@lokomotif/blueprint-langgraph

Adapt a ComposedPrompt to a LangGraph StateGraph.

The blueprint produces a graph with three nodes — compose, execute, and audit — wired in series. The composed RTCSG prompt is the compose node's output; execute calls a caller-supplied LLM function; audit records the composition hash and forwards the model's response.

Install

pnpm add @lokomotif/sdk @lokomotif/blueprint-langgraph @langchain/core @langchain/langgraph

LangGraph and @langchain/core are peer dependencies — bring your own version aligned with whichever model adapter you use.

Usage

import { compose } from '@lokomotif/sdk';
import { buildStateGraph, type LokomotifGraphState } from '@lokomotif/blueprint-langgraph';

const composed = compose(modules);

const graph = buildStateGraph(composed, {
  llm: async (state) => callYourLLM(state.system_prompt, state.user_input),
});

const compiled = graph.compile();
const result = await compiled.invoke({ user_input: 'analyze this' });
console.log(result.llm_output);
console.log(result.composition_hash);

API

adaptToLangGraph(composed)              → LokomotifGraphLayout
buildStateGraph(composed, options)      → StateGraph<LokomotifGraphState>

adaptToLangGraph is the pure form — no LangGraph dependency, returns the layout description. buildStateGraph constructs the actual StateGraph and is the typical entry point.

Graph layout

START → compose → execute → audit → END

• compose — emits the composed system prompt and user input into state. Pure.
• execute — calls options.llm(state). The caller's contract: read state.system_prompt and state.user_input, return { llm_output: string }.
• audit — appends composition_hash and the module manifest to state for downstream observability.

Phase status

Ships in Phase 7 of IMPLEMENTATION_PLAN.md.