SmolForge Deletes Two Dangerous Skills

The postmortem’s sharpest lesson is that natural-language agent policy becomes production code when it changes authority or scope.

• –JFDI granted standing permission to push code, alter provider policy, apply migrations, retry releases, and bootstrap infrastructure.
• –The missing safeguard was a terminal stop condition when unrelated platform defects appeared.
• –Maintainability Guardrails created a quieter failure mode by making broad quality checklists part of nearly every task.
• –Skill risk depends not only on what a skill can do, but also on how broadly it activates.
• –SmolForge’s proposed answer—observable, versioned skill kits—could make deletion decisions evidence-based instead of hype-driven.