The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs persistent filesystem operations on other skill files located in ~/.copilot/skills/ using commands such as cat, cp, and git. It explicitly modifies existing skill logic and commits these changes to git repositories.
[COMMAND_EXECUTION]: The 'Eval Harness' mechanism encourages the generation and execution of arbitrary shell commands (e.g., grep, ls, wc) to verify success, which presents a risk of executing unintended or dangerous operations during autonomous loops.
[COMMAND_EXECUTION]: The skill recommends the use of the --allow-all flag and high autopilot continuation limits (--max-autopilot-continues 100), which effectively bypasses human-in-the-loop safety checkpoints for a large number of automated iterations.
[EXTERNAL_DOWNLOADS]: Suggests using npx lighthouse to evaluate performance metrics, which involves downloading and executing software from the npm registry at runtime.
[PROMPT_INJECTION]: Includes instructions for the user to override standard safety protocols by using specific CLI flags that suppress confirmation prompts, enabling the agent to operate without restrictions.
[DATA_EXFILTRATION]: While primarily focused on local optimization, the suggested use of tools like lighthouse or custom node scripts for evaluation can involve network requests to external endpoints.

autoresearch