The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a Python script (scripts/goal-state-tracker.py) and a shell wrapper (scripts/goal-state-tracker.sh) to manage task states. These scripts use subprocess.run to execute local git commands (e.g., git rev-parse --show-toplevel) to identify the project root. This is a standard and safe operation for development tools.
[INDIRECT_PROMPT_INJECTION]: The skill describes workflows that ingest data from external sources such as web_search, gh search, and official documentation. While processing untrusted external content is a known surface for indirect prompt injection, the skill mitigates this through strict documentation templates (e.g., atomic-task-template.md), mandatory human review gates (User Nod), and a '3-Tries' failure recovery protocol that limits automated retries. Additionally, it recommends using isolated sub-agents for research tasks to prevent context pollution in the main session.
[SAFE]: The skill follows security best practices for AI agents, including:
Explicitly forbidding destructive operations like git reset, git rebase, or git clean -fd.
Enforcing 'Read-Modify-Verify' cycles to ensure changes are correctly applied.
Using a file-based 'source of truth' for state, which prevents the agent from hallucinating progress.
Implementing a gated re-entry system for failed tasks to prevent infinite loops and resource exhaustion.

goal-loop