The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the execution of local binaries such as vibereview, dot, and dot-html. In skills/randroid-vibereview/SKILL.md, the instructions explicitly direct the agent to locate and verify these binaries, including a check for unknown symlinks.
[DYNAMIC_EXECUTION]: The randroid-loop workflow uses scripts/randroid-loop.sh to dynamically invoke the codex exec command, launching new model iterations to process tasks.
[PERSISTENCE]: A platform-specific Stop hook is defined in hooks/hooks.json and implemented in hooks/stop-hook.sh. This hook intercepts the agent's session exit to re-trigger the loop, maintaining the autonomous process until a completion condition is met.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data which presents an injection surface.
Ingestion points: The agent reads external content from GitHub pull request comments (randroid-address-pr-comments) and project task files stored in .dots/ (randroid-loop).
Boundary markers: The instructions do not define specific delimiters or "ignore" rules to prevent the agent from following instructions embedded in PR comments or task descriptions.
Capability inventory: Across its scripts, the skill can write to the local filesystem, perform git operations (commit, push, merge), and execute shell commands.
Sanitization: No sanitization or validation of external input from PRs or task ledgers is mentioned, allowing potential malicious instructions in those sources to influence the agent's behavior.

randroid