Struct Daemon

server/internal/daemon/daemon.go:153–240

Daemon is the local agent runtime that polls for and executes tasks.

// Daemon is the local agent runtime that polls for and executes tasks.
type Daemon struct {
	cfg        Config
	client     *Client
	repoCache  repoCacheBackend
	skillCache *SkillBundleCache
	logger     *slog.Logger

	mu           sync.Mutex
	workspaces   map[string]*workspaceState
	runtimeIndex map[string]Runtime // runtimeID -> Runtime for provider lookups
	// profileLaunchSpecs maps a custom runtime profile_id -> the absolute
	// executable path plus fixed launch args resolved for that profile
	// (MUL-3284). Populated in registerRuntimesForWorkspace when a profile's
	// command resolves; read by runTask to launch the custom command for a
	// claimed task. Guarded by mu.
	profileLaunchSpecs map[string]profileLaunchSpec
	reloading          sync.Mutex         // prevents concurrent workspace syncs
	runtimeSet         *runtimeSetWatcher // multi-subscriber pub/sub for runtime-set changes

	versionsMu    sync.RWMutex      // guards agentVersions
	agentVersions map[string]string // provider -> detected CLI version (set during registration)

	wsHBMu      sync.RWMutex         // guards wsHBLastAck
	wsHBLastAck map[string]time.Time // runtime_id -> last successful WS heartbeat ack timestamp

	// reconcile fans out a "re-check server state now" signal to subscribers
	// (watchTaskCancellation, workspaceSyncLoop) so the WS connect/reconnect
	// path can shrink the 5s / 30s reconciliation gap to sub-second. See
	// reconcile.go and runTaskWakeupConnection.
	reconcile *reconcileBroadcaster

	// runtimeGoneMu guards runtimeGoneInflight, reregisterNextAttempt, and
	// reregisterLastCompletedAt. The state lets heartbeat / poller / WS-ack
	// handlers converge on a single recovery path when they each detect that a
	// runtime row was deleted server-side without three of them stampeding
	// registerRuntimesForWorkspace.
	runtimeGoneMu             sync.Mutex
	runtimeGoneInflight       map[string]struct{}  // runtime_id -> currently recovering
	reregisterNextAttempt     map[string]time.Time // workspace_id -> earliest time the next re-register attempt may run
	reregisterLastCompletedAt map[string]time.Time // workspace_id -> wall-clock at which the last SUCCESSFUL re-register call returned (failures intentionally not stamped — see recordRegisterCompletion)

	cancelFunc    context.CancelFunc // set by Run(); called by triggerRestart
	rootCtx       context.Context    // set by Run(); used by long-running recoveries that must survive per-runtime ctx cancellation
	restartBinary string             // non-empty after a successful update; path to the new binary
	updating      atomic.Bool        // prevents concurrent update attempts
	activeTasks   atomic.Int64       // number of tasks currently in handleTask; exposed via /health
	ready         atomic.Bool        // false until preflight completes; gates /health status (starting -> running)

	// claimMu guards pauseClaims and claimsInFlight. It is held only for the
	// microseconds it takes to make a decision; ClaimTask itself runs without
	// the lock so a slow per-runtime claim cannot stall auto-update or any
	// other poller.
	//
	// The pair is the auto-update path's barrier against the issue's
	// requirement that "升级过程中如果有 task 进来，会延后升级而不是中断 task":
	// runRuntimePoller refuses to call ClaimTask while pauseClaims is set, and
	// tryAutoUpdate refuses to flip pauseClaims while any poller is mid-claim
	// or any task is in handleTask. Together that closes the fetch-then-claim