[antonyprasad-db]

What and why

The Databricks Power runs the ai-dev-kit installer, copies its skills into the Power's steering/ directory, and ships a disabled entry for ai-dev-kit's local MCP server, which upstream has deprecated. Databricks CLI v1.20.0 (released 2026-10-07) added Kiro as a databricks aitools target (databricks/cli#6908), so this moves the Power to the official Databricks agent skills, installed and tracked by the CLI, and repackages it in the Agent Plugins v1.0.0 format, following power-builder's migrate-plugin guide.

Changes

• databricks/plugin.json: replaces the POWER.md frontmatter (name, display name, description, author), with six specific keywords.
• databricks/skills/databricks-kiro-setup/SKILL.md: the former POWER.md body as one skill, per migrate-plugin for a single POWER.md under 800 lines. Activation checks, CLI version and databricks auth login, databricks aitools install --agents kiro, the eight MLflow skills the ai-dev-kit installer also installed, a new chat session, governed data access through Databricks-provided MCPs in Unity Gateway, updating and removing skills, migration, and troubleshooting. Kiro doesn't load the dev.kiro namespace yet, so the activation guidance lives in the skill's On activation section.
• databricks/POWER.md, databricks/mcp.json, databricks/steering/: removed. The Databricks skills live in ~/.kiro/skills/, managed by the CLI.
• README.md: catalog entry renamed to Databricks, with MCP Servers: None (Knowledge Base Power).

Why the Power ships no MCP server

The governed option is a Databricks-provided MCP such as system.ai.genie_one_mcp, served by the user's own workspace at https://<workspace-hostname>/ai-gateway/mcp-services/<name> and called with the user's credentials. A plugin mcp.json can't express that entry: the spec expands only ${PLUGIN_ROOT} and ${PLUGIN_DATA}, and only in stdio args, env and cwd, so neither the workspace host nor a token reference can be filled in, and there is no disabled field to ship it off by default (Kiro 1.1.237's loader skips a server that has one). The skill's governed data access section walks users through adding the entry to their own ~/.kiro/settings/mcp.json, where Kiro does expand ${...} from the environment.

Upgrade path for existing users

Updating the Power reinstalls it, which deletes the old installed folder (including the skill copies in steering/) and its powers.mcpServers entry. The skill then has users run the CLI install and lists what lives outside the Power: a top-level MCP entry from ai-dev-kit's optional MCP installer, old credential exports in the shell profile, and the ~/.ai-dev-kit clone.

Testing

• Packaging: Kiro 1.1.237's own Agent Plugin loader code, run against the directory, loads the manifest, display name and the databricks-kiro-setup skill with no issues (tested at 54af359). As a check on the harness, a wrong $schema, an unknown manifest field, an invalid skill directory name and a disabled MCP server each produce the expected rejection or issue. The review follow-up only removes dev.kiro/INSTRUCTIONS.md and the one license line that named it; the manifest and the skill's frontmatter are unchanged.
• Commands: two commands changed from the previous commit and were re-run: the leftover check now matches only ai-dev-kit's databricks-mcp-server/run_server.py (it matches that entry and not an unrelated uvx databricks-mcp-server), and the MLflow step restores a copy left by an interrupted run (with the old block that copy was deleted; with the new one it survives). With the released Databricks CLI v1.20.0 in a clean HOME, each one was run in an interactive zsh (zsh -f -i, where # is not a comment):
  • databricks aitools install --agents kiro --scope global installs 33 skills into ~/.kiro/skills/; aitools update, update --force and uninstall --agents kiro behave as documented.
  • The MLflow step installs all eight skills and exits 0; with a failed clone it keeps existing copies and exits non-zero.
  • None of the 41 skills exceeds Kiro's SKILL.md name and description limits.
  • The activation checks run offline and print no secrets.

Older CLIs (v1.19 and earlier) reject --agents kiro with unknown agent "kiro", so the skill checks the version first. The Agent Plugins format needs Kiro IDE 1.0.288 or later (changelog).

@nadetastic, would you take a look when you have a moment? Thanks!

This pull request and its description were written by Isaac.