microdao-daarion/docs/ops/experience_bus_phase2.md

Experience Bus Phase-2 (Lessons Extractor)

Scope

• Source stream: EXPERIENCE
• Source subjects: agent.experience.v1.>
• Consumer mode: durable pull + explicit ack
• Output table: agent_lessons (append-only)
• Output subject: agent.lesson.v1 (optional publish)

Service

• Container: dagi-experience-learner-node1
• Endpoint:
  • GET /health
  • GET /metrics

Environment

• NATS_URL=nats://nats:4222
• EXPERIENCE_STREAM_NAME=EXPERIENCE
• EXPERIENCE_SUBJECT=agent.experience.v1.>
• EXPERIENCE_DURABLE=experience-learner-v1
• EXPERIENCE_ACK_WAIT_SECONDS=30
• EXPERIENCE_MAX_DELIVER=20
• EXPERIENCE_FETCH_BATCH=64
• EXPERIENCE_FETCH_TIMEOUT_SECONDS=2
• EXPERIENCE_WINDOW_SECONDS=1800
• EXPERIENCE_OK_SAMPLE_PCT=10
• EXPERIENCE_LATENCY_SPIKE_MS=5000
• EXPERIENCE_ERROR_THRESHOLD=3
• EXPERIENCE_SILENT_THRESHOLD=5
• EXPERIENCE_LATENCY_THRESHOLD=3
• EXPERIENCE_EVENT_DEDUP_TTL_SECONDS=3600
• LEARNER_DATABASE_URL=postgresql://<user>:<pass>@<host>:5432/daarion_memory
• LESSON_SUBJECT=agent.lesson.v1
• LESSON_PUBLISH_ENABLED=true

Deploy

1. Apply migration migrations/055_agent_lessons.sql.
2. Deploy service experience-learner.
3. Verify service health and metrics.

Smoke

# Generate event traffic (Phase-1 router path)
for i in $(seq 1 50); do
  agent=$([ $((i%2)) -eq 0 ] && echo "aistalk" || echo "devtools")
  curl -sS -m 8 -o /dev/null \
    -X POST "http://127.0.0.1:9102/v1/agents/${agent}/infer" \
    -H "content-type: application/json" \
    -d "{\"prompt\":\"phase2-smoke-${agent}-${i}-$(date +%s%N)\"}" || true
done

Verify

# Lessons rows
docker exec dagi-postgres psql -U daarion -d daarion_memory -tAc \
  "SELECT count(*) FROM agent_lessons WHERE ts > now()-interval '30 minutes';"

# Idempotency check (run again, duplicates should not explode)
docker exec dagi-postgres psql -U daarion -d daarion_memory -tAc \
  "SELECT count(*), count(distinct lesson_key) FROM agent_lessons;"

# Learner metrics
curl -sS http://127.0.0.1:9109/metrics | grep -E 'lessons_|js_messages_'

Acceptance

• agent_lessons receives rows under live event flow.
• Reprocessing/redelivery does not duplicate lessons (lesson_key unique).
• js_messages_acked_total increases.
• js_messages_redelivered_total is observable when replay/redelivery occurs.