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#

bash

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#

bash

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:

bash

#!/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.