@tracegate/adapters

Framework adapters and observability exports for TraceGate.

Adapters stay thin: @tracegate/core remains the source of truth for contract validation, policy, approval, evidence, redaction, and trace semantics.

Install

pnpm add @tracegate/adapters

Framework Adapters

import {
  createTraceGateClientFunctionTool,
  createTraceGateFunctionRegistry,
} from "@tracegate/adapters/plain-functions";
import { createTraceGateOpenAIAgentsTool } from "@tracegate/adapters/openai-agents";
import { createTraceGateLangGraphTool } from "@tracegate/adapters/langgraph";
import { createTraceGateVercelAITool } from "@tracegate/adapters/vercel-ai-sdk";

• Plain functions: wraps an existing in-process tool registry with createRuntimeGate(). createTraceGateClientFunctionTool() supports pre-call and post-call evidence flows for client/host-dispatched tools.
• OpenAI Agents SDK: creates a real function tool with the contract name, description, and Zod input schema.
• LangGraph JS: creates a LangChain/LangGraph-compatible structured tool for ToolNode-style workflows.
• Vercel AI SDK: creates an AI SDK tool() with TraceGate validation, summaries, and JSONL traces at the tool boundary.

OpenAI and LangGraph adapters keep harness semantics. Plain function and Vercel adapters use runtime-gate semantics for observe, shadow, and targeted enforce rollout.

Run local examples:

pnpm --filter tracegate-example-openai-agents start
pnpm --filter tracegate-example-langgraph-js start

Run local stack templates:

pnpm templates:check

Templates cover existing registries, observe mode, shadow comparison, validation-only enforcement, JSONL traces, and runtime replay without model/API credentials.

Observability Exports

import { createOpenTelemetryTraceSink } from "@tracegate/adapters/opentelemetry";
import { toBraintrustEvalRows } from "@tracegate/adapters/braintrust";
import { toLangfuseTraceEvents } from "@tracegate/adapters/langfuse";

• OpenTelemetry sink emits one span per TraceGate event with stable tracegate.* attributes.
• Braintrust mapper emits JSON-serializable eval rows from matrix/replay results.
• Langfuse mapper emits trace/event payloads and a no-op-by-default sink.

No exporter sends network data unless the caller provides a configured SDK/exporter writer.

JSONL traces remain the local source of truth for replay and CI.