Reka — Claude Code Plugin

RAG-powered AI development for any codebase
Semantic search · Project memory · Architecture decisions · Coding workflows

---

Overview

Reka is a Claude Code plugin that connects your AI assistant to a shared RAG (Retrieval-Augmented Generation) infrastructure. It gives Claude persistent project memory, semantic codebase search, architecture awareness, and structured development workflows — across sessions and team members.

What you get

Capability	How it works
Semantic search	Search code, docs, and Confluence by meaning, not just keywords
Project memory	Decisions, patterns, and insights persist across sessions via Ebbinghaus-inspired retention
Architecture awareness	ADRs, dependency graphs, blast radius analysis before every change
Structured workflows	5-phase coding, deep investigation, tribunal debates, and more
Auto session lifecycle	Hooks start/end RAG sessions automatically, trigger memory consolidation

---

Quick Start

1. Add marketplace

/plugin marketplace add getreka/reka-plugin

2. Install plugin

/plugin install reka@reka-plugins

3. Configure

On first enable, Claude Code prompts for three settings:

Setting	Description	Example
RAG API URL	Your RAG API server	http://localhost:3100
RAG API Key	Project-scoped key (rk_{project}_…)	rk_myapp_…
Project Name	Used to mint the key; the namespace is resolved from the key	myapp

Multiple projects on one machine? The single bundled config above is for one project. For several repos, run npx @getreka/cli init (or /reka:onboard) once per project — each writes its own gitignored .mcp.json with its own rk_{project}_… key and resolves its own namespace from that key, independently.

Team auto-install

Add to your project's .claude/settings.json so everyone gets it automatically:

{
  "extraKnownMarketplaces": {
    "reka-plugins": {
      "source": { "source": "github", "repo": "getreka/reka-plugin" }
    }
  },
  "enabledPlugins": {
    "reka@reka-plugins": true
  }
}

---

Commands

Coding & Review

Command	Description
/reka:code	5-phase workflow: context → plan → implement → verify → remember
/reka:investigate	Deep research — find, trace, debug. Saves to memory, never modifies code
/reka:review	Architecture-aware code review against patterns and ADRs

Architecture

Command	Description
/reka:arch	Record and analyze architecture decisions (ADRs)
/reka:debate	Adversarial tribunal debate for complex decisions (2-4 positions)

Session & Memory

Command	Description
/reka:start	Display session status and project stats (sessions auto-start via hook)
/reka:end	Save knowledge, close session, trigger memory consolidation
/reka:memory-review	Triage quarantine queue, promote/reject auto-extracted memories

Setup

Command	Description
/reka:onboard	Set up RAG for a new project: configure, index, verify

---

Agents

Internal subagents used by the workflows — /reka:code, /reka:investigate, and /reka:review delegate to them via the Task tool. For direct requests, prefer the corresponding /reka:* command.

Agent	Model	Used by
reka:feature-builder	Sonnet	/reka:code Phase 3 — multi-file implementations with RAG context
reka:code-reviewer	Sonnet	/reka:review — per-file review against patterns and ADRs
reka:rag-researcher	Haiku	/reka:investigate — parallel semantic search and graph traversal
reka:rag-ops	Haiku	Operations: indexing, collections, memory maintenance

All agents have persistent memory (memory: project) — they learn your codebase patterns across sessions.

---

Hooks

Event	Action
SessionStart	Auto-starts RAG session, injects RAG_SESSION_ID env var, injects the session digest into context
SessionEnd	Ships the session transcript to your RAG API, ends the RAG session, triggers consolidation agent

Transcript capture sends the last 2 MiB of the Claude Code transcript to POST /api/capture/transcript on the RAG API URL you configured — your own server is the only destination; the plugin sends data nowhere else. Set REKA_TRANSCRIPT_CAPTURE=0 to disable it.

---

Skills

Reference skills loaded automatically by commands and agents:

• memory-protocol — Session lifecycle, smart remember with relationship detection, memory type selection, structured facts
• rag-workflows — Search tool priority guide (Grep → find_symbol → hybrid_search → search_graph → context_briefing)
• obsidian-sync — Bidirectional sync between RAG memories and Obsidian vault

---

Architecture

Claude Code
  │
  └── Reka Plugin
        ├── 9 Commands (/reka:code, /reka:investigate, ...)
        ├── 4 Agents (feature-builder, code-reviewer, ...)
        ├── 3 Skills (memory-protocol, rag-workflows, obsidian-sync)
        ├── 2 Hooks (session lifecycle)
        │
        └── MCP Server (@getreka/mcp)
              │  29 tools: search, memory, architecture, sessions, agents
              │
              └── RAG API
                    ├── Qdrant — vector database
                    ├── Ollama — embeddings (qwen3-embedding) + local LLM
                    └── Claude — complex tasks (optional, hybrid routing)

---

Prerequisites

• RAG API server running (getreka/reka)
• Qdrant vector database
• Ollama — embeddings + local LLM (default provider)

Optional but recommended:

• typescript-lsp@claude-plugins-official for TypeScript code intelligence
• Obsidian for memory visualization (via /reka:obsidian-sync skill)

---

Local Development

claude --plugin-dir ./reka-plugin

Use /reload-plugins after making changes to pick up updates without restarting.

Test the hooks (plain bash, mocks curl via a PATH shim):

bash tests/session-start.test.sh
bash tests/session-end.test.sh

---

License

BSL-1.1