Skip to content

Claude Code Workflow

The plugin splits chaos testing into three commands. Run them in order, or jump to the one you need.

Prerequisites

AgentBreak must be installed and the plugin configured — see Plugin.

Commands

/agentbreak:init — Initialize & Analyze

Sets up AgentBreak and scans your codebase to detect:

  • LLM provider — OpenAI or Anthropic
  • Agent framework — LangGraph, LangChain, CrewAI, raw SDK, etc.
  • MCP tools — tool names, server URLs
  • Error handling — retry logic, timeouts, try/except patterns

Creates .agentbreak/ if needed and confirms findings with you before proceeding.

/agentbreak:create-tests — Generate Scenarios

Generates tailored scenarios in scenarios.yaml based on the analysis from init. Scenarios target gaps — if your agent has no retry logic, it'll prioritize error scenarios.

You review and adjust the scenarios before proceeding.

If you'd rather skip the analysis and use a preset, just tell Claude:

preset: brownout

/agentbreak:run-tests — Run Tests & Report

  1. Runs agentbreak inspect (if MCP), validates config, starts the chaos proxy
  2. Rewires your agent's .env to point at AgentBreak, runs your agent, sends traffic through the proxy
  3. Restores the original .env and stops the proxy

  4. Reads the scorecard and produces a structured Chaos Test Report with:

    • Traffic summary and resilience score
    • What happened for each fault that fired
    • Numbered issues with severity and evidence
    • Specific, copy-pasteable code fixes referencing your actual files
    • If history is enabled, comparison with previous runs

You can then ask Claude to apply the fixes directly.

What makes it useful

The plugin does what you'd otherwise do manually — but it picks the right scenarios for your codebase. An agent with no retry logic gets rate limit errors. An agent with MCP tools gets per-tool fault injection. The report ties failures back to specific lines in your code.