recall
recall is the manual intervention surface of the Acme control plane. It is the operator's hand on the system — the ability to act immediately, precisely, and safely when automated recovery is insufficient or you need direct control.
recall is an Acme Agent Supply product. It requires an Acme account.
Installation
curl -fsSL https://acmeagentsupply.com/install/recall | bash
Installs the recall command globally at /usr/local/bin/recall.
The Operator Mental Model
stall = pause the brain, keep the ears open
sleep = disconnect
freeze = stop spawning only
stun = full intervention (heavy)
reset = safe gateway restart
recall stall is the dominant operator command in real production.
Most incidents are not catastrophic. They are agents stepping on each other, message storms beginning, config edits mid-flight. You rarely need to kill an agent — you need to pause it.
recall stall <agent> stops the agent from acting but keeps it listening. Investigate. Then decide: recall wake (minor issue) or recall stun (truly broken).
recall stun is heavy — it kills sub-agents, archives sessions, compacts memory. Use it when something is genuinely broken, not as a first response.
Command Reference
System Commands
recall lockdown
Activate global lockdown. No new model calls, tool executions, or sub-agent spawning across the entire system. Existing in-progress work drains normally. Use first in a runaway storm.
recall unlock
Lift global lockdown. System returns to normal operation.
recall status
Fleet state snapshot — all agents, their states, active locks.
Agent State Commands
recall stall <agent> Pause execution. Agent keeps receiving messages.
recall stall --all Emergency brake. Pauses the entire fleet.
recall freeze <agent> Block sub-agent spawning only. Agent stays active.
recall unfreeze <agent> Restore sub-agent spawning.
recall sleep <agent> Disconnect from all channels.
recall sleep <agent> --channel <id> Disconnect from one channel.
recall sleep --all
recall stun <agent> Full intervention: disconnect, drain, kill sub-agents,
archive session, compact memory.
recall quarantine <agent> Isolate for inspection. Memory read-only, no execution.
recall wake <agent> Reverse most recent sleep/stall (or stun after recover).
recall wake --all
Recovery Commands
recall recover <agent>
Post-stun recovery wizard. Shows memory diff from before the stun, allows quarantine of suspect entries, confirms readiness before reconnect. Operator confirmation required.
recall reset
Safe gateway restart. Captures a triage bundle, flushes memory, takes a snapshot, runs backup, announces to active sessions, restarts, verifies health.
Focus Commands
recall focus <agent>
Single-leader mode. The named agent acts; all others are stalled. Use when agents are stepping on each other — parallel decision-making failures, conflicting memory writes, tool contention, duplicated work.
Output on activation:
Fleet Focus Active
primary_agent: hendrik
stalled_agents: heike, gerrit, soren
recall unfocus
Release focus. All stalled agents return to their pre-focus state.
Observation Commands
recall log <agent>
Full intervention history for an agent.
Standard Incident Response
openclaw watch # something looks odd
recall stall heike # pause the agent immediately
radcheck risk # assess reliability score
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
Lockdown Workflow (Runaway Storm)
openclaw watch # signals runaway tokens / alert storm
recall lockdown # freeze the whole system first
recall stun heike # surgical per-agent cleanup
recall stun hendrik # (if multiple agents involved)
recall status # confirm all clear
recall unlock # restore normal operation
recall recover heike # review, then wake
recall wake heike
Lockdown makes surgical intervention safe. Without it, stunning one agent can trigger other agents to investigate — amplifying the storm.
Reversibility
| Command | Reversal | What is Restored |
|---|---|---|
lockdown |
unlock |
Full system execution |
freeze |
unfreeze |
Sub-agent spawning |
stall |
wake |
Processing; queued messages delivered |
sleep |
wake |
Channel connections |
stun |
wake (after recover) |
Channels; clean session |
quarantine |
recover → wake |
Full active state |
reset |
n/a (non-destructive) | Triage bundle + backup exist |
Session context is not restored after stun. If the session caused the problem, restoring it restores the problem. Archived sessions remain available for operator review.
Integration with the Reliability Stack
recall sits in the Acme reliability stack alongside RadCheck, triage, and Agent911:
triage → what was the system state when it broke?
radcheck → what is the current reliability score?
agent911 → automated: diagnose → plan → execute
recall → manual: operator-commanded intervention
recall stun calls Agent911 primitives directly — same primitives, manual trigger. No divergence between automated and manual recovery paths.
Every recall operation emits a control plane event, ingested by Bonfire and visible in openclaw watch, RadCheck signals, triage bundles, and Agent911 diagnosis context.
