Sphinx: quiz yourself before shipping agent code
Added NewCopyable prompt below
Published Last verified Released
Our verdict
Worth trying if agent-written diffs are shipping faster than you can absorb them: on a push to a protected branch, Git waits while a separate reviewer model reads the commit and asks you to predict failure modes, not file names. Pass and the same push continues; miss and you get an explanation plus a new scenario. It is practice, not a security gate: anyone local can skip or bypass hooks, the examiner is an LLM and can be wrong, and the reviewer provider receives your changed code. Global install sets core.hooksPath for every repo on the machine, so run sphinx init per project and edit .sphinxignore before the first check.
- Who it's for
- Developers using Claude Code, Codex, Cursor or plain git who want a forced understanding check on main before agent code lands
- Leverage
- A push pauses until you can explain what the change does under failure
- Effort
- Medium
- Date added
What it does
Sphinx (parthh01/sphinx, npm package sphinx-understand, MIT) installs a user-level Git hooks path that chains with existing hooks. A reviewer model (Claude Code or Codex plan, OpenRouter, or any OpenAI-compatible endpoint) reads the committed change through read-only Git tools and never sees the chat that produced the code. Checks open in a local browser questionnaire or in the terminal; Claude Code can get native questionnaire hooks plus an MCP server. Needs Node.js 22.19+, Git, and macOS, Linux or WSL (native Windows not supported yet). A typical check makes several model calls; set provider spending limits.
How builders use it
- npm install -g sphinx-understand, then sphinx install (or sphinx install --yes --preset claude-economy --agents claude).
- In each repo: sphinx init, edit .sphinxignore for secrets, fill .sphinxcontext.md, then sphinx doctor.
- Push to main or master and answer the questionnaire; use sphinx yolo only for an explicit hotfix skip.
- For Claude Code, confirm /mcp shows sphinx after install; for Codex, trust the hooks under /hooks.
What to expect
Protected pushes only continue after you pass an open-book quiz on the behaviour of the change.
Copy this: Install + first protected push
# Official source: github.com/parthh01/sphinx (site: sphinx-six-theta.vercel.app)
npm install -g sphinx-understand
sphinx install --yes --preset claude-economy --agents claude
# Codex: --preset codex-economy --agents codex
# No coding agent: --agents none
cd your-repo
sphinx init --yes --branches main
# Edit .sphinxignore now if the repo has private or generated paths
sphinx doctor
git push origin main # Git waits on the local questionnaire URL until you pass, pause, or skipRelated discoveries
See the newest picks →- SetupMulti-Step Agent Setups
Deterministic guardrails with Claude Code hooks
Run your own scripts at agent lifecycle events: format, lint, block risky commands.
Effort: Medium
- WorkflowCoding & Shipping Agents
Auto lint-and-test fix loop with Aider
Make the agent run your linter and tests after each edit and fix what breaks.
Effort: LowPrompt
- ToolSmall Tools, Repos & Plugins
agent-device: verify mobile apps from your coding agent
CLI, MCP server and Node API that let coding agents inspect, control and verify iOS, Android and other apps with accessibility snapshots and evidence.
NewEffort: MediumPrompt#Plugin / extension#MCP
- ToolSmall Tools, Repos & Plugins
ccusage: see what your coding agents cost
A zero-install CLI that reads local logs from Claude Code, Codex, OpenCode and more and reports tokens and costs.
NewEffort: LowPrompt#Local / BYO model#Open source
Want more like this?
Practical discoveries, without the noise.