agent-chaos · packages

Framework adapters for injecting fault-tolerance testing into LangChain, LlamaIndex, Vercel AI SDK, or any custom agent with a tool-call interface, using a shared interceptor pattern that wraps tools without modifying agent code. Each adapter exports a framework-specific wrapper (e.g., a tool wrapper for LangChain, a tool spec wrapper for LlamaIndex) that pairs with the `@reaatech/agent-chaos-core` engine at runtime.

A CLI that validates, generates, and runs chaos engineering scenarios for AI agents, providing `init`, `generate`, `validate`, and `run` commands for managing fault-injection test scenarios defined in YAML or JSON.

Middleware-based fault injection engine for agent systems that intercepts tool calls and injects latency, timeout, malformed output, token exhaustion, and other failure modes via a single `engine.intercept()` call, with scenario-driven YAML/JSON configuration, probability-based fault selection, and seeded randomness for reproducible testing.

End-to-end test suite that validates the full agent-chaos pipeline—scenario loading, schema validation, fault injection, engine event recording, CLI execution, and cross-package integration—against mock agent implementations.

Structured logging, metrics collection, OpenTelemetry tracing, and report generation for chaos engineering experiments, exposed as a set of classes (`MetricsCollector`, `Tracer`, `ReportGenerator`) that consume events from `@reaatech/agent-chaos-core`.

A scenario loader and validator for the agent-chaos fault injection toolkit that parses YAML and JSON scenario files, validates them against a JSON Schema, supports composition via `extends` with circular reference detection, and provides hot-reloading via file watching. Exports `ScenarioLoader` and `SchemaValidator` classes, plus a `createScenarioLoader` factory function; depends on `yaml`, `ajv`, and `ajv-formats` at runtime.