interrogate

Point the skill at specific files, a diff, or a feature branch and it runs git diff against the base branch. It spawns one reviewer per configured model with the same prompt and rubric, then synthesizes their findings into a single verdict.

Reviewers apply a rubric covering correctness, root causes versus symptoms, structural integrity, verification, complexity budget, and security, plus a code-quality lens targeting structural simplification and spaghetti growth. Each finding includes severity, evidence, and an optional suggestion.

Findings raised by two or more models independently are treated as highest-confidence signal, while lone-model findings are weighted lower. Disagreements between models are noted explicitly in the agreement map section of the verdict.

No, the skill explicitly does not auto-apply changes. It delivers a synthesized verdict categorizing findings as act on, consider, noted, or dismissed, leaving all fix decisions to the user.

The skill checks the valid slugs in the Task tool error message, picks the closest equivalent preferring the highest-reasoning tier of the same family, and continues the review. It does not block on the slug issue and suggests updating the configured value separately.