Struct ContainerRuntimeHandle 

pub struct ContainerRuntimeHandle {
    slot: RwLock<RuntimeSlot>,
    dev_proxy: RwLock<Option<DevProxyRegistry>>,
}

Self-healing holder for the container runtime connection.

A plain Option<ContainerManager> fixed at process boot means a Docker daemon that starts (or a socket/named-pipe that appears) AFTER AgentMux launched is never picked up without an app restart — every consumer (the availability probe AND the actual container-launch code paths) would see a permanent None. This type retries on demand instead, so “start Docker Desktop while AgentMux is already running” is picked up within one call, everywhere container_manager is read.

See docs/retro/RETRO_DOCKER_DETECTION_DIVERGENCE_2026_07_04.md.

Fields

slot: RwLock<RuntimeSlot>

dev_proxy: RwLock<Option<DevProxyRegistry>>

Dev-proxy registry to attach to every ContainerManager this handle ever produces — including across reconnects, each of which creates a fresh manager instance with its own unset attach_dev_proxy_registry state. None until set_dev_proxy_registry is called (bootstrap wires this in once both this handle and the registry exist); tests using disabled() never set it, and get() never even constructs a manager in that case. See docs/specs/SPEC_NATIVE_CONTAINER_DEV_PROXY_2026_09_19.md.

Implementations

impl ContainerRuntimeHandle

pub fn connect_at_startup() -> Self

Boot-time constructor. Attempts one connect (same as the prior behavior) but never permanently gives up on failure — a failed attempt leaves the slot Empty so get()/is_available() retry later instead of staying stuck for the process lifetime.

pub fn set_dev_proxy_registry(&self, registry: DevProxyRegistry)

Wire a dev-proxy registry into this handle so every ContainerManager it produces — now and after any future Docker reconnect — has stop/remove clear that agent’s routes. Call once at bootstrap, after both this handle and the registry exist (see bootstrap::build_app_state).

pub async fn get(&self) -> Option<ContainerManager>

Returns a connected manager, retrying ContainerManager::connect() if the slot is currently Empty. Cheap on the happy path (a read-lock plus a cheap Arc-backed clone); only takes the write lock and re-dials when we don’t already have a manager.

pub async fn is_available(&self) -> bool

True iff a manager is available AND its daemon answers a live ping right now. Never trusts a cached Connected slot alone — the daemon can go down again after a successful connect, and check_available is a real docker.ping(), not a cached flag.