ContextOS Vault is a set of files for Cursor and Claude Code for six situations developers meet when an AI assistant works on real code. Some files are hooks that block or ask before an action. The rest are rules, skills and checklists that steer the assistant.

Find your situation

- It ignores my rules: short rules that load only for matching files, plus an auditor that finds vague, duplicate and conflicting rules.

- It runs a destructive command: a hook that denies catastrophic commands and asks before git reset --hard, force push, recursive delete, destructive SQL and cloud delete.

- It writes a secret into a file: a hook that blocks secret-shaped content and real .env files, plus permission rules that block reading them.

- It adds a dependency I did not ask for: a hook that asks before installs and manifest edits and denies hand-edits of lockfiles.

- It says "all tests pass" and did not run them: a skill that finds the commands your repo really defines, runs them and reports what ran and what did not.

- It changed 15 files and I asked for one: a skill that measures the change against a file and line budget and asks to split it.

What you get (33 files in one zip)

- 3 hook scripts for Claude Code, and the same 3 wired for Cursor hooks. You can read all three in minutes (338 lines in total). Standard library only, no packages to install.

- 4 Claude Code skills: verify-before-done, review-diff, explore-before-edit, emergency-stop

- Path-scoped rules for TypeScript/JavaScript, Python and tests, for Claude Code and Cursor

- 4 playbooks: destructive-action tiers, prompt-injection rules, long-session handoff, project setup

- A pre-commit checklist

- install.py: copies everything into your project and never overwrites your files

- The ContextOS Audit Kit rules linter

- 35 self-tests you can run yourself

Requirements

Python 3 and git. The hooks run as python3 (change to python on Windows).

What was tested, and what was not

- 35 automated tests pass on Python 3.11, 3.12 and 3.13 on Linux. They feed each hook simulated payloads shaped like the documented hook input, and run the other tools in a throw-away git repository.

- The hooks configuration, the hook output format, the skill format, the permission syntax and Cursor's hooks.json were checked against the official Claude Code and Cursor documentation on 2026-10-20. The Claude Code docs confirm that Write and Edit hooks receive the file path but do not document their other fields, so the secrets hook scans every text field.

- NOT tested inside the real Claude Code or Cursor applications. Real payloads can differ from the simulated ones, and the tools change their formats over time. The README includes a smoke test to run on your machine.

- Not tested on Windows, macOS, very large repositories or monorepos.

- No model benchmark. Nothing here guarantees that a model will follow a rule.

Limits

- Hooks read command and file text. They are a safety net, not a security boundary. For anything that must hold, also use least-privilege credentials, a sandbox, branch protection, backups and review.

- Secret detection finds common shapes. It can miss a secret and can flag a harmless string.

- Rules and CLAUDE.md guide the assistant; they are context, not enforcement.

Updates: new versions are listed in CHANGELOG.md inside the zip. Version 1.0.0, 2026-10-20.

License: use in your own personal and commercial projects. Do not resell or redistribute the files. Not affiliated with Anthropic or Cursor.

30-day refund

You can request a refund within 30 days of purchase if the product is broken, not as described, or was never delivered. Open the order in your Library, choose "Request refund" and describe the problem. A Getly admin reviews every request. A change of mind after downloading may be declined.