agent-dev-kit

Builds one TypeScript AI agent or RAG increment from a spec-dev-kit spec (and optional html-generator-kit prototype) — scoped UPSTREAM_SPEC import of agent-surface, hub-and-spoke (scaffold-agent or embed-agent, tools, @openai/agents or LangGraph.js, eval goldens), quality gates, and a mandatory human review. Writes kit-result.json. Never opens a PR. Use when implementing a named agent from an approved spec, not for choosing architecture alone (that is agent-architect) and not for a greenfield tree alone (that is scaffold-agent).

Set up Vitest evaluation for TypeScript agents and RAG — golden JSON datasets, Zod LLM-as-judge, trajectory tool-correctness and max-turns checks. Use before shipping or when changing prompts, tools, or chunking. Not for implementing Agent or StateGraph.

Embed an @openai/agents chat route into an existing Express HTTP app. Merge optional OPENROUTER_API_KEY (503 if unset), Zod request body, OpenAPI path, auth, inject host services through RunContext, mount the router in compose.ts. Use when compose.ts already exists. Do not overwrite host env, tracing, or index.ts.

Build a production LangGraph.js agent — StateGraph, ToolNode, MemorySaver or Postgres checkpointing, interruptBefore human-in-the-loop, Plan-Execute. LLM is OpenRouter ChatOpenAI. Use when custom graph topology or durable resume is required, not for a simple @openai/agents Runner loop.

Build a production agent with @openai/agents in TypeScript — Agent, Runner.run, tool(), handoffs, guardrails, sessions, and structured outputType. Wire the OpenAI SDK to OpenRouter Chat Completions. Use for tool-calling loops that do not need LangGraph.js checkpointing.

Build a TypeScript RAG pipeline with OpenRouter embeddings, Qdrant hybrid dense+BM25 search, reranking, parent-child chunking, and citation generation. Use for knowledge-base Q&A, not for Agent/Runner loops or Vitest judges.

Scaffold a Node.js TypeScript agent project layout — package.json, src/llm OpenRouter client factory, env parsing, agents/tools/rag folder tree, and .env.example. Use when creating the initial directory structure for a new AI agent service, not when implementing Runner or StateGraph loops.

Design and implement agent tools with Zod input/output schemas, structured error fields, a read/write split, and a name registry for @openai/agents tool() or LangGraph.js structured tools. Use when adding or auditing a tool function, not when wiring Runner or StateGraph.