The Agent Skills Directory

[SAFE]: The skill is entirely focused on implementing safety guardrails for autonomous agents. Notable security features include:
Destructive Command Prevention: scripts/deny-destructive-git.sh acts as a pre-tool-use hook that explicitly blocks commands like git reset --hard or git push --force to prevent the agent from accidentally deleting work.
Execution Integrity: scripts/run.sh implements a snapshotting mechanism that copies the main poller script before execution, preventing shell parsing errors or race conditions if the script is edited while running.
Fail-Closed Logic: The scripts scripts/pr-gate.sh and scripts/watch.sh are designed to fail closed, ensuring that if external tools like gh (GitHub CLI) fail or return empty results, the agent does not proceed with risky actions like merging code.
Data Integrity: scripts/backup.sh automatically creates git bundle and tar snapshots of the repository, providing a local recovery mechanism independent of remote services.
Indirect Injection Mitigation: Parsing logic in scripts/watch.sh and scripts/pr-gate.sh uses strict markers (e.g., first occurrence of specific headers) to prevent data from inside reports or briefs from being interpreted as control instructions.
[COMMAND_EXECUTION]: The skill uses local shell scripts to manage the agent loop and health checks. These operations are restricted to the local environment and the project repository.
[DATA_EXFILTRATION]: scripts/notify.sh uses osascript to send iMessage notifications to a configured phone number. This is a legitimate notification channel for an autonomous system, and the provided number is a placeholder.

autonomous-loop-safety