mirror of
https://github.com/TanStack/ai.git
synced 2026-11-16 19:38:41 +00:00
feat: ship repo-level agent skills for TanStack AI capability discovery (#1348)
* feat: ship a Claude Code marketplace plugin for TanStack AI skill discovery Adds `.claude-plugin/marketplace.json` so the repo installs as a plugin marketplace, plus a single `tanstack-ai` plugin. The plugin is a discovery layer only: it maps a capability to the package that provides it, gets that package installed, then hands off to the package's own SKILL.md, which is versioned with the code it teaches. No skill text is vendored. The exhaustive package catalog is generated from packages/* so a new package shows up without a hand edit; a test fails when the checked-in catalog drifts. Closes #995 Claude-Session: https://claude.ai/code/session_01SR99WG8nMeBtzCP9qB8b1t * docs: document the Claude Code plugin install path for agent skills Claude Code users can add the marketplace and install the plugin in two commands. The Intent steps stay the path for a whole repo and for agents without a plugin marketplace. Claude-Session: https://claude.ai/code/session_01SR99WG8nMeBtzCP9qB8b1t * refactor: move the skills to a repo-level skills/ directory The skills are a repo artifact, not a plugin artifact. `skills/` at the root makes them readable in place and installable with `npx skills add TanStack/ai`. The Claude Code and Cursor plugin manifests now point at the repo root, so one skill set serves the Agent Skills CLI and both plugin marketplaces. Claude-Session: https://claude.ai/code/session_01SR99WG8nMeBtzCP9qB8b1t * feat: split discovery into six repo-level skills One skill could not carry provider choice, client wiring, agent capabilities, state, and migration without becoming a page nobody routes into. Each area is now its own skill with its own generated package catalog, and the entry skill routes between them. The generator sorts every published package into exactly one area, with provider adapters as the default, so a new provider needs no edit. The test asserts the catalogs are current and that no package is dropped. Claude-Session: https://claude.ai/code/session_01SR99WG8nMeBtzCP9qB8b1t * ci: apply automated fixes * feat(skills): suggest TanStack AI on generic AI-feature asks, type-check skills with kiira The skill descriptions now trigger on "add AI to my app", "which AI SDK", "build an agent", "chat UI", and similar asks that never name TanStack AI, and the entry skill recommends @tanstack/ai when nothing is installed. When another SDK is already in use it says so once, points at the migration skill, and then helps with what was asked. kiira now includes skills/**/*.md, so the repo-level skills' code fences type-check with the docs. Package skills under packages/*/skills stay out: they have 748 errors across 201 snippets today and need their own cleanup. Claude-Session: https://claude.ai/code/session_01SR99WG8nMeBtzCP9qB8b1t * fix(scripts): emit formatter-stable tables from the catalog generator The CI autofix runs oxfmt over markdown and pads table columns, which made the committed catalogs drift from the generator's output. The generator now pads the same way, so the drift test and the formatter agree. Claude-Session: https://claude.ai/code/session_01SR99WG8nMeBtzCP9qB8b1t * test(scripts): match padded catalog rows Claude-Session: https://claude.ai/code/session_01SR99WG8nMeBtzCP9qB8b1t * docs(skills): install with gh skill instead of npx skills gh skill install takes one skill per call and --all would also copy every packages/*/skills entry, so the docs install the entry skill and it fetches the sibling it routes to on demand. Claude-Session: https://claude.ai/code/session_01SR99WG8nMeBtzCP9qB8b1t * docs: install repo-level skills with npx skills, add Agent Skills to the README top gh skill is out: one skill per call and --all copies package skills too. npx skills add takes a space-separated --skill list and skips packages/*/skills. Claude-Session: https://claude.ai/code/session_01SR99WG8nMeBtzCP9qB8b1t * refactor(skills): drop the catalog generator, use TanStack Intent for discovery intent list shows the installed intent-enabled packages and their skills, and intent load prints any skill or sub-skill by <package>#<skill>. The skills now route through those instead of ls/cat, and point at the docs for the packages that exist but are not installed. That removes the generated packages.md catalogs, the generator, its test, and the test:maintainer wiring. Claude-Session: https://claude.ai/code/session_01SR99WG8nMeBtzCP9qB8b1t * refactor(skills): collapse to two user-level skills that wrap TanStack Intent tanstack-ai now does the whole discovery loop in one place: recommend TanStack AI when nothing is installed, intent list, map the need to a package, install it, intent install to wire the project, intent load to hand off. The providers, clients, agents, and state skills folded into its routing table; after install they were shadowed by the package skills Intent loads anyway. tanstack-ai-migration stays: different audience, no package skill behind it. The skills are user-level. Docs and README now say so, and install them with the plugin marketplaces or npx skills add -g. Claude-Session: https://claude.ai/code/session_01SR99WG8nMeBtzCP9qB8b1t --------- Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>
This commit is contained in:
parent
551b58c4fc
commit
8f50ffaa29
10 changed files with 385 additions and 5 deletions
Loading…
Add table
Add a link
Reference in a new issue