Speckit Guard — Testing skill for Claude Code

speckit-guard

[](LICENSE)

**Your coding agent should not be the one grading its own work.**

speckit-guard is a [Claude Code](https://docs.claude.com/en/docs/claude-code) plugin for [GitHub Spec Kit](https://github.com/github/spec-kit). It has acceptance tests written from the spec before any code exists, locks them while `/speckit-implement` runs, and then has a reviewer with a fresh context check the implementation against the spec, criterion by criterion.

[Version française](README.fr.md)

The problem

In spec-driven development with Spec Kit, the agent that implements a feature also writes its tests and decides when the feature is done. When a test fails, the shortest path to green is often to edit the test instead of the code. The default flow has no step that checks for this, so "all tests pass" says little about whether the spec is met.

What speckit-guard adds

1. /speckit-tests: a skeleton of the planned interfaces is created first, so that the tests compile. Then an isolated subagent (test-writer) reads only the spec, writes the acceptance tests before implementation, and proves they fail at runtime. It also flags contradictions with the other features' specs. The tests are committed, and that commit becomes the reference.
2. A lock on the tests: during /speckit-implement, a PreToolUse hook prevents the agent from modifying those tests. It has to change the code, not the target. If a locked test is itself broken, /speckit-fix-test has the test-writer repair it and records the change as an amendment, without stopping the run.
3. /speckit-verify: mechanical checks first (no change to the tests since the reference commit other than amendments, full test suite, mutation testing if a tool is installed), then a spec-reviewer subagent that never saw the implementation reasoning judges each acceptance criterion as OK, PARTIAL or MISSING. Gaps ar