If you discover a security vulnerability in Correctless, please report it responsibly:

1. Do NOT open a public GitHub issue for security vulnerabilities.
2. Email: jft.inbox@gmail.com with subject line [SECURITY] Correctless: <brief description>
3. Include:
   • Description of the vulnerability
   • Steps to reproduce
   • Potential impact
   • Suggested fix (if any)

You should receive a response within 48 hours. We will work with you to understand the issue and coordinate disclosure.

Correctless is a set of Claude Code skill prompts and bash hooks. Key security considerations:

• Hooks execute shell commands — the workflow gate and audit trail hooks run as bash scripts. They are registered in .claude/settings.json and execute with the user's permissions.
• No network access — hooks do not make network requests. Skills may use WebSearch/WebFetch for research but only when explicitly configured.
• No credential storage — Correctless does not store or transmit credentials. The health check detects hardcoded secrets and helps remove them.
• Output redaction — external-facing skills (PR review, contribution, maintainer review) redact paths, credentials, and hostnames before posting to GitHub/GitLab.
• Red team assessments (/credteam) require an isolated environment — the skill verifies isolation before executing.

The following are NOT security vulnerabilities in Correctless:

• The workflow gate can be bypassed via workflow-advance.sh override — this is intentional and logged
• Bash write commands that bypass the gate's pattern detection — this is a known limitation documented in the defense-in-depth section of the README
• Skills that spawn subagents inherit the user's permissions — this is by design