Vibe-Coded Repo Cleanup With Claude Code or Codex

Vibe coding rewards local progress. A generated feature compiles, the demo works, and nobody asks whether the new helper duplicated an older helper, skipped the migration path, or hid a state bug behind a happy-path browser session.

• Parallel abstractions appear because each prompt sees a partial repo.
• Tests are added around generated behavior instead of business behavior.
• Old files stay alive because nobody knows which imports are safe to remove.
• Env vars, background jobs, and auth callbacks drift away from docs.
• Agents keep patching symptoms because the repo has no shared operating rules.

Give the agent a narrow inspection prompt before allowing edits. The first output should be a map, not a patch.

Inspect this repo without editing files.

Return:
1. Entry points and user-visible workflows
2. Build, lint, typecheck, test, and e2e commands
3. Duplicated concepts or modules
4. Dead files or unreachable paths
5. State, auth, billing, and data migration risks
6. Three smallest cleanup PRs with rollback plans

That prompt keeps Claude Code or Codex in investigator mode. Once the map is useful, convert each cleanup into a bounded task with a test requirement and a rollback note.

## Cleanup claim
This PR removes or simplifies: ...

## Proof
- [ ] Existing behavior covered by tests
- [ ] New regression test added when behavior was unclear
- [ ] Lint/typecheck/build run
- [ ] Manual smoke path checked

## Risk
- Files touched:
- State or data migration:
- Rollback:

## Agent notes
- What was intentionally left alone:
- What should be cleaned next:

This is where an always-on workspace helps. Long cleanup passes need repeated test runs, dependency installs, browser smoke tests, and review loops. If the machine sleeps or loses terminal state, the agent loses the thread.

Hyperbox is useful when cleanup becomes an operating loop: an agent runs tests, waits for review, updates docs, retries failed checks, and keeps repo state warm across hours or days. The Mac is not magic. It is the stable place where the work keeps living.

• Keep dependency caches, browser sessions, local services, and repo state warm.
• Run Claude Code, Codex, Cursor, and smoke-test browsers from one persistent workspace.
• Use SSH for command work and VNC when cleanup requires visual inspection.
• Preserve logs and terminal history after your laptop closes.

Should I rewrite a vibe-coded repo from scratch?

Usually no. First freeze behavior, map dead paths, add tests around the paths that matter, then refactor in small PRs. Rewrite only when the dependency graph, data model, or security model cannot be recovered safely.

Can Claude Code or Codex clean up AI-generated code?

Yes, if you give the agent a bounded runbook, tests, repo context, and a persistent workspace. Treat the agent as a refactoring operator, not as a one-shot prompt box.

Why use an always-on Mac for repo cleanup?

Cleanup runs often span hours of tests, migrations, browser checks, and review passes. An always-on Mac keeps the repo, caches, credentials, logs, and agent sessions available after your laptop sleeps.