Local Integration Testing — All Agents

Run end-to-end integration tests against the full CAIPE multi-agent stack in the local Docker Compose dev environment.

Prerequisites

• Docker Desktop running with sufficient resources (16 GB+ RAM recommended for all 15 agents)
• docker-compose.dev.yaml present in the repo root
• .env file configured with required API keys and agent enable flags
• MongoDB container (caipe-mongodb-dev) running

Instructions

Phase 1: Enable All Agents

1. Edit .env — ensure all agent flags are enabled:

   ENABLE_ARGOCD=true
   ENABLE_AWS=true
   ENABLE_BACKSTAGE=true
   ENABLE_CONFLUENCE=true
   ENABLE_GITHUB=true
   ENABLE_GITLAB=true
   ENABLE_JIRA=true
   ENABLE_KOMODOR=true
   ENABLE_NETUTILS=true
   ENABLE_PAGERDUTY=true
   ENABLE_SLACK=true
   ENABLE_SPLUNK=true
   ENABLE_VICTOROPS=true
   ENABLE_WEATHER=true
   ENABLE_WEBEX=true
   

2. Verify MongoDB checkpoint type:

   grep "^LANGGRAPH_CHECKPOINT_TYPE=" .env
   # Should output: LANGGRAPH_CHECKPOINT_TYPE=mongodb
   

Phase 2: Start the Stack

1. Bring up all containers:

   IMAGE_TAG=latest docker compose -f docker-compose.dev.yaml up -d
   

2. Wait for agents to be healthy (agents take 15-30 seconds to initialize):

   # Check all agent containers are running
   docker ps --filter "name=agent-" --format "table {{.Names}}\t{{.Status}}" | sort
   

3. Restart supervisor after agents are warm (avoids race condition where supervisor starts before agents are ready):

   docker restart caipe-supervisor
   

4. Verify supervisor discovered all agents:

   docker logs caipe-supervisor 2>&1 | grep -E "(ONLINE|subagents|tools)" | tail -5
   

   Expected: Deep agent updated with 15 tools and 15 subagents

Phase 3: Validate Per-Agent Checkpoint Isolation

Run the checkpoint validation script:

./skills/persistence/validate_agent_checkpoints.sh

This checks:

• Each agent container is running
• Auto-prefix log present (per-agent MongoDB collection names)
• Collections exist with documents
• No InMemorySaver fallback
• No cross-contamination between agent collections

Phase 4: Multi-Agent Routing Tests

Open the CAIPE UI at http://localhost:3000 and test multi-agent routing:

Test	Query	Expected Agents	Verification
AWS	"list EKS clusters"	supervisor → aws	checkpoints_aws has docs
ArgoCD	"show argocd version"	supervisor → argocd	checkpoints_argocd has docs
Jira	"show recent Jira issues"	supervisor → jira	checkpoints_jira has docs
Splunk	"check latest splunk logs"	supervisor → splunk	checkpoints_splunk has docs
Weather	"what's the weather in San Jose?"	supervisor → weather	checkpoints_weather has docs
Multi-agent	"list EKS clusters and show ArgoCD version"	supervisor → aws + argocd	Both checkpoint collections grow

After each query, verify checkpoint writes:

docker exec caipe-mongodb-dev mongosh "mongodb://admin:changeme@localhost:27017/caipe?authSource=admin" --quiet --eval '
  db.getCollectionNames().filter(c => c.includes("checkpoint")).sort().forEach(function(c) {
    print(c + ": " + db.getCollection(c).countDocuments() + " docs");
  });
'

Phase 5: Persistence Across Restarts

1. Note the current checkpoint counts from Phase 4
2. Restart agent containers:

   docker restart caipe-supervisor agent-aws agent-argocd agent-jira
   

3. On the same thread, ask a follow-up question:
   • "Where were you working on?" or "What did we discuss?"
4. Verify the supervisor recalls prior context from MongoDB checkpoints without re-calling subagents

Phase 6: Cross-Contamination Verification

docker exec caipe-mongodb-dev mongosh "mongodb://admin:changeme@localhost:27017/caipe?authSource=admin" --quiet --eval '
  var colls = db.getCollectionNames().filter(c => c.startsWith("checkpoints_"));
  var threadMap = {};
  colls.forEach(function(coll) {
    db.getCollection(coll).distinct("thread_id").forEach(function(tid) {
      if (!threadMap[tid]) threadMap[tid] = [];
      threadMap[tid].push(coll);
    });
  });
  var shared = 0;
  Object.keys(threadMap).forEach(function(tid) {
    if (threadMap[tid].length > 1) {
      shared++;
      print("thread " + tid.substring(0,8) + "... → " + threadMap[tid].join(", "));
    }
  });
  if (shared > 0) {
    print(shared + " threads shared across collections (expected — supervisor forwards context_id)");
  } else {
    print("No shared threads — each agent is fully isolated");
  }
'

Shared thread IDs between supervisor and subagent collections are expected — the supervisor forwards its context_id as the subagent's thread_id. What matters is that each agent's graph state is only in its own collection.

Troubleshooting

Agent not discovered by supervisor

# Check agent logs for startup errors
docker logs agent-<name> 2>&1 | tail -20

# Verify the ENABLE flag in .env
grep "ENABLE_<NAME>" .env

# Restart supervisor after agent is running
docker restart caipe-supervisor

Agent using InMemorySaver instead of MongoDB

# Check for InMemorySaver fallback in logs
docker logs agent-<name> 2>&1 | grep -i "InMemorySaver\|checkpointer"

# Verify the agent imports get_checkpointer()
grep -r "get_checkpointer\|MemorySaver\|InMemorySaver" ai_platform_engineering/agents/<name>/

Container name conflict on startup

# Remove stale containers
docker rm -f agent-<name>
docker compose -f docker-compose.dev.yaml up -d agent-<name>

Supervisor race condition (agents show OFFLINE)

Agents take 15-30 seconds to initialize. If supervisor starts first, it marks agents as offline.

# Wait for all agents, then restart supervisor
sleep 30
docker restart caipe-supervisor

Quick Reference

# Full validation (run after stack is up)
./skills/persistence/validate_agent_checkpoints.sh

# Validate specific agents
./skills/persistence/validate_agent_checkpoints.sh aws jira argocd

# Check supervisor agent discovery
docker logs caipe-supervisor 2>&1 | grep -E "subagents|ONLINE|OFFLINE" | tail -20

# MongoDB checkpoint overview
docker exec caipe-mongodb-dev mongosh "mongodb://admin:changeme@localhost:27017/caipe?authSource=admin" --quiet --eval 'db.getCollectionNames().filter(c => c.includes("checkpoint")).sort().forEach(c => print(c + ": " + db.getCollection(c).countDocuments() + " docs"))'

# Watch agent logs in real-time
docker logs -f agent-aws 2>&1 | grep -i "checkpoint\|error"

Examples

• "Test that all agents write checkpoints to isolated MongoDB collections"
• "Verify checkpoint persistence survives container restarts"
• "Run the full integration test suite locally"
• "Check if any agent fell back to InMemorySaver"

Guidelines

• Always restart the supervisor after agents are warm — race condition is a known issue
• Agent source code is cross-mounted via Docker volumes — code changes only require docker restart, not rebuild
• The validate_agent_checkpoints.sh script is the single source of truth for checkpoint health
• Shared thread_id values across supervisor and subagent collections are expected behavior
• For agents that require API keys (AWS, Jira, Splunk, etc.), ensure credentials are set in .env