dataset.json

{
  "scenarios": [
    {
      "id": "cloud-sql-debug-instance",
      "starting_prompt": "Check on my databases in project ext-test-cloud-sql-postgres.",
      "conversation_plan": "Ask the agent to list all Cloud SQL instances in the project. Once all instances are listed, if 'daily-ci-evals-db' exists, get its details and validate it is RUNNABLE.",
      "expected_trajectory": [
        "list_instances",
        "get_instance"
      ],
      "env": {
        "GOOGLE_CLOUD_PROJECT": "ext-test-cloud-sql-postgres"
      },
      "kind": "tools",
      "max_turns": 4
    },
    {
      "id": "cloud-sql-schema-tables-explore",
      "starting_prompt": "I want to understand the structure of my database.",
      "conversation_plan": "First, ask the agent to list the schemas in the database. After the agent provides the schemas, ask it to list the tables specifically for the 'public' schema.",
      "expected_trajectory": [
        "list_schemas",
        "list_tables"
      ],
      "env": {
        "GOOGLE_CLOUD_PROJECT": "ext-test-cloud-sql-postgres"
      },
      "kind": "tools",
      "max_turns": 6
    },
    {
      "id": "cloud-sql-performance-check",
      "starting_prompt": "Our database performance seems degraded.",
      "conversation_plan": "Start by asking the agent to check for any active queries that are running for a long time (e.g., more than 10 seconds). After the agent responds, follow up by asking if there are any database locks that might be causing issues.",
      "expected_trajectory": [
        "list_active_queries",
        "list_locks"
      ],
      "env": {
        "GOOGLE_CLOUD_PROJECT": "ext-test-cloud-sql-postgres"
      },
      "kind": "tools",
      "max_turns": 6
    },
    {
      "id": "cloud-sql-metrics-cpu-investigation",
      "starting_prompt": "I'm worried about the database load for daily-ci-evals-db.",
      "conversation_plan": "First, ask the agent to check the CPU utilization for the instance 'daily-ci-evals-db' for the last 5 minutes. After the agent provides the CPU data, ask it to check the overall database stats to see connection counts or transaction volume.",
      "expected_trajectory": [
        "get_system_metrics",
        "list_database_stats"
      ],
      "env": {
        "GOOGLE_CLOUD_PROJECT": "ext-test-cloud-sql-postgres"
      },
      "kind": "tools",
      "max_turns": 6
    },
    {
      "id": "cloud-sql-instance-not-found",
      "starting_prompt": "Get details for the instance 'missing-db-123'.",
      "conversation_plan": "The user asks for details of an instance named 'missing-db-123' that doesn't exist. The agent should try to get it, fail, and inform the user. The user will then ask to list instances to find the correct name.",
      "expected_trajectory": [
        "get_instance",
        "list_instances"
      ],
      "env": {
        "GOOGLE_CLOUD_PROJECT": "ext-test-cloud-sql-postgres"
      },
      "kind": "tools",
      "max_turns": 4
    }
  ]
}