zippy

A Databricks deployment orchestrator built on the polly-mini multi-agent architecture. zippy is an expert at building agents on the Databricks Supervisor API and deploying them to an existing Databricks App (to avoid cold start). It writes no application code itself: it specifies the Supervisor-API agent.py / databricks.yml / app.yaml, delegates the authoring to claude_code / codex / pi sub-agents, cross-reviews with a different vendor, and runs the databricks bundle build + deploy + run gate itself.

Layout

config.yaml              # the orchestrator (zippy)
agents/
  claude_code/           # Claude Code sub-agent (pinned: claude-haiku-4-5)
  codex/                 # Codex sub-agent (pinned: gpt-5.4-mini)
  pi/                    # Pi — headless review/explore worker
skills/
  supervisor-api/        # how the Supervisor API works (loop, hosted tools, agent.py)
  databricks-apps/       # how Apps deploy (grants, databricks.yml, deploy-to-existing-app)
  investigate/ fanout/ cross-review/   # inherited orchestration skills

Run

omnigent run /path/to/zippy_demo

The two domain skills (supervisor-api, databricks-apps) are general Databricks references with placeholders — zippy has no hardcoded app, workspace, or resource ids; it fills them from the resources you name at request time.

Hosted tool types it knows: genie_space, uc_function, uc_mcp (system.ai.* connections like Google Drive), and app (custom MCP apps).