Codex Cleaner

Problem

Codex inside ChatGPT desktop can leave runtime files, logs, caches, temporary artifacts, and old daily-report threads under paths such as ~/.codex, %LOCALAPPDATA%, or ~/Library. Recursive deletion is risky because it can reach SQLite databases, sessions, credentials, skills/plugins, automations, and app package state. Manual audits often lack clear boundaries. Codex Cleaner separates read-only inspection from write operations, starting with health checks, storage audits, and dry-run cleanup plans.

How It Works

• Health checks and environment detection verify the host, install path, PowerShell environment, and Codex runtime area before cleanup.
• Storage and boundary audits inspect .codex size, large directories, Chat/Work/browser/app package boundaries, and Codex leftovers in user TEMP. New boundary areas are audited read-only and are not added to deletion candidates.
• Runtime cleanup planning defaults to dry-run and proposes whitelisted candidates, especially under .codex/.tmp; execution requires explicit user confirmation after the dry-run.
• Daily-report retention matches only AI 应用端日报 and GitHub 工具日报, reads minimal metadata, avoids message bodies, previews, and rollout JSON, and plans archive/delete actions based on 14/30-day rules while skipping pinned, running, non-report, or ambiguous threads.
• Update checks run after the workflow, verify the official GitHub Release, and report failures without invalidating the current cleanup result.

Boundaries

It is best used when ChatGPT desktop is closed and the skill directory is run by a human or external agent. Do not treat coverage maps or runtime-integrity results as deletion authorization. App package state, SQLite/state databases, auth/config files, full bin/runtimes directories, and third-party app directories should remain read-only or avoid-listed. Stop and review if paths are unexpected, scripts fail, candidates exceed the whitelist, or the target belongs to third-party apps such as CC Switch or CodeX++.