The Agent Skills Directory

[SAFE]: The skill incorporates a robust 'Forbidden Without Explicit User Approval' policy that prevents the agent from executing destructive commands like 'rm -rf' or 'git reset --hard', modifying environment variables, or installing new dependencies without user confirmation.
[COMMAND_EXECUTION]: The skill utilizes shell commands for repository inspection (git status), verification (lint, typecheck, tests), and version control management. These actions are performed within the local repository context to ensure implementation quality.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process implementation plans authored by other AI models, which serves as a potential vector for indirect prompt injection. This risk is effectively mitigated by the 'Reality-Check' requirement, which forces the agent to validate all instructions against the existing codebase rather than following the plan blindly.
[DATA_EXFILTRATION]: An auto-push feature is included to synchronize verified work with remote branches. While this involves outbound data transfer, the skill limits its scope to non-protected working branches and mandates a diff inspection phase to ensure no sensitive data like secrets or debug logs are included in the commits.

frontier-plan-opencode-executor