcho(**kwargs) construction), and the second write overwrites the first — leaving an open connection leaked.

Impact

In multi-threaded agent sessions sharing one process (e.g. delegated tool calls, background workers), Honcho clients can be constructed multiple times, leaking connections and causing non-deterministic behavior if the first client established state (e.g. session contexts) that the replacement doesn't know about.

Fix

Add import threading, _honcho_client_lock = threading.Lock(), and apply double-checked locking in get_honcho_client().