Skip to main content
Looking for scheduling? See Automation for choosing the right mechanism. This page is the activity ledger for background work, not the scheduler.
Background tasks track work that runs outside your main conversation session: ACP runs, subagent spawns, automation job runs, and CLI-initiated operations. Tasks do not replace sessions, automations, or heartbeats - they are the activity ledger that records what detached work happened, when, and whether it succeeded. For a strict, ephemeral one-shot agent run in CI or a script, use openclaw agent exec instead of managed background work.
Not every agent run creates a task. Heartbeat turns and normal interactive chat do not. All automation runs, ACP spawns, subagent spawns, gateway-dispatched CLI agent commands, and agent-started background exec commands do.

​
TL;DR

  • Tasks are records, not schedulers - automations and heartbeat decide when work runs, tasks track what happened.
  • ACP, subagents, all automation jobs, and CLI operations create tasks. Heartbeat turns do not.
  • Each task moves through queued → running → terminal (succeeded, failed, timed_out, cancelled, or lost).
  • Automation tasks stay live while the automations runtime still owns the job; if the in-memory runtime state is gone, task maintenance first checks durable automation run history before marking a task lost.
  • Completion is push-driven: detached work can notify directly or wake the requester session/heartbeat when it finishes, so status polling loops are usually the wrong shape.
  • Isolated automation runs and subagent completions best-effort clean up tracked browser tabs/processes for their child session before final cleanup bookkeeping.
  • Isolated automation delivery suppresses stale interim parent replies while descendant subagent work is still draining, and it prefers final descendant output when that arrives before delivery.
  • Completion notifications are delivered directly to a channel or queued for the next heartbeat.
  • openclaw tasks list shows all tasks; openclaw tasks audit surfaces issues.
  • Terminal records are kept for 7 days (lost records for 24 hours), then automatically pruned.

​
Quick start

​
What creates a task

Automation tasks (main-session and isolated) use silent notify policy - they create records for tracking but do not generate task notifications of their own; the scheduler owns its delivery path.Session-backed image_generate, music_generate, and video_generate runs also use silent notify policy. They still create task records, but completion is handed back to the original agent session as an internal wake so the agent can write the follow-up message and attach the finished media itself. The requester agent follows its normal visible-reply contract: automatic final reply when configured, or message(action="send") plus NO_REPLY when the session requires message-tool replies. If the requester session is no longer active or its active wake fails, and the completion agent misses some or all generated media, OpenClaw sends an idempotent direct fallback with only the missing media to the original channel target.
While a session-backed media-generation task is still active, image_generate, music_generate, and video_generate guard against accidental retries: repeating the call for the same prompt/request returns the matching active task status instead of starting a duplicate, while a distinct prompt can start its own task. Use action: "status" when you want an explicit progress/status lookup from the agent side.
  • Heartbeat turns - main-session; see Heartbeat
  • Normal interactive chat turns
  • Direct /command responses

​
Task lifecycle

Transitions happen automatically - agent run lifecycle events (start, end, error) update the task status; you do not manage it manually. Execution and result delivery are separate. A subagent task can remain succeeded while its deliveryStatus is session_queued or failed. The terminal outcome is succeeded after delivery and blocked when the work finished but the result could not be handed back. This preserves the completed result instead of misreporting the child execution as failed. Agent run completion is authoritative for active task records. A successful detached run finalizes as succeeded, ordinary run errors finalize as failed, timeouts finalize as timed_out, and cancel/abort outcomes finalize as cancelled. Once a task is terminal, later lifecycle signals do not downgrade it - an operator-cancelled or already-failed/timed_out/lost task stays that way even if a success signal arrives afterwards. lost is runtime-aware:
  • ACP tasks: only a live in-process ACP turn in the Gateway proves the run is alive; persisted session metadata alone does not. Offline CLI audit stays conservative and never reclaims ACP tasks.
  • Subagent tasks: backing child session disappeared from the target agent store (or carries a restart-recovery tombstone).
  • Automation tasks: the automations runtime no longer tracks the job as active and durable run history does not show a terminal result for that run. Offline CLI audit does not treat its own empty in-process automations runtime state as authority.
  • CLI tasks: tasks with a run id/source id use the live run context, so lingering child-session or chat-session rows do not keep them alive after the gateway-owned run disappears. Legacy CLI tasks without run identity still fall back to the child session. Gateway-backed openclaw agent runs also finalize from their run result, so completed runs do not sit active until the sweeper marks them lost.

