haiven-orchestrator User Guide

Health Check

curl -sf http://localhost:8500/health

Basic Orchestration Call

curl -s -X POST http://localhost:8500/orchestrate \
  -H "Content-Type: application/json" \
  -d '{
    "message": "What does my day look like?",
    "input_modality": "text",
    "output_format": "markdown"
  }' | jq .

Session Continuity

Pass the returned session_id back on later requests:

curl -s -X POST http://localhost:8500/orchestrate \
  -H "Content-Type: application/json" \
  -d '{
    "message": "Which of those tasks should I do first?",
    "session_id": "e7d2a1f9-3b4c-5d6e-7f8a-9b0c1d2e3f4a"
  }' | jq .

Runtime Classifier Note

The deployed runtime currently uses:

The source fallback in app/config.py is still qwen3.5-27b, so if you are debugging source-level defaults, keep that distinction in mind.

When Clarification Happens

If classifier confidence falls below 0.7, the service returns:

No agent dispatch happens in that case.

Troubleshooting

All responses fail

curl -sf http://localhost:4000/health/liveliness
docker logs --tail 100 haiven-orchestrator
docker logs --tail 100 litellm

All responses have low confidence

docker logs --tail 100 haiven-orchestrator 2>&1 | grep -i "classifier_parse_failed\|classifier_unknown_intent"

Verify the deployed classifier setting:

docker inspect haiven-orchestrator --format '{{range .Config.Env}}{{println .}}{{end}}' | grep ORCH_CLASSIFIER_MODEL

Sessions are not persisting

docker exec redis redis-cli -n 1 PING
docker exec redis redis-cli -n 1 KEYS "orch:session:*"

Backend dispatch failures

docker logs --tail 50 agent-briefing
docker logs --tail 50 work-hub
docker logs --tail 50 haiven-knowledge
docker logs --tail 50 research-agent

Intent Set

The current runtime taxonomy has 17 intents: