What problem does it solve? Prompt and tool changes in an AI agent can silently alter its decision-making behavior, and standard integration tests cannot catch these regressions. This Skill provides a structured workflow for writing, fixing, and promoting behavioral evaluations that verify the agent's tool choices and reasoning trajectory. ## Core Features & Use Cases - Eval Creation Guidance: Choose between evalTest (subprocess TestRig) and appEvalTest (in-process AppRig), seed realistic workspaces, and write assertions using breakpoints, tool-log audits, and mock MCP servers. - Failure Diagnosis: Follow a step-by-step investigation process using nightly CI logs, local Vitest runs, and prompt-tuning strategies grounded in instructional generality principles. - Test Promotion: Identify stable incubated tests (USUALLY_PASSES) with 100% pass rates over 7 nightly runs and promote them to ALWAYS_PASSES regression suites. - Use Case: After editing the system prompt in snippets.ts, write a behavioral eval that asserts the agent calls read_file within its first five turns, run it locally three times across Gemini models, then promote it once stable. ## Quick Start Ask the agent to create a behavioral eval that verifies the agent reads a file before editing it in a seeded NodeJS workspace.
