deployshield CLAUDE.md

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

What This Is

DeployShield is a production safety guardrail for Claude Code and Gemini CLI. It blocks write/mutating CLI commands while allowing read-only ones. It works as a hook on shell commands — every command the AI tries to run goes through the validator, which default-denies any recognized CLI action that isn't on the safe-list. Unrecognized CLIs pass through untouched.

Commands

# Run tests
uv run pytest -v

# Run a single test file or class
uv run pytest tests/test_integration.py -v
uv run pytest tests/test_cloud_providers.py -v

# Lint
uv run ruff check hooks/scripts/ tests/
uv run ruff format --check hooks/scripts/ tests/

# Pre-commit (runs automatically on commit after install)
uv run pre-commit install
uv run pre-commit run --all-files

# Manual smoke test (pipe JSON to stdin, exit 0 = allow, JSON output = deny)
echo '{"hook_event_name": "PreToolUse", "tool_input":{"command":"kubectl get pods"}}' | ./hooks/scripts/validate-cloud-command.py
echo '{"hook_event_name": "BeforeTool", "tool_input":{"command":"terraform apply"}}' | ./hooks/scripts/validate-cloud-command.py

# Test session start message
python3 hooks/scripts/validate-cloud-command.py --session-start claude
python3 hooks/scripts/validate-cloud-command.py --session-start gemini

Architecture

The entire validation engine is a single Python script (hooks/scripts/validate-cloud-command.py) with no dependencies beyond the stdlib.

• hooks/hooks.json — Primary hook file for Gemini CLI (uses BeforeTool event).
• hooks/claude-hooks.json — Dedicated hook file for Claude Code (uses PreToolUse event).
• .claude-plugin/plugin.json — Plugin manifest. Points to hooks/claude-hooks.json.
• hooks/scripts/validate-cloud-command.py — Core logic. Flow: read JSON from stdin → detect platform from hook_event_name → split_compound_command (state-machine shell parser) → normalize_segment (unwraps sudo, env, xargs) → dispatch to per-provider check_* function → deny() with platform-specific JSON output if unsafe.
• skills/deployshield/SKILL.md — Context injected into Claude when working with infra files.
• skills/gemini/SKILL.md — Context injected into Gemini CLI.