recovery

name	recovery
description	Recover orchestrator state after a session restart. Use at the start of any new session to discover running agents, pending tasks, and recent activity.

Session Recovery Protocol

When starting a new orchestrator session (after restart, /reload, or fresh start), follow this protocol to recover state before doing anything else.

Prerequisites

• VERS_INFRA_URL and VERS_AUTH_TOKEN must be set
• The agent-services extension must be loaded

Steps

1. Check infrastructure health

registry_discover { role: "infra" }

If no infra VM found, the coordination layer is down — you'll need to restore or recreate it before proceeding.

2. Read the board

board_list_tasks {}

Categorize tasks by status: open, in_progress, blocked, done. This is your source of truth for what work exists.

3. Read recent activity

feed_list { limit: 50 }

Scan for recent agent_started, agent_stopped, blocker_found, task_completed events to understand what happened since last session.

4. Discover running agents

registry_list { status: "running" }

Cross-reference with in_progress board tasks to see which agents are still alive and working.

If using swarm tools:

vers_swarm_status {}

5. Resume coordination

• in_progress tasks: Check if the assigned agent is still running (in registry). If not, reassign or spawn a new agent.
• blocked tasks: Read the blocker notes, resolve, and reassign.
• open tasks: Assign to available agents or spawn new ones.
• done tasks: Review results, merge PRs, close out.
• Stale registry entries: VMs with no heartbeat in 5+ min are likely dead. Clean up or re-spawn.

6. Publish recovery event

feed_publish {
  agent: "orchestrator",
  type: "custom",
  summary: "Orchestrator recovered — N agents running, M tasks in progress"
}

What Happens Automatically

The agent-services extension handles these on every agent boot — no manual action needed:

• Publishes agent_started / agent_stopped to the feed
• Registers the agent in the registry (using VERS_VM_ID)
• Sends heartbeats every 60s to keep the registration alive
• Updates registry status to stopped on shutdown

You only need to manually register VMs that were created outside the extension (e.g., infra VMs set up via vers_vm_use).

Additionally, the extension syncs skills from the SkillHub (/skills/*) to ~/.pi/agent/skills/_hub/ on session start and subscribes to SSE for live updates — so recovered agents automatically get the latest skill definitions.