Security contact: coding@hazli.dev

Markdown skills and optional Python scripts that run inside the user’s agent environment. They can instruct the agent to read client files the user points at. Treat installed skills as code that influences access to sensitive financial data.

Please do not open a public issue for:

• Prompt-injection patterns in skills that exfiltrate data
• Scripts that execute unsafe shell by default
• Secrets accidentally committed to the repo
• Ways to bypass mathematical or finalisation gates in a way that could mislead non-experts

Prefer private disclosure via GitHub Security Advisories on the repository, or email coding@hazli.dev (see MAINTAINERS.md).

Include:

1. Affected plugin/skill path
2. Description of the behavior
3. Impact (data exfiltration, silent imbalance, etc.)
4. Minimal reproduction

We aim to acknowledge within 7 days and ship fixes as soon as practical.

• No credentials, API keys, or client data in the repo
• Community skills must go through accounting-builder-hub:skills-qa before recommendation
• scripts/ must not download and execute remote code
• Hooks, if added, must be documented and least-privilege

• Install plugins from sources you trust
• Prefer user-scoped installs you control
• Never paste production secrets into chat logs you don’t control
• Review draft financial statements before any external issuance

Skills and scripts may write workpapers/, Excel packs, and ledger/*.beancount containing client financials. Treat these as confidential. Do not commit real client folders. Use fixtures/ only with synthetic data.