Skip to content

v2.0.7

Latest

Choose a tag to compare

@CloudAI-X CloudAI-X released this 02 Oct 05:54
3b5a89e

Security

  • protect-files.py now protects .git/ and warns on .github/workflows/ for the absolute paths Claude Code actually sends (the patterns only matched relative paths before), and blocks every .env.* file except the .env.example, .env.template and .env.sample templates.
  • security-check.py no longer skips secret scanning for any path that merely contains "test" or "spec" (for example a project folder named latest-app). Test and fixture locations are now matched on whole path segments and file-name conventions. Added the current OpenAI (sk-proj-…) and Anthropic key shapes.
  • format-on-edit.py and typescript-check.py no longer call npx. They run the project's own node_modules/.bin tool (or one on PATH) and skip silently otherwise. Outside a TTY, npx tsc downloaded and ran an unrelated npm package named tsc.
  • security-patterns skill: corrected examples that taught insecure code (mass assignment from req.body, bcrypt's 72-byte limit, OAuth without PKCE/state, login timing that reveals which emails exist, a zod trim() applied after the length check, unpinned JWT algorithm).
  • security-scan command: the secret-detection greps used basic regular expressions and matched nothing; they now use -E, scan .env, JSON, YAML and JSX/TSX files, and exclude .git and node_modules by directory.

Fixed

  • Hook warnings were invisible. pre-commit-check.py, branch-protection.sh, typescript-check.py, the warning path of protect-files.py, verify-on-complete.py and suggest-doc-updates.py printed plain text and exited 0, which Claude Code only writes to its debug log. They now return JSON: hookSpecificOutput.additionalContext to inform Claude, systemMessage to inform the user.
  • Six commands aborted before reaching the model in the default permission mode (add-tests, commit-push-pr, quick-fix, refactor-guided, summarize-changes, sync-branch), and commit aborted in a repository with no commits, because their !`…` context commands failed Claude Code's permission pre-check or exited non-zero. The context commands are now simple read-only forms.
  • npm installer did not activate hooks. npx install-claude-workflow-v2 copied the hook scripts to .claude/hooks/ but never registered them. It now merges them into .claude/settings.json (additive, idempotent, existing settings untouched).
  • argument-hint values are quoted strings in all ten commands that used a bare [...] (a YAML list, and a parse error in three files), so strict loaders such as GitHub Copilot CLI load them again (#24).
  • hooks/hooks.json quotes ${CLAUDE_PLUGIN_ROOT} so a plugin path containing spaces works.
  • suggest-doc-updates.py reported "New directory created" for every new file in an existing directory, counted the plugin's own log files as project changes, and would have repeated the same suggestions every turn; it now reports only genuinely new top-level directories, once per session, and ignores .claude/.
  • validate-prompt.py: the security hint never matched "vulnerability".
  • verify-on-complete.py detects Bun's text lockfile (bun.lock).
  • track-metrics.py crashed on Python 3.8 (type hints evaluated at import); branch-protection.sh and log-commands.sh no longer print a jq parse error on malformed input.
  • templates/mcp.json.template pointed at eight npm packages that do not exist; it now uses the vendors' remote MCP endpoints and the maintained @modelcontextprotocol packages. templates/settings.local.json.template no longer installs a hook that relied on a non-existent $CLAUDE_FILE_PATH variable.
  • Skills: corrected examples that did not work as written in devops-infrastructure (GHCR push without login or permission, Postgres containers without a password, dev dependencies left in the image, uv sync before the source is copied, end-of-life Node 20), database-design (zero-downtime migration lost writes, keyset pagination without a tie-breaker, trigger never attached), error-handling (pino calls that dropped their fields, a retry example that never retried), optimizing-performance, designing-tests, designing-apis, managing-git, web-design-guidelines and vercel-react-best-practices (two SWR imports that do not exist).
  • Documentation: plugin install commands (claude plugin marketplace add + project-starter@claude-workflow), hook table triggers and behaviour, /install-github-app, claude plugin validate usage, the contributor hook-registration example, changelog years for 1.0.0 and 1.1.0, the privacy policy's description of the local command and metrics logs, and the security policy, which now points to private vulnerability reporting.

Changed

  • orchestrator agent: delegates when auto-delegated as a subagent (current Claude Code allows nested subagents), and task tracking no longer depends on TodoWrite.
  • parallel-execution skill and the orchestration examples describe the current Agent tool (Task remains an alias), automatic delivery of background subagent results, and running dependent work in a second wave.
  • commit-push-pr refuses to commit or push on the default branch without creating a feature branch or asking first. dependency-upgrade and refactor-guided stop on a dirty working tree instead of stashing or committing the user's changes.
  • The four "output styles" are described as mode commands, which is what they are.
  • CI: also validates hook script references in hooks/hooks.json, that argument-hint values are strings, and that the version agrees across all manifests and the changelog. actions/checkout updated to v7.0.1.

npm installer: install-claude-workflow-v2@2.0.7 is published and carries the hook-registration fix (npx install-claude-workflow-v2@latest).