WorkspaceTaskRepository

Workspace-backed TaskRepository that uses WorkspaceManager as the authoritative truth source for task state, while maintaining in-memory BackgroundTask handles as a local performance overlay for tasks running on the current node.

Storage layout: agents/<parentAgentId>/tasks/<sessionId>.json — a JSON map of taskId → TaskRecord, consistent with how sessions are stored. In distributed deployments using RemoteFilesystemSpec, this path is automatically routed to shared storage, making task state visible to any node. In sandbox mode the injected filesystem is a per-call proxy with no live sandbox between calls, so task records are persisted on the host workspace instead of inside the sandbox — see the task-record routing in WorkspaceManager.

The in-memory localTasks map is keyed by "<sessionId>:<taskId>" to preserve session isolation when multiple sessions coexist in the same process.

Distributed semantics:

• Task execution is sticky to the originating node (the node that called #putTask).

• Any node can read task status via #getTask or #listTasks by falling back to workspace records when no local future exists.

• block=true on a non-originating node degrades gracefully to reading the latest persisted terminal state without hanging.

• Cancellation sets a TaskRecord#isCancelRequested() flag in workspace storage; the originating node checks this flag before invoking the subagent for best-effort cancel.

• Remote TaskRunSpec.RemoteTaskRunSpec tasks use a RemoteSubagentTransport (Agent Protocol by default) and persist TaskRecord#getRemoteBaseUrl() for cross-node resume.

• Orphan sweeping is throttled via PeriodicGate: use io.agentscope.harness.agent.coordination.StoreBackedPeriodicGate for cross-replica deduplication, or LocalPeriodicGate (the default) for single-process deployments. Prefer a control-plane hosted TaskRepository when available — that path runs a leader-only sweep and does not use this workspace repository at all.