Skip to content

README.md

Latest commit

 

History

History
79 lines (62 loc) · 3.74 KB

README.md

File metadata and controls

79 lines (62 loc) · 3.74 KB
Raw

LoreKit plugins

Deterministic shared-memory integrations for three coding agents. All three share one engine — the lorekit hook command in @lorekit/cli — and differ only in the thin, declarative config that wires it to each host.

                    ┌───────────────────────────────┐
  host hook event   │  npx @lorekit/cli hook          │
  (JSON on stdin) ─→│    --adapter <claude|cursor|…>  │─→ inject lessons /
                    │    --event  <SessionStart|…>    │   retrospective nudge
                    │  (scope + MCP fetch + heuristic)│   (JSON on stdout)
                    └───────────────────────────────┘
                                   ↕
                       LoreKit MCP server (memory.*)
Bundle Read on start Failure / retrospective Notes
lorekit-claude SessionStart hook PostToolUseFailure + Stop hooks Full Claude Code plugin (skill + hooks + MCP). Installable via the marketplace.
lorekit-cursor rule (beforeSubmitPrompt best-effort) stop hook Cursor has no session-start / post-exec-result events; the rule carries the read path.
lorekit-codex SessionStart hook PostToolUse + Stop hooks Experimental (feature-flagged); AGENTS.md fallback included.

Why one engine and not three copies: every host invokes hooks as an external process with JSON over stdio, so the logic (scope derivation, MCP fetch, failure heuristic, retrospective) lives once in the zero-dependency CLI, and each per-host adapter only reshapes input/output to that host's contract. No build step is required — the CLI is plain ESM.

Claude Code (marketplace)

# From this repo (or point at the GitHub repo)
/plugin marketplace add mthines/lorekit
/plugin install lorekit-memory@lorekit

Set LOREKIT_MCP_URL and LOREKIT_TOKEN in your environment, or run npx @lorekit/cli install in the project to write a .mcp.json. Then npx @lorekit/cli doctor to verify.

Local mode (offline, no hosted server)

Set LOREKIT_MODE=local and lessons live in markdown files in two tiers: a per-user home tier at ~/.lorekit/ and an opt-in per-repo project tier at <repo>/.lorekit/ (commit or gitignore it to share or keep private). See the @lorekit/cli README for the two-tier merge / write-routing / control model.

To give the model memory.* tool calls against that local store (offline, Bash-restricted contexts included) instead of the hosted endpoint, point the lorekit MCP server at the CLI's own stdio server rather than mcp-remote:

{
  "mcpServers": {
    "lorekit": { "command": "npx", "args": ["-y", "@lorekit/cli", "mcp"] }
  }
}

lorekit mcp resolves the same control model, so LOREKIT_MODE=local (or a .lorekit.json) serves the .lorekit/ files with no network.

Cursor / Codex

See each bundle's README for copy-in instructions: Cursor · Codex.

Keeping the Claude skills in sync

The lorekit-claude/skills/* directories (lorekit-memory/, lorekit-setup/, lorekit-groom/) are vendored from the single sources at packages/cli/skill/*. sync-plugin-skill.mjs mirrors every skill it finds, so a newly added source skill is vendored automatically. Re-sync after editing any source:

node scripts/codegen/sync-plugin-skill.mjs        # copy source → plugin
node scripts/codegen/sync-plugin-skill.mjs --check # CI: fail if they differ