The Agent Skills Directory

[PROMPT_INJECTION]: Indirect prompt injection surface detected in the task orchestration workflow.\n
Ingestion points: Task descriptions and notes are read from JSON files in the .tasks/ directory, which are populated via CLI arguments in scripts/task_tracking.sh.\n
Boundary markers: scripts/run_task_loop.sh interpolates these values into a prompt using only basic Markdown headers, which are insufficient to prevent an agent from following instructions embedded in the data.\n
Capability inventory: The agent is empowered to execute arbitrary shell commands (via verify-command) and perform git operations.\n
Sanitization: No validation or escaping is applied to untrusted task data before it is included in the AI prompt.\n- [COMMAND_EXECUTION]: The skill facilitates the execution of arbitrary shell commands.\n
scripts/run_task_loop.sh provides a verify-command to the agent and explicitly instructs it to execute the command for verification.\n
The orchestration script uses the --dangerously-skip-permissions flag when calling the Claude CLI, disabling safety prompts for tool execution.

task-tracking