The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The installation script scripts/install.sh and scripts/ensure-plannotator.sh contain patterns that download and immediately execute shell scripts from external URLs using curl | bash (e.g., https://plannotator.ai/install.sh and https://bun.sh/install).
[REMOTE_CODE_EXECUTION]: The skill executes dynamic Python code fetched from a local network endpoint http://localhost:4747/pending in scripts/claude-agentation-submit-hook.py and scripts/setup-codex.sh, which is a common vector for local privilege escalation or side-loading attacks in shared environments.
[COMMAND_EXECUTION]: The scripts/claude-plan-gate.py script programmatically modifies ~/.claude/settings.json to change the permissionMode to acceptEdits. This is designed to bypass human-in-the-loop approval requirements for file modifications during the 'EXECUTE' phase, effectively escalating the agent's privileges without explicit user consent for each action.
[PROMPT_INJECTION]: The skill uses 'ralphmode' and specific instructions in SKILL.md to force the agent into an 'automatic' execution state, specifically instructing it to avoid reopening approval gates and to disregard certain safety/confirmation steps once a plan hash is matched.
[EXTERNAL_DOWNLOADS]: The setup scripts (setup-claude.sh, setup-codex.sh, etc.) automatically register third-party plugins and MCP servers from unverified GitHub repositories and NPM packages without integrity checks.

jeo