菜单
Commit changes with staged-diff secret scanning and branch protection guardrails. Use when user says "safe commit", "safe-commit", or "commit with checks". Do NOT use for regular commits — this is an opt-in extra layer on top of normal git workflow.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Lint, query, and maintain the project knowledge base. Use when auditing KB health, searching for compiled knowledge, or checking governance hygiene. Do not use when ingesting new research (use research-intake if available).
Ingest external sources into structured knowledge base. Use when user provides research material, URLs, papers, or says "research intake", "ingest this", "add to knowledge base". Do NOT use for querying existing knowledge (use project-wiki instead).
Create a structured handoff document for cross-session or cross-agent work tracking. Use when user says "new handoff", "create handoff", "track work item", "document this for next session", or needs to pass work to another session/agent. Do NOT use for reading or updating existing handoffs.
Guided project initialization wizard. Use when .needs-init exists in project root (detected by session-start hook). Do NOT use on already-initialized projects. Walks through 9 interactive setup steps.
Scaffold a new Claude Code skill with correct folder structure, YAML frontmatter, gotchas section, and progressive disclosure. Use when user says "create a skill", "new skill", "scaffold skill", "add a skill", or wants to build a reusable capability. Do NOT use when user wants to edit an existing skill.
Review changed code for reuse, quality, and efficiency, then fix any issues found. Use when user says "simplify", "review code", "clean up", "refactor", "reduce complexity", or after completing a significant code change. Do NOT use for general code questions, debugging, or feature implementation.
| name | safe-commit |
| description | Commit changes with staged-diff secret scanning and branch protection guardrails. Use when user says "safe commit", "safe-commit", or "commit with checks". Do NOT use for regular commits — this is an opt-in extra layer on top of normal git workflow. |
Adds two guardrails that existing hooks don't cover:
git added (e.g., .env files, hardcoded tokens)| Check | Existing hook | safe-commit |
|---|---|---|
| Secrets in Write/Edit content | pre-edit-guard.sh | -- |
| Schema validation | agents_schema_guard.sh | -- |
| Filesystem sandboxing | check_filesystem_path.sh | -- |
Secrets in git diff --cached | -- | scan-staged.sh |
| Direct commit to main/master | -- | check-branch.sh |
scripts/scan-staged.sh to check staged diff against shared secret patternsscripts/check-branch.sh to verify you're not on a protected branchThis skill demonstrates the on-demand hook pattern: a PreToolUse hook registered in .claude/settings.json that intercepts git commit Bash commands and runs both checks automatically before the commit proceeds.
Hook: scripts/pre-commit-hook.sh
PreToolUse hook with "matcher": "Bash"git commitscan-staged.sh (blocks on secrets) then check-branch.sh (warns on main/master)This is the reference implementation for skill-scoped hooks. The pattern:
You can also run the scripts directly:
bash agents/skills/safe-commit/scripts/scan-staged.sh
bash agents/skills/safe-commit/scripts/check-branch.sh
scripts/hooks/lib/secret-patterns.txt — do not duplicate patterns into a separate filesecret-patterns.txt so all hooks benefit