The Agent Skills Directory

[SAFE]: The skill implements a comprehensive and structured framework for autonomous iteration. No malicious patterns or data exfiltration attempts were detected.
[COMMAND_EXECUTION]: The skill relies heavily on executing shell commands for verification (Verify:) and safety gating (Guard:). However, it includes a mandatory 'Verify Safety Screen' (in autoresearch.md) designed to block dangerous patterns like rm -rf, fork bombs, curl|sh, and embedded credentials before any execution occurs.
[CREDENTIALS_SAFE]: The learn.md (wiki mode) protocol explicitly instructs the agent to never include verbatim values from .env or credentials. It employs a two-layer safety check: first, a prompt instruction to extract only variable names; second, a post-generation regex scan (grep) to detect and warn about leaked keys (AWS, Stripe, GitHub, etc.).
[PRIVILEGE_ESCALATION]: All deployment and publishing actions (in ship.md) are strictly human-gated. The skill explicitly states: 'Pause for explicit user approval before any push, publish, deploy, PR mutation, unpublish, or rollback action.'
[DATA_EXPOSURE]: The regression.md component includes a safety check for database migrations, enforcing an allowlist for DB URLs (localhost, CI containers, or _test/_ci suffixes) and refusing access to production environments.
[PROMPT_INJECTION]: No prompt injection or behavior override markers were detected. The skill uses natural-language goal parsing as a configuration step (plan.md) rather than an execution bypass.

autoresearch