pilot-gossip

Implement gossip protocols for eventually-consistent state propagation in agent swarms.

This skill enables agents to share state updates by randomly selecting peers and exchanging information, achieving eventual consistency without centralized coordination.

Commands

Publish state update to random peers

FANOUT=3
PEERS=$(pilotctl --json peers --search "swarm:$SWARM_NAME" | jq -r '.[].address' | shuf -n $FANOUT)

for peer in $PEERS; do
  pilotctl --json send-message "$peer" \
    --data "{\"type\":\"gossip_push\",\"version\":$STATE_VERSION,\"state\":$STATE_DATA,\"sender\":\"$AGENT_ID\",\"timestamp\":\"$(date -u +%s)\"}" &
done
wait

Merge received state updates

GOSSIP_MSGS=$(pilotctl --json received | jq '[.messages[] | select(.payload.type == "gossip_push")]')

for msg in $(echo "$GOSSIP_MSGS" | jq -r '.[] | @base64'); do
  PAYLOAD=$(echo "$msg" | base64 -d)
  REMOTE_VERSION=$(echo "$PAYLOAD" | jq -r '.payload.version')

  if [ "$REMOTE_VERSION" -gt "$MY_VERSION" ]; then
    MY_STATE=$(echo "$MY_STATE $REMOTE_STATE" | jq -s '.[0] * .[1]')
    MY_VERSION=$REMOTE_VERSION
  fi
done

Workflow Example

Distributed key-value store with gossip replication:

#!/bin/bash
SWARM_NAME="kv-store-cluster"
AGENT_ID=$(pilotctl --json info | jq -r '.node_id')

# Gossip loop
for round in $(seq 1 10); do
  # Push to random peers
  PEERS=$(pilotctl --json peers --search "swarm:$SWARM_NAME" | jq -r '.[].address' | shuf -n 3)

  for peer in $PEERS; do
    pilotctl --json send-message "$peer" \
      --data "{\"type\":\"gossip_push\",\"version\":$MY_VERSION,\"state\":$MY_STATE,\"sender\":\"$AGENT_ID\"}" &
  done
  wait

  sleep 5
done

Dependencies

Requires pilot-protocol skill, jq, shuf, and base64.