(WIP) Migrate DAIS agents to Apps AI Gateway

[djliden]

Summary

• Migrates the DAIS refund and complaint custom agents from MLflow model registration + agents.deploy to Databricks Apps served with MLflow AgentServer /responses.
• Treats LLM_MODEL as a Unity AI Gateway endpoint name; refund uses ChatDatabricks(model=LLM_MODEL, use_ai_gateway=True), complaint configures DSPy through ${DATABRICKS_HOST}/ai-gateway/mlflow/v1.
• Updates stream jobs, eval notebooks, and the ops dashboard to call Apps via app URLs /responses, while leaving support_request_agent, MAS, and Knowledge Assistant serving endpoints unchanged.
• Adds a shared app helper for deterministic app names, notebook-token exchange, /responses calls, Gateway probing, and response text extraction.

Validation

• python3 -m py_compile utils/agent_app_client.py apps/refund-agent/agent.py apps/refund-agent/start_server.py apps/complaint-agent/agent.py apps/complaint-agent/start_server.py apps/caspers-ops-dashboard/app/main.py
• Parsed edited notebook code cells with ast.parse
• Static scan: no refund/complaint migrated paths use agents.deploy, DatabricksServingEndpoint, custom-agent endpoint params, or /serving-endpoints LLM invocation routes. Remaining /serving-endpoints references are MAS/KA/support/terminal paths intentionally left unchanged.
• databricks bundle validate -t all --profile sandbox
• databricks bundle deploy -t all --profile sandbox
• Verified sandbox sync contains apps/refund-agent and apps/complaint-agent; verified .git remains excluded

Not run

• Did not run the full initializer DAG or smoke-call the deployed Apps after stages. The bundle source is deployed; the agent Apps are created when the refunder_agent and complaint_agent stages run.