The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructs the agent to ignore interactive safety patterns, such as 'ask the user' or 'wait for the user' prompts, effectively bypassing human-in-the-loop controls. It specifies trigger phrases like 'answer the questions yourself' or 'run it to the end without me' that transition the agent into this autonomous state.
[INDIRECT_PROMPT_INJECTION]: The activation mechanism relies on text matching within the agent's context. If these trigger phrases appear in untrusted data (e.g., code comments, documentation, or PR descriptions) that the agent processes, it may move into an unattended mode without the user's intent. 1. Ingestion points: User prompts and project files processed by the pipeline in SKILL.md. 2. Boundary markers: No explicit delimiters are defined to separate user commands from processed data regarding triggers. 3. Capability inventory: Autonomous use of git push, PR creation, and issue management (SKILL.md). 4. Sanitization: No sanitization is performed on external content to prevent activation.
[COMMAND_EXECUTION]: The skill facilitates the automated execution of shell-based workflows, including git operations and PR management, without interactive confirmation, relying on boundaries defined at the start of the session.

unattended-run