Add new Compute Management skill for cluster and warehouse selection

[CheeYuTan]

Summary

New skill documenting 6 MCP tools for Databricks compute management. No compute skill existed — agents had no guidance on how to select, start, or monitor clusters and warehouses.

Tools documented: list_clusters, list_warehouses, get_best_cluster, get_best_warehouse, get_cluster_status, start_cluster

What's in the skill

• Decision table — when to use cluster vs warehouse (SQL → warehouse, notebooks → cluster, streaming → cluster)
• Quick Start — 2-line patterns to get compute and run code
• MCP Tools Reference — all 6 tools with params, return formats, and response shapes
• Workflows — compute selection, cluster startup with polling, warehouse auto-start
• Serverless vs Classic guidance
• Common Issues — 6 real error patterns with solutions

Test evidence — 10 iterations of test-and-fix

Live MCP tool tests

Tool	Input	Output	Status
list_clusters	{}	1 cluster: Steven Tan's Personal Compute Cluster, state=RUNNING	PASS
list_warehouses	{}	1 warehouse: Serverless Starter Warehouse, state=RUNNING, size=Small	PASS
get_best_cluster	{}	{"cluster_id": "1203-135841-22k8xamq"}	PASS
get_best_warehouse	{}	"d41ad9fd669499ed" (bare string)	PASS
get_cluster_status	{cluster_id: "1203-135841-22k8xamq"}	{state: "RUNNING", message: "...running and ready for use."}	PASS
get_cluster_status	{cluster_id: "invalid-id"}	ResourceDoesNotExist error	PASS (edge case)
start_cluster	{cluster_id: "1203-135841-22k8xamq"} (already running)	Returns {state: "RUNNING"} — no previous_state field	PASS (edge case)

MCP tool schema verification

All 6 tool JSON schemas read and cross-referenced with skill documentation — parameter names, types, and required/optional status all match.

What the 10 iterations found and fixed

Iteration	What was wrong	Fix
3	execute_sql param was sql= — actual param is sql_query=	Fixed Quick Start example
4	execute_databricks_command param was command= — actual param is code=	Fixed Quick Start example
6	get_best_warehouse selection logic was wrong ("prefers smaller sizes") — actual logic is a 7-step priority algorithm	Rewrote with correct logic from MCP source code
8	start_cluster on already-running cluster returns different format (no previous_state)	Added variant to docs

Edge cases verified

Scenario	Verified via	Result
Invalid cluster ID	Live test	ResourceDoesNotExist error
start_cluster on RUNNING cluster	Live test	Returns current state, no error
No clusters available	Source code review	get_best_cluster returns {"cluster_id": null}
No warehouses available	Source code review	get_best_warehouse returns null

Test plan

• CI validation passes (validate_skills.py — 27 skills)
• SKILL.md frontmatter valid
• Registered in install_skills.sh
• All 6 MCP tool schemas verified against JSON descriptors
• 5 of 6 tools tested live (start_cluster only tested on already-running)
• All response shapes verified against real output
• 2 edge cases tested live, 2 verified via source code
• Cross-reference to databricks-config skill verified