reasoning

Getting Started · What You Get · How It Works · Verification · Layout · License

This monorepo unifies four projects:
Cortex — persistent memory with biological consolidation
automatised-pipeline — Rust codebase-intelligence graph
zetetic-team-subagents — 97 reasoning patterns + 19 team agents
prd-spec-generator — stateless PRD reducer with multi-judge verification

---

Claude Code is powerful in one session and amnesiac the next. It can reason about a function but not the call graph it sits in. It can draft a PRD but not measure whether the PRD is actionable. Each of these problems has a project; each project has its own install, its own update path, its own MCP server, its own bug-report surface.

agentic-ai is the four projects merged into one TypeScript monorepo with a single Claude Code marketplace install. One pnpm command builds everything. One /plugin install enables any of the four capabilities. The MCP servers are wired against the unified TS/Rust outputs, not the original separate repos. The Cortex retrieval pipeline runs end-to-end in TypeScript and exceeds the Python baseline on the LoCoMo benchmark (MRR 0.851 vs 0.696, hit-rate 98.5% vs 95.9%).

4 plugins. 87+ MCP tools across them. 3870 tests. Real-subprocess parity verification against every source repo. pnpm audit --prod clean.

---

Getting Started

Install all four plugins from inside Claude Code:

/plugin marketplace add cdeust/agentic-ai
/plugin install memory@agentic-ai
/plugin install codebase@agentic-ai
/plugin install reasoning@agentic-ai
/plugin install prd@agentic-ai

Restart Claude Code. /mcp should show four servers connected:

MCP server name	Plugin	What it provides
memory	memory@agentic-ai	persistent memory across sessions (45+ tools)
codebase	codebase@agentic-ai	codebase graph + semantic search (23 tools)
reasoning	reasoning@agentic-ai	97 reasoning patterns + 19 specialist agents (2 tools + 63 skills)
prd	prd@agentic-ai	9-file PRD pipeline with multi-judge verification (17 tools)

Install only the plugins you want — they're independent. No monorepo checkout, no extra build, no pnpm install on the user's side.

The four MCP server names are deliberately chosen to NOT collide with the standalone source repos' server names (cortex, ai-architect, prd-gen, reasoning). If you have any of cortex@cortex-plugins, automatised-pipeline@automatised-pipeline-marketplace, or prd-spec-generator@prd-spec-generator-marketplace installed, both can coexist — Claude Code routes tool calls to the right server because the names differ.

What each plugin does on its first launch

Plugin	First-launch path	Subsequent launches
memory	runs npm install --omit=dev once to fetch native bindings (better-sqlite3, onnxruntime-node, @xenova/transformers, pg, sqlite-vec)	exec node dist/index.js immediately
codebase	downloads the prebuilt automatised-pipeline-<os>-<arch> binary from the latest GitHub Release (codebase-v* tag), caches it under bin/	exec the cached binary immediately
reasoning	exec node dist/index.js immediately — no native deps	same
prd	runs npm install --omit=dev once to fetch ajv	exec node dist/index.js immediately