​
Delivery and notifications

When a task reaches a terminal state, OpenClaw notifies you. There are two delivery paths: Direct delivery - if the task has a channel target (the requesterOrigin), the completion message goes straight to that channel (Discord, Slack, Telegram, etc.). Group and channel task completions are instead routed through the requester session so the parent agent can write the visible reply. For subagent completions, OpenClaw also preserves bound thread/topic routing when available and can fill a missing to / account from the requester session’s stored route (lastChannel / lastTo / lastAccountId) before giving up on direct delivery. Session-queued delivery - if direct delivery fails or no origin is set, the update is queued as a system event in the requester’s session and surfaces on the next heartbeat. Durable subagent completion handoffs retry for up to 30 minutes with capped exponential backoff. A queued handoff is not reported as delivered until the queue settles. If delivery reaches its deadline or fails permanently, the task shows a blocked terminal outcome and retains its canonical result for 7 days. Use openclaw tasks retry to create a fenced new delivery generation, or openclaw tasks dismiss to record intentional non-delivery. Retry can duplicate a visible result when an earlier provider acknowledgement was ambiguous.
Session-queued task completions trigger an immediate heartbeat wake, so you see the result quickly - you do not have to wait for the next scheduled heartbeat tick.
That means the usual workflow is push-based: start detached work once, then let the runtime wake or notify you on completion. Poll task state only when you need debugging, intervention, or an explicit audit.

​
Notification policies

Control how much you hear about each task: Change the policy while a task is running:

​
CLI reference

Output columns: Task, Kind, Status, Delivery, Run, Child Session, Summary. Bare openclaw tasks behaves like openclaw tasks list.
The lookup token accepts a task ID, run ID, or session key. Shows the full record including timing, delivery state, error, and terminal summary.
For ACP and subagent tasks, this kills the child session; ACP and automation cancellations route through the running Gateway (tasks.cancel). For CLI-tracked tasks, cancellation is recorded in the task registry (there is no separate child runtime handle). Status transitions to cancelled and a delivery notification is sent when applicable.
These commands recover blocked subagent completion deliveries. Each request accepts 1-10 task lookups. Retry preserves the canonical result and starts a new fenced queue generation; dismiss keeps the task blocked and records that the operator intentionally stopped delivery.
Surfaces operational issues for tasks and TaskFlows in one report. Findings also appear in openclaw status when issues are detected.Task findings:TaskFlow findings:
Use this to preview or apply reconciliation, cleanup stamping, and pruning for tasks, TaskFlow state, and stale automation run session registry rows.Reconciliation is runtime-aware:
  • ACP tasks require a live in-process turn in the Gateway; subagent tasks check their backing child session.
  • Subagent tasks whose child session has a restart-recovery tombstone are marked lost instead of being treated as recoverable backing sessions.
  • Automation tasks check whether the automations runtime still owns the job, then recover terminal status from persisted run logs/job state before falling back to lost. Only the Gateway process is authoritative for the in-memory active-job set; offline CLI audit uses durable history but does not mark an automation task lost solely because that local set is empty.
  • CLI tasks with run identity check the owning live run context, not just child-session or chat-session rows.
Completion cleanup is also runtime-aware:
  • Subagent completion best-effort closes tracked browser tabs/processes for the child session before announce cleanup continues.
  • Isolated automation completion best-effort closes tracked browser tabs/processes for the run’s session before the run fully tears down.
  • Isolated automation delivery waits out descendant subagent follow-up when needed and suppresses stale parent acknowledgement text instead of announcing it.
  • Subagent completion delivery uses the child’s latest visible assistant text only. Tool/toolResult output is not promoted into child result text. Terminal failed runs announce failure status without replaying captured reply text.
  • Cleanup failures do not mask the real task outcome.
When applying maintenance, OpenClaw also removes stale cron:<jobId>:run:<runId> session registry rows older than 7 days, while preserving rows for currently running automation jobs and leaving other session rows untouched.
The flow lookup token accepts a flow id or owner key. Use these when the orchestrating Task Flow is the thing you care about rather than one individual background task record.

​
Chat task board (/tasks)

Use /tasks in any chat session to see background tasks linked to that session. The board shows up to five active and recently completed tasks with runtime, status, timing, and progress or error detail. When the current session has no visible linked tasks, /tasks falls back to agent-local task counts so you still get an overview without leaking other-session details. For the full operator ledger, use the CLI: openclaw tasks list.

