Приёмочное тестирование MCP-инструментов живым агентом

★ 7.4 · testing

Разовые задачи в процессе разработки is a Claude Code skill that runs bot acceptance tests (BAT) to validate MCP tools from a real AI agent's perspective, not from a mock or unit-test harness. Reach for it when you need to confirm a PR hasn't broken tools before merging, when you want to catch a regression by comparing branch behavior against master, or when a tool has changed and you need an end-to-end check driven by a live Claude or Gemini CLI. The skill dynamically builds a JSON scenario with setup, test, and teardown prompts, pipes it to `tests/uat/run_uat.py`, and returns a concise summary with an `all_passed` flag per agent, tool-call success rates, turn counts, and timings. On failure it can re-run with `--branch master` to produce a side-by-side regression comparison. Built on Bash, Read, and Write tools; requires a configured environment with Claude and/or Gemini CLI access and the Python `run_uat.py` runner. The direction is agent-driven integration testing of MCP tools, not static code analysis or unit tests. Suitable for teams building or maintaining MCP servers; each scenario invocation consumes API credits per agent, so keep prompts focused.