feat(plugin): per-provider bundles under plugins/databricks/<provider>/

[simonfaltum]

Why

The plugin shipped from a single whole-repo source (marketplace.json source: "./"), which caused real problems. Copilot crashed copying the repo's .git/fsmonitor--daemon.ipc socket (upstream #3842), Codex mis-resolved hook paths from the repo root, and every install cloned the entire repo even though each agent only uses its own dialect's files. There was also no easy way to see, per provider, exactly what ships.

Changes

Before: one source: "./" per catalog; all four agents cloned the whole repo; commands/ and commands-cursor/ were maintained as two near-duplicate trees.

Now: each agent fetches a generated, self-contained plugins/databricks/<provider>/ folder (claude, codex, copilot, cursor) holding only what it uses: its plugin.json, a copy of skills/, its hook wiring as hooks/hooks.json plus the scripts it references, and where applicable commands/, rules/, assets/. The four root marketplace.json catalogs (Cursor's is new) each point a scoped source at their own subfolder, so an install fetches only that provider's payload.

Implementation:

• skillsgen/bundle.py: assembles each provider folder and byte-exact drift-checks it (a missing, stale, hand-edited, or extra file all fail CI).
• Uniform hooks/hooks.json per provider, auto-discovered from the plugin root, so no plugin.json declares a "hooks" path. The per-dialect names (codex-hooks.json, ...) remain only at the repo root so the single generation step does not collide; the bundle renames each into its folder.
• skillsgen/commands.py + templated commands/*.md: one source per command with a {{claude-or-codex | cursor}} alternation, rendered per provider. commands-cursor/ removed.
• plugins.py: catalogs scoped per provider via metaplugin/plugin.meta.json marketplace.source; ref: main for now (a mechanical follow-up flips it to v{version} tag-pinning).
• CLI files-channel unchanged: the stable repo_dir stays skills (root skills/), so manifest.json and the released CLI are untouched.
• .gitattributes marks plugins/** linguist-generated; release.yml stages plugins/databricks; validate-manifest.yml triggers on plugins/**.
• Docs (CLAUDE / AGENTS / CONTRIBUTING / READMEs / metaplugin) and tests updated.

Everything under plugins/ is generated, committed, and drift-checked (not gitignored), so a reviewer sees the exact bytes each agent fetches.

Test plan

• python3 scripts/skills.py validate clean
• 110 unit tests pass (python3 -m unittest discover -s tests -p '*_test.py')
• Verified empirically against a throwaway GitHub repo: scoped install on all four agents; tag-freeze on Claude/Codex/Copilot (a tagged add/remove sequence serves only the tagged contents); plugin load on Cursor (cursor-agent --plugin-dir); the Copilot fsmonitor fix (the whole-repo control crashes, the scoped install does not); the CLI files-channel path; and the ./ -> scoped migration on a version bump.
• Codex still needs its policy block plus a live router-fire check (an open question tracked in the proposal).

Follow-up (separate, mechanical PR): flip ref_template from main to v{version} so the ref-capable tools (Claude, Codex, Copilot) serve frozen release tags instead of main HEAD. Cursor cannot pin a ref and always tracks the default branch.

This pull request and its description were written by Isaac.