arize-instrumentation

Adds Arize AX tracing to an LLM application for the first time using a two-phase, agent-assisted flow. The goal is to set up LLM observability by integrating OpenTelemetry and OpenInference instrumentation without disturbing the application's behavior. It supports Python, TypeScript/JavaScript, Java, and Go, mapping each stack to the appropriate packages: openinference-instrumentation packages for Python and JS auto-instrumentation, arize-otel plus per-provider instrumentors for Go, and the OpenTelemetry SDK with manual OpenInference spans for Java.

The process is deliberately staged. A preflight step clarifies repo scope and runtime surface. Phase 1 is read-only analysis: inspect dependency manifests, scan imports, detect existing OTel or vendor tracing, and identify language, package manager, LLM providers (OpenAI, Anthropic, LiteLLM, Bedrock and others), and frameworks (LangChain, LangGraph, LlamaIndex, Vercel AI SDK, Mastra). Phase 2 implements the instrumentation only after the user confirms the analysis, installing packages first and then following the matched integration docs.

Guiding principles keep changes safe and minimal: prefer inspection over mutation, never alter business logic since tracing is purely additive, favor auto-instrumentation and add manual spans only for custom logic, and follow existing code style. Credentials are never hardcoded; generated code always references environment variables such as ARIZE_API_KEY. When a framework is detected alongside a provider, the framework-native integration is preferred, and manual CHAIN and TOOL spans are added when the framework does not emit tool spans so each tool call appears with input and output.