@peculis/provenance-langgraph

LangGraph.js / LangChain capture adapter for @peculis/provenance-sdk. Pass one callback handler into graph.invoke and model_call / tool_call events are recorded automatically — declared governance calls (gate / accessed / humanApproval) inside your nodes resolve the active session via async context.

Early/WIP (0.x).

Install

pnpm add @peculis/provenance-sdk @peculis/provenance-langgraph @langchain/core

@peculis/provenance-sdk and @langchain/core are peer dependencies.

Use

import { init } from "@peculis/provenance-sdk";
import { withSession, currentSession } from "@peculis/provenance-langgraph";

const provenance = init({
  projectKey: process.env.PROVENANCE_PROJECT_KEY!,
  endpoint: "https://ingest.your-provenance.example",
});

await withSession(provenance, { agentId: "refund-agent" }, async ({ handler }) => {
  // pass `handler` as a callback — model_call / tool_call captured automatically
  await graph.invoke({ input }, { callbacks: [handler] });
});

// declared calls inside your tools/nodes resolve the active session:
const lookup = tool(async ({ id }) => {
  currentSession().accessed("crm.customers", "lookup", { recordCount: 1 });
  return fetchCustomer(id);
}, { name: "lookup_customer", schema: z.object({ id: z.number() }) });

withSession starts/ends the session (one graph invocation = one session), records an error_abort on throw, and binds the writer so declared calls and delegated sub-graphs attach to the right session. A delegated sub-graph calls withSession again with metadata.parent_session_id to link the child.

See the quickstart.