This repo is a public kit for directing coding agents. It contains no production services, credentials, or customer data — but it is the kind of code that teaches agents to handle untrusted input, so vulnerabilities in the skills, prompts, or scripts still matter.

Please report security issues through GitHub Private Vulnerability Reporting (repo → Security → Report a vulnerability) rather than a public issue, so the report is visible only to the maintainer.

What to include:

• The affected file(s) and a minimal reproduction (a prompt, a skill, or a script input that triggers it).
• Impact: what an attacker could do, and under what assumptions (e.g. "an agent runs this skill on a malicious PR description").
• Suggested fix, if you have one.

• Findings are triaged privately and fixed on main; a fix is announced in the commit/PR that lands it.
• Skills that read untrusted input (diffs, web pages, issues, PR bodies) are expected to follow the prompt-injection defense in docs/prompt-defense.md — a skill that doesn't is a bug worth reporting.
• Third-party skills are pinned in skills-lock.json and restored via the skills CLI; update them with the documented flow, never by editing this tree's copies.