The Agent Skills Directory

[SAFE]: The skill implements a high-security posture through several design choices:
Agent Isolation: It requires spawning subagents with no inherited conversation history (fork_turns: 'none'), preventing context leakage between runs or from the coordinator.
Local-First Default: The skill explicitly prohibits remote publication or writes to services like Vercel, Cloudflare, or GitHub unless the user provides a specific, narrow instruction. It treats ambient tool authentication as a capability, not authorization.
Provenance and Integrity: It uses a coordinator-owned directory (.oneshot-provenance) to store immutable records of run identities and prompt digests, ensuring that worker-owned scripts cannot modify their own history or masquerade as successful runs.
Artifact Scanning: The included validate_catalog.py script automatically scans generated artifacts for sensitive files, such as SSH private keys, environment files, and project configuration data, before they are considered valid.
Path Traversal Defenses: The scripts use strict path resolution and child-validation logic to ensure that file operations (such as cleanup of temporary files) are confined to the specific run directory.

oneshot-websites