The Agent Skills Directory

[SAFE]: The skill is strictly instructional and focuses on safety hardening. It advocates for a 'graduation-by-guardrail' approach, requiring deterministic blocks on destructive operations (such as force-pushes or hard resets) via repository hooks like .claude/hooks/git-guard.py before unattended execution is permitted.
[SAFE]: It implements 'Context Firewalls' to mitigate indirect prompt injection risks. By processing work items in isolated, disposable sub-agent contexts and returning only distilled summaries to the main orchestrator, the skill prevents the accumulation of untrusted external data (such as issue bodies or comments) in the long-term context.
[SAFE]: The methodology mandates the use of automated stop conditions, including cost ceilings, iteration caps, and wall-clock timeouts. This ensures that looping or stuck agents are automatically terminated, preventing resource exhaustion.

autonomous-loop