A city team is changing Claude tool schemas for a civic integration. What engineering practice should they apply?

City teams changing Claude tool schemas for a civic integration should version the Claude integration client in Git and require code review before merging tool schemas. That software-engineering practice works because tool definitions are production contracts: parameter names, types, and descriptions drive what Claude will call against municipal systems. Git history plus review creates an audit trail, enables rollback, and catches breaking schema edits before they reach residents or internal operators. Allowing direct commits to main for urgent schema fixes with retrospective approvals afterward fails by concept: bypassing pre-merge review means breaking schema edits can reach production before anyone catches them, and after-the-fact approval is not a substitute for a code-review gate. Generating SDK types from updated tool schemas in CI and committing generated files alongside local source edits fails by concept: automating type generation does not provide a human review gate for the schema contract itself, and committing generated output with manual edits can obscure exactly what changed and who approved it. Packaging tool schema JSON inside container image labels and redeploying services to distribute updates fails by concept: bundling schemas into image labels couples schema changes to container rebuilds and redeploys rather than keeping them versioned and reviewable in Git before they reach production. Exam caveat: code review should include security and privacy checks on what tools can read or write—not only style—and schema versioning should coordinate with downstream API consumers so Claude tools and municipal endpoints remain compatible. Operational check: open a pull request that changes a Claude tool schema in the city integration repo, require at least one review approval, run automated validation of the schema against the client tests, and only then merge to the branch that deploys.