Your coding agent shouldn't run pytest
First post in a build-in-public series about verdict , an MCP server that gives coding agents structured, sandboxed test feedback. The problem Watch a coding agent work and you'll see it run pytest in your shell, unsandboxed, and then push 40,000 tokens of raw output through its context window to answer one question: did my change break anything? That's three problems in one command: Token waste.…
Coding agents run pytest unsandboxed, causing token waste, lack of controlled testing environment, and difficulty in determining if failures were introduced by the agent. verdict is an MCP server that replaces pytest with four tools: verify, explain_failure, history, and run_checks. Verify returns typed JSON with counts, per-failure message and location, while explain_failure provides only on-demand full tracebacks.
History assigns fingerprints to failures using normalized signatures, and run_checks performs lint and type checks with the same verdict shape. Each failure is stored in an SQLite db per project, allowing verdict to identify whether the failure was present before the change or introduced by the agent. Checks run in an ephemeral container with read-only worktree, and sandboxing posture provides a tighter setup for security.
Dogfooding (testing verdict with Claude Code) revealed bugs in impact selection and test scope, leading to improvements in the system.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
