Testing Distributed Systems with AI Coding Agents

Two complementary AI coding-agent skills have been released for testing distributed and stateful systems. One designs a claim-driven test plan; the other executes it and produces a findings report. Both ship as plain SKILL.md files usable with any agent that reads Markdown and runs shell commands, including Claude Code, Codex, Copilot CLI, Cursor, and Gemini. A reviewer can decide whether to ship by reading just the plan and findings artifacts, without re-running anything.

The approach responds to the observation that typical integration testing misses most of the failure modes that actually break distributed systems in production—partial network partitions, non-deterministic concurrency, crash-recovery, and idempotency under replay. Instead, every scenario is named after the single claim it tries to falsify under a specific fault. For consistency-critical claims such as safety, durability, isolation, or ordering, each scenario must bind an abstract model (register, queue, log, lock, lease, ledger, etc.) to an operation-history schema, a named checker (linearizability, serializability, exactly-once, etc.), and observable evidence that the fault actually landed.

To eliminate silent passes, every PASS must cite evidence that the fault was actually triggered, verdicts are reported through a 10-state classification, and every FAIL is tagged with blame assigned to the SUT, harness, checker, or environment. This lets engineers distinguish 'the chaos script ran' from 'the claim survived the fault' and routes reproducers to the right team.

Installation is a single instruction that any coding agent can follow: it clones the repository and wires the skills in via symlinks, with updates applied idempotently through git pull.