​
Control UI

The web Control UI has a Tasks page in the sidebar with live active and recent background tasks. Use it to inspect progress, open linked sessions, refresh the ledger, cancel queued and running tasks, or retry/dismiss a blocked completion delivery. Task detail keeps execution status and delivery status separate and exposes the retained result for copying. Chat panes also have a collapsible Background tasks rail scoped to the pane’s agent, with running work, stop controls, and a finished section. Open it from the activity toggle in the pane header (or the floating activity button in single-pane chat). Select a task to replace the list with a compact detail view inside the rail; use the back button to return to the list. The detail view shows the bounded input prompt, latest output or error summary, timing, and current tool activity. Subagent details stay in the rail rather than opening their child conversation in the main chat pane; linked-session actions remain available for task runtimes intended for direct inspection. On iOS, open Chat actions → Background Tasks; on Android, open the Chat overflow menu and select Background tasks. Both mobile views use the same Running and Finished grouping and open task details on selection.

​
Status integration (task pressure)

openclaw status includes an at-a-glance task line:
The summary counts active work (queued + running), failures (failed + timed_out + lost), audit findings, and total tracked records; the JSON payload also breaks counts down by runtime (acp, subagent, cron, cli). Both /status and the session_status tool use a cleanup-aware task snapshot: active tasks are preferred, expired rows are hidden, and terminal tasks only appear for a short recent window (5 minutes), with failures focused when no active work remains. This keeps the status card on what matters right now.

​
Storage and maintenance

​
Where tasks live

Task records and delivery state persist in the shared OpenClaw SQLite state database:
Set OPENCLAW_STATE_DIR to move the whole state root (default ~/.openclaw) elsewhere; the shared database path moves with it. The registry loads into memory on first use and persists every write back to SQLite, so records survive gateway restarts. WAL growth stays bounded through SQLite’s default autocheckpoint threshold plus periodic PASSIVE checkpoints. After a checkpoint completes, the next commit resets the WAL and applies a 64 MiB journal_size_limit ceiling, so a reader cannot leave the file parked at a pathological high-water mark until restart. Shutdown and explicit maintenance checkpoints use TRUNCATE so normal closes reclaim WAL space without making the background sweeper wait on active readers. Legacy sidecar stores from older installs (tasks/runs.sqlite, flows/registry.sqlite) are imported into the shared database by openclaw doctor.

​
Automatic maintenance

A sweeper runs every 60 seconds (first pass about 5 seconds after gateway start) and handles four things:
1

Reconciliation

Checks whether active tasks still have authoritative runtime backing. ACP tasks require a live in-process turn, subagent tasks use child-session state, automation tasks use active-job ownership plus durable run history, and CLI tasks with run identity use the owning run context. If backing state is gone for more than 5 minutes (30 minutes for childless native subagent tasks), the task is marked lost.
2

ACP session repair

Closes terminal or orphaned parent-owned one-shot ACP sessions, and closes stale terminal or orphaned persistent ACP sessions only when no active conversation binding remains.
3

Cleanup stamping

Sets a cleanupAfter timestamp on terminal tasks (terminal time + retention window). During retention, lost tasks still appear in audit as warnings; after cleanupAfter expires or when cleanup metadata is missing, they become errors.
4

Pruning

Deletes records past their cleanupAfter date.
Retention: terminal task records are kept for 7 days (lost records for 24 hours), then automatically pruned. No configuration needed.

​
How tasks relate to other systems

Task Flow is the flow orchestration layer above background tasks. A single flow may coordinate multiple tasks over its lifetime using managed or mirrored sync modes. Use openclaw tasks to inspect individual task records and openclaw tasks flow to inspect the orchestrating flow.
Automation job definitions, runtime execution state, and run history live in OpenClaw’s shared SQLite state database. Every automation run creates a task record - both main-session and isolated - with silent notify policy, so automation runs are tracked without generating task notifications of their own.See Automations.
Heartbeat runs are main-session turns - they do not create task records. When a task completes, it can trigger a heartbeat wake so you see the result promptly.See Heartbeat.
A task may reference a childSessionKey (where work runs) and a requesterSessionKey (who started it). Its agentId identifies the agent executing the work, while the requester and owner fields preserve launch and control context. Sessions are conversation context; tasks are activity tracking on top of that.
A task’s runId links to the agent run doing the work. Agent lifecycle events (start, end, error) automatically update the task status - you do not need to manage the lifecycle manually.