Add Resource Tracking reference for demo/POC cleanup

[CheeYuTan]

Summary

Documents list_tracked_resources and delete_tracked_resource MCP tools which had zero skill coverage. These tools automatically track every resource created via MCP (dashboards, jobs, pipelines, catalogs, volumes, etc.) for cleanup after demos and POCs.

What's in the file (258 lines)

• Overview — what resource tracking is and why it matters
• 15 tracked resource types — dashboard, job, pipeline, genie_space, catalog, schema, volume, app, lakebase_instance, vs_endpoint, vs_index, metric_view, knowledge_assistant, multi_agent_supervisor, lakebase_project
• MCP tools reference — list_tracked_resources (type filter) and delete_tracked_resource (manifest-only vs Databricks deletion)
• 4 workflows — post-demo cleanup, targeted cleanup by type, manifest-only removal, audit before workspace handoff
• Behavior notes — safe defaults, cross-session persistence, upsert on recreate, deletion support matrix
• Troubleshooting — 5 real error patterns

Test evidence — 10 iterations

Verification method

MCP server was intermittently unavailable during testing. Verification done via:

1. MCP tool JSON schemas — parameter names, types, defaults all confirmed
2. MCP server source code — track_resource(), list_tracked_resources(), delete_tracked_resource() implementations reviewed
3. Prior live tests — 7 MCP calls from earlier session (list all, filter by dashboard/volume, delete non-existent)

What testing found and fixed (6 inaccuracies)

Fix	Before (wrong)	After (correct)	Evidence
Resource types count	9 types listed	15 types (added app, lakebase_instance, lakebase_project, vs_endpoint, vs_index, metric_view)	Source code: 15 modules call track_resource()
Schema ID format	"Three-level name"	Two-level name (catalog.schema)	Source: resource_id=f"{catalog}.{schema}"
URL field presence	"dashboards and some other types"	Dashboards only	Source: only aibi_dashboards.py passes url=
Overview resource list	Missing catalogs, MAS, apps, Lakebase, VS	All 15 types listed	Source code audit
Volume vs schema ID	Both described as "three-level"	Schema=two-level, Volume=three-level	Source code
Deletion support	Not documented	10 types support Databricks deletion, 5 are manifest-only	Source: _RESOURCE_DELETERS dict has 10 entries

Prior live test results (from earlier session)

Tool	Input	Result	Status
list_tracked_resources	{}	5 resources (1 catalog, 2 volumes, 1 dashboard, 1 genie_space)	PASS
list_tracked_resources	{type: "dashboard"}	1 dashboard with url field	PASS
list_tracked_resources	{type: "volume"}	2 volumes, no url field	PASS
list_tracked_resources	{type: "job"}	{"resources": [], "count": 0}	PASS
delete_tracked_resource	{type: "job", resource_id: "nonexistent"}	{"success": false, "error": "Resource job/nonexistent not found"}	PASS
Re-list after failed delete	{}	All 5 resources intact	PASS

Test plan

• CI validation passes
• Both MCP tool schemas verified against JSON descriptors
• All 15 resource types verified against source code
• Response shapes verified against source code
• Safe-by-default (delete_from_databricks: false) confirmed in schema
• 6 inaccuracies found and fixed
• install_skills.sh updated with new extra file