Skip to main content

​
Recall

Recall is the operator’s hand on the system — the CLI for direct intervention when automated systems are insufficient or you need immediate control over a running agent fleet. 
recall status          # fleet state snapshot
recall stall heike     # pause an agent immediately
recall wake heike      # resume
recall lockdown        # stop everything, now

​
The Three Control Surfaces

Recall completes the operator control plane alongside two existing commands: 
openclaw health    →  snapshot: what is the system's current state?
openclaw watch     →  live: what is changing right now?
recall             →  action: intervene directly

​
The Three Operator Primitives

These three commands cover every coordination scenario:
CommandEffectWhen to Use
recall stall <agent>Pause — agent stops acting, keeps listeningAlmost always the right first move
recall focus <agent>Leader — named agent acts, all others stallAgents stepping on each other
recall stun <agent>Reset — full intervention on a broken agentSomething is genuinely broken
stall is the dominant operator command in real production. Most incidents are coordination failures, not catastrophic failures. Stop the agent, investigate, then decide: recall wake (minor) or recall stun (broken).

​
Full Command Reference

​
System Commands

recall lockdown          # global lockdown — no new model calls, tool calls, or sub-agent spawning
recall unlock            # lift lockdown
recall status            # fleet state snapshot
recall log [agent]       # intervention history

​
Agent State

recall freeze <agent>    # stop sub-agent spawning only (agent keeps working)
recall unfreeze <agent>
recall stall <agent>     # queue messages, agent stops processing
recall stall --all       # emergency brake — pauses everything
recall wake <agent>      # reverse stall or sleep
recall wake --all
recall sleep <agent>     # disconnect from channel(s)
recall stun <agent>      # full intervention (see below)
recall quarantine <agent> # isolate for inspection

​
Recovery

recall recover <agent>   # post-stun return-to-service wizard
recall reset             # safe gateway restart with triage + backup
recall focus <agent>     # single-leader mode — all others stall
recall unfocus           # return to normal multi-agent operation

​
What Stun Actually Does

recall stun is not just a stop command. It runs a 7-step sequence:
  1. Disconnect channel connections
  2. Drain in-progress tool calls (max 10 seconds)
  3. Kill all sub-agents
  4. Archive the active session
  5. Compact memory
  6. Emit a control plane event (visible in openclaw watch)
  7. Hold — agent offline until recall recover + operator confirmation
Why step 4 matters: A gateway restart doesn’t fix a runaway agent. The bad state lives in the active session, active sub-agents, and memory files. Stun targets all three. Gateway restart misses all three.

​
Standard Incident Response

openclaw watch          # something looks odd
recall stall heike      # pause the agent immediately
radcheck risk           # assess
triage                  # capture proof bundle
agent911 diagnose       # classify the incident

# If minor:
recall wake heike

# If serious:
recall stun heike
recall recover heike
recall wake heike

​
Reversibility

Every operation has a reversal:
OperationReversal
lockdownunlock
freezeunfreeze
stallwake (queued messages delivered)
sleepwake
stunwake (after recover + confirmation)
Session context is not restored after stun — by design. If the session caused the problem, restoring it restores the problem.

​
Observability

Every Recall operation emits a structured event that flows into the reliability stack:
  • Visible in openclaw watch incident timeline
  • Ingested by RadCheck signal analysis
  • Available in OCTriage proof bundles
  • Read by Agent911 during incident diagnosis

Agent911

Recall calls Agent911 primitives. They share the same execution layer.

Triage

Run triage before stun to capture a proof bundle at the moment of intervention.