How Claude Code Debugs and Repairs Code: the Verification Stack

The hard part of a fix isn't writing it — it's proving it

When a build breaks, an AI coding agent like Claude Code can propose a repair in seconds. The real work is finding the actual cause and verifying that the change fixed it without breaking something else. These are not AI tools — they are the standard engineering toolkit that makes an AI-proposed fix trustworthy. Point Claude Code at them and you get a repair loop you can actually rely on.

The recovery flow:

broken build → dependency graph → root cause → AST repair → affected rebuild → tests → security scan → verified patch

Trace the damage

• dependency-cruiser — map dependencies and detect architectural violations.
• Madge — expose circular dependencies and module relationships.
• Knip — find unused files, exports, and dependencies.
• ts-morph — inspect and modify TypeScript code structurally.

Find the root cause

• TypeScript — catch unresolved imports and type errors.
• ESLint — detect code problems through configurable rules.
• Semgrep — identify risky code patterns with static analysis.
• ast-grep — search and rewrite code using AST patterns.

Repair the codebase

• jscodeshift — automate large code transformations.
• Comby — rewrite code using structural matching.
• Biome — lint and format code consistently.
• Renovate — automate dependency updates.

Rebuild only what changed

• Nx — identify affected projects and run their tasks.
• Turborepo — coordinate builds with task caching.
• Dagger — create reproducible build and test pipelines.
• Bazel — build and test large projects with precise dependency tracking.

Prove the fix works

• Vitest — run automated tests and catch regressions.
• Playwright — verify real browser behavior after changes.
• StrykerJS — mutation-test the tests themselves.
• CodeQL — detect security vulnerabilities through code analysis.

Three repair stacks worth stealing

• Import failures: dependency-cruiser → Madge → ts-morph → TypeScript
• Large refactors: Semgrep → ast-grep → jscodeshift → Nx → Vitest
• Production recovery: Renovate → Dagger → Playwright → CodeQL

An honest word

Claude Code can investigate a failure and propose a repair, but these tools are how you confirm what actually changed — the dependency graph shows the blast radius, the tests and mutation tests prove behavior, and the security scan catches what a fast fix might introduce. Don't merge an AI-generated patch on trust alone; let the graph, the tests, and the scan back it up. That is the difference between "it compiles" and "it's fixed."

Pairing these with ready-made agents and skills? See our hand-picked Claude Code agents, skills & plugins.