
lunora
- GitHub stars
- 302
Install
/plugin install lunora@lunoraRun it inside Claude Code after adding the marketplace with /plugin marketplace add anolilab/lunora.
Contents
Discovered from sourceAgents 1
Lunora agent plugin
The first-party Lunora plugin for Claude Code and Codex. It ships two things:
| Component | What it is |
|---|---|
| Skills | The 14 first-party skills from packages/cli/skills/, symlinked in — lunora routes, the rest do the work. |
| Hook | A Stop gate that re-runs lunora verify at turn end and blocks with the compiler's own errors when it fails. |
Install
Claude Code
/plugin marketplace add anolilab/lunora
/plugin install lunora@lunora
Codex — search for lunora in the plugin directory, or add this repo as a
marketplace (.agents/plugins/marketplace.json).
Neither — lunora rules install copies the same skills into the project's
.agents/skills/, which Cursor, Copilot and Codex all read. That path gives you
the skills without the hook.
MCP is a separate install, on purpose
The plugin declares no MCP server. lunora mcp install already owns that
job, and it makes two decisions a plugin manifest cannot express: it spawns the
project's binary through the project's package manager (a plugin-level npx
would fetch whatever is newest on npm and point a mismatched CLI at this
project's .lunora/dev.json), and it deliberately refuses to write a
machine-wide lunora entry, because the stdio spec carries no cwd and one
global entry would serve whichever directory the editor happened to start in.
See the comments on McpServerPlan in packages/cli/src/commands/mcp/install.ts.
lunora mcp install # every MCP client configured on this machine
lunora mcp install --print # show what it would write
The Stop hook
scripts/verify-turn.mjs is the only new behaviour here. The skills already
say "run lunora verify before you call it done"; agents skip it, and the type
error surfaces on the user's next lunora dev instead of in the turn that
caused it. The hook makes the rule a mechanism: on turn end it runs the same
command and, on failure, blocks the stop with the output, so the next turn
starts from the real errors.
It stays quiet unless it has something to say:
- No
lunora/directory next to a wrangler config → not a Lunora project, exit. The walk up stops at the enclosing.git, so a sibling checkout namedlunoracan't capture an unrelated repo. git status --porcelain -- lunoraclean → the turn didn't touch the backend, so skip without paying for atscrun. Non-git projects always verify.stop_hook_active→ Claude Code is already continuing because this hook blocked. It also caps a hook at 8 consecutive blocks, which is why there is no counter here to get wrong.- No
lunorain the project's (or a parent's)node_modules/.bin→ an environment fact, not a type error.
A verify that times out or cannot be spawned returns a systemMessage rather
than a silent {}, so "could not check" never reads as "checked, and it
passed".
node --test plugins/lunora/scripts/verify-turn.test.mjs # also runs in postinstall
Adding a skill
Skills are not authored here. packages/cli/skills/<name>/ is the source of
truth; this directory holds symlinks, as do .agents/skills/ and
.claude/skills/. See packages/cli/skills/README.md for the hops — or just
run pnpm install, and scripts/check-skill-mirrors.js prints the exact
ln -s command for whichever mirror is missing.
Claude Code dereferences symlinks that resolve inside the marketplace when it copies a plugin into its cache, so installed users get file contents, not links.
Not included, and why
- A reviewer skill.
lunora advisoralready runs the static and runtime lint set — RLS coverage, ownership checks, validators, index usage, secrets — and scores them per procedure. A markdown checklist restating a working linter is a checklist that drifts from it. Thelunorarouter skill points at the command instead. - A runtime-error monitor. Plugin monitors need something to tail, and
lunora devwrites no error log. Add one when it does.
Hooks 1
Sponsors
Details
- Marketplace
- anolilab/lunora
- Source
- plugins/lunora
- Category
- Engineering