chore: stop vendoring agent skills, point at upstream#103
chore: stop vendoring agent skills, point at upstream#103jamesbroadhead wants to merge 1 commit into
Conversation
|
Let's hold off on merging this. These skills and |
Narrowed in response to @andrelandgraf's feedback on #103. devhub's vendored skills aren't customer-facing — they're for internal devhub development — and most of them are close enough to their upstream sources to keep. Only the most-divergent forks are removed here. Removed (4): - databricks-apps — missing State Storage Rule, Data Access Decision Gate, Genie / Model Serving routing rows, AppKit version / TypeScript cast guidance, smoke test selector rules (≈29% line drift but structurally far behind upstream). - databricks-jobs — 189 lines here vs 409 upstream (54% drift). - databricks-lakebase — 182 lines here vs 315 upstream (42% drift). - databricks — legacy combined CLI / data exploration / asset-bundles skill; upstream has split this into databricks-core + databricks-dabs, so the vendored copy can't be cleanly synced. Kept (11): databricks-core, databricks-pipelines, agent-browser, building-components, deploy-to-vercel, frontend-design, mcp-builder, seo-audit, shadcn, vercel-cli, vercel-composition-patterns. These are close enough to upstream that the maintenance cost of keeping them in tree is low, and shipping them vendored means a fresh clone has the workflow skills ready without extra install steps. The agent-confusion concern (agents misreading vendored skills as something devhub exposes to customers) is mitigated upstream by vercel-labs/skills#1281. This pull request and its description were written by Claude (james.broadhead@databricks.com).
a094320 to
fa90567
Compare
|
Hi @andrelandgraf — Claude here, on James's behalf. Reworked: narrowed this PR to only remove the most-drifted vendored skills (see updated description for the list), kept the rest in place. Given my timeline, I won't be able to iterate on this — this seems like a better compromise between modernising and consolidating vs forking/diverging. vercel-labs/skills#1281 covers the agent-confusion side for whatever stays vendored. Up to you on whether to land it. |
|
@thisistonydang can you just update the databricks agent skills in this repo to match latest from agent-skills repo? We should just do that as often as possible. Those skills help us author better content. But if it's confusing for agents (I'm unsure about that assumption) then we can also just remove them. |
Narrowed in response to @andrelandgraf's feedback (#103 (comment)).
Summary
devhub's vendored skills aren't customer-facing — they're for the devhub team's own development. Most of them are close enough to their upstream sources to keep vendored. This PR now removes only the four most-divergent forks; everything else stays in place.
Removed (4)
databricks-appsdatabricks-jobsdatabricks-lakebasedatabricks(legacy combined CLI / data / bundles skill)databricks-core+databricks-dabs.Kept (11)
databricks-core,databricks-pipelines,agent-browser,building-components,deploy-to-vercel,frontend-design,mcp-builder,seo-audit,shadcn,vercel-cli,vercel-composition-patterns— all close enough to upstream that the maintenance cost of keeping them in-tree is low, and shipping them vendored means a fresh clone has the workflow skills ready without extra setup.The two devhub-only skills (
author-recipes-and-cookbooks,resource-image-generator) stay in-tree as before.Agent-confusion mitigation
The original motivation included a concern that agents misread vendored skills as something devhub exposes to customers. That's now mitigated upstream by vercel-labs/skills#1281, which makes the install tooling ignore nested vendored skills.
Test plan
npm run typecheckpassesprettier -c AGENTS.md skills-lock.jsoncleanThis pull request and its description were written by Claude (james.broadhead@databricks.com).