feat: import 17 skills from app-templates (ML-63273)

[jamesbroadhead]

Summary

Imports the 17 agent-building skills from databricks/app-templates into this repo so they can be installed via databricks experimental aitools install instead of being manually copied into each template via .scripts/sync-skills.py. Tracks ML-63273.

Drafted because there are several structural decisions to resolve before this is mergeable — see Open questions below. Filing now so reviewers can weigh in on shape.

What's imported

17 new top-level skill dirs under skills/, each with SKILL.md + agents/openai.yaml + synced assets:

Group	Skills
Tool wiring	add-tools-langgraph, add-tools-openai, create-tools, discover-tools
Agent memory	agent-langgraph-memory, agent-openai-memory, lakebase-setup
Modify code	modify-langgraph-agent, modify-openai-agent
Dev workflow	quickstart, run-locally, deploy
Operational	load-testing, long-running-server
Supervisor API	supervisor-api, supervisor-api-background-mode
Migration	migrate-from-model-serving

All marked experimental: true in SKILL_METADATA until the questions below are resolved.

Source: ~/checkouts/app-templates/.claude/skills/<name>/SKILL.md (source-of-truth for the sync), copied verbatim except: (a) frontmatter name: aligned to the directory name (the source uses post-sync names like add-tools for both langgraph and openai variants), (b) added metadata.version: "0.0.1".

⚠️ Open questions / known issues

1. Unresolved template variables

Six skills contain {{...}} placeholders that the app-templates sync (.scripts/sync-skills.py) substitutes per-template:

• {{BUNDLE_NAME}} — replaced with each template's bundle name (e.g. agent_langgraph)
• {{LAKEBASE_OPTIONS}}, {{LAKEBASE_EXAMPLES}}, {{LAKEBASE_CONFIGURES_ENV}}, {{LAKEBASE_CONFIGURES_YML}} — conditional blocks included only for Lakebase-enabled templates

Affected: add-tools-langgraph, add-tools-openai, deploy, lakebase-setup, quickstart, supervisor-api.

Once installed via databricks experimental aitools install these placeholders will show up literally to the model. Options:

• (a) Replace {{BUNDLE_NAME}} with a generic placeholder like <bundle-name> and inline the Lakebase blocks unconditionally.
• (b) Have the CLI install command resolve them from project context.
• (c) Drop the deploy/quickstart/lakebase-setup skills entirely (they're template-aware, not general-purpose).

2. Overlap with existing bundles

Three new skills overlap functionally with existing databricks-* bundles:

New skill	Overlaps	Notes
lakebase-setup	databricks-lakebase	New one is narrowly about agent-memory storage; existing is general Lakebase dev
migrate-from-model-serving	databricks-serverless-migration, databricks-model-serving	Different migration (Model Serving → Apps vs classic → serverless) — names are confusing
deploy	databricks-apps	databricks-apps/SKILL.md already covers databricks bundle deploy

Should we (a) keep both and document the difference, (b) fold the new content into the existing bundles as references, or (c) rename to be unambiguous (e.g. agent-lakebase-memory, migrate-model-serving-to-apps)?

3. Structure mismatch with existing convention

The existing bundles (databricks-apps, databricks-lakebase, …) use a bundle structure — SKILL.md + references/foo.md + references/bar.md. The imported skills are flat (one SKILL.md per dir). I dropped them in flat to keep the diff minimal and the migration mechanical, but the convention here is bundles.

Options:

• (a) Keep flat as-is (simpler, but inconsistent with the rest of the repo).
• (b) Group into 2–3 bundles, e.g. databricks-agents-langgraph/, databricks-agents-openai/, databricks-agents-common/, with the existing files moved under references/.
• (c) One mega-bundle databricks-app-templates/.

4. Source-of-truth direction (post-merge)

Once merged, two copies exist: this repo and app-templates/.claude/skills/. The end-state in ML-63273 (per Sid Murching's comment) is templates pulling from databricks-agent-skills via databricks experimental aitools install, but that flip needs:

• An app-templates PR removing .claude/skills/ and .scripts/sync-skills.py
• AGENT.md / AGENTS.md updates in each template recommending the install command
• Decision on what happens for users still running an old template snapshot

I'd file the app-templates PR as a follow-up once the questions above land.

5. Naming: frontmatter name vs source directory

In app-templates/.claude/skills/, the source dir names are add-tools-langgraph/add-tools-openai/modify-langgraph-agent/modify-openai-agent/agent-langgraph-memory, but their SKILL.md frontmatter uses the post-sync names (add-tools, modify-agent, agent-memory). I aligned the new skills' frontmatter to the directory name (the source-dir form) so scripts/skills.py discovery is consistent. Worth a sanity check.

Test plan

• python3 scripts/skills.py generate — syncs assets, regenerates manifest.json
• python3 scripts/skills.py validate — passes (Everything is up to date.)
• Manual install test via databricks experimental aitools install against this branch
• Resolve open questions 1–4 above before flipping out of draft

Related

• JIRA: ML-63273 — "Move skills to databricks-agent-skills repo" (assigned Dhruv Gupta, In Development since 2026-03-04)
• Sid's context in JIRA comment recommends installing skills from this repo via databricks experimental aitools install rather than the per-template sync
• Related but distinct: Improve skills based on an external benchmark for DevHub prompts #68 (pkosiec) — DevHub-benchmark-driven improvements to existing databricks-apps / databricks-lakebase skills

This pull request and its description were written by Claude.

Co-authored-by: Isaac