Skip to content

README.md

Latest commit

 

History

History
120 lines (84 loc) · 4.2 KB

README.md

File metadata and controls

120 lines (84 loc) · 4.2 KB
Raw

linka-skillhub

One local registry for every coding-agent skill on your machine.

中文 · Quick start · Why · Safety · CLI


License: MIT Node 20+ pnpm 9 Local first



linka-skillhub Web console

Why It Exists

Coding agents are getting useful enough that their skill libraries become real infrastructure. Then the copies multiply: Codex, Claude Code, OpenCode, Mavis, Cursor, OpenClaw, Hermes, and shared .agents/skills paths all start carrying slightly different versions of the same skill.

linka-skillhub turns that sprawl into one Git-backed source of truth.

What It Does

  • Finds every local skill across agent-specific and shared directories.
  • Imports canonicals into registry/skills/<name>/, with normal Git history.
  • Reviews before sharing using deterministic rules and optional code-agent reviewers.
  • Distributes safely from the registry back to Codex, Claude, OpenCode, Mavis, Cursor, OpenClaw, Hermes, or shared paths.
  • Reconciles drift with pull, push, fork, push-all, and agent-assisted merge.
  • Works headless or visual: the Web console and CLI call the same core workflows.

Quick Start

pnpm install
pnpm mirror:local
pnpm build
pnpm serve

Open http://127.0.0.1:4873.

By default, mirror writes only to .sandbox/local-mirror, so you can test with mirrored real skills before touching your actual agent directories.

Workflow

scan -> registry -> review -> distribute -> sync drift

CLI

lsh list
lsh registry import --create
lsh review --reviewer rules
lsh distribute preview --target codex,claude --skill writing-plans
lsh distribute apply --target codex,claude --skill writing-plans --yes
lsh sync status
lsh sync pull writing-plans --from claude
lsh sync merge writing-plans --from claude,opencode --by codex
lsh serve

During local development, use pnpm lsh -- <command> after pnpm build, or pnpm lsh:dev -- <command> to run the TypeScript entrypoint.

Profiles

Profile Writes to real agent dirs? Best for
mirror No Default testing with mirrored real skills.
sandbox No Small deterministic fixtures.
local Yes Intentional writes to your real machine.

Real skill directories are opt-in: use --profile local, and pass --yes for non-interactive writes.

Safety Model

  • Preview first, apply second for distribution writes.
  • Web apply requires a live server-side confirmToken.
  • Unsafe, invalid, and agent-bound skills are excluded unless explicitly included.
  • Registry writes are path-checked inside the active profile repo.
  • registry/instances.json is ignored by Git because it contains machine-local paths.
  • Overwrites create backups under the active stateDir.

Web Console

The browser UI is intentionally thin. It presents the same operations the CLI runs:

Page Use it for
Overview Search skills, inspect status, source distribution, history, and evidence.
Intersect Copy selected skills from one agent to another.
Distribute Push registry skills to several agents with preview and confirmation.
Repo Import, review, switch registry, bind a remote, pull, push, and browse history.

Development

pnpm typecheck
pnpm test
pnpm build
pnpm test:webui:functional
pnpm verify

pnpm test:webui:functional builds the Web console, starts an isolated sandbox server, and clicks through language switching, registry loading, import, distribute, intersect, sync pull/push/push-all, merge gating, and remote push flows.

pnpm verify runs typecheck, tests, production build, UI audit, and UI flow checks.

License

MIT