Struct ContainerManagerInner 

struct ContainerManagerInner {
    docker: Docker,
    ensure_locks: Mutex<HashMap<String, Arc<Mutex<()>>>>,
    exec_identities: Mutex<HashMap<String, (u64, u64)>>,
    cli_ready: Mutex<HashSet<String>>,
    cli_locks: Mutex<HashMap<String, Arc<Mutex<()>>>>,
    dev_proxy: OnceLock<DevProxyRegistry>,
}

Fields

docker: Docker

ensure_locks: Mutex<HashMap<String, Arc<Mutex<()>>>>

Per-container serialization lock: prevents concurrent ensure_running calls for the same container from both seeing “not found” in Docker and both attempting create_container (which would fail with a 409 name-conflict on the second caller).

exec_identities: Mutex<HashMap<String, (u64, u64)>>

Per-container (uid, gid) of the user execs run as, resolved once and cached — see ContainerManager::exec_identity.

cli_ready: Mutex<HashSet<String>>

(container, install marker) pairs already confirmed installed, so a running agent pays the check once, not on every turn. Dropped when the container is created or removed.

cli_locks: Mutex<HashMap<String, Arc<Mutex<()>>>>

Serializes provision_cli per container: two quick messages must not run two npm installs into the same directory.

dev_proxy: OnceLock<DevProxyRegistry>

Dev-proxy routing table, wired in post-construction by ContainerManager::attach_dev_proxy_registry (called by ContainerRuntimeHandle, which re-attaches after every reconnect since each reconnect creates a fresh ContainerManagerInner). stop/remove clear this agent’s registrations when set — see cleanup_dev_proxy_routes. Unset (None via OnceLock) in tests and any path that never wires it up; cleanup is then simply skipped, never an error. See docs/specs/SPEC_NATIVE_CONTAINER_DEV_PROXY_2026_09_19.md.