The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external task data from a controller, which introduces a surface for instructions that could attempt to manipulate the agent's behavior.
Ingestion points: The skill utilizes the $ARGUMENTS placeholder in SKILL.md and explicitly directs the agent to read the "complete assigned task" from a canonical controller in references/implement.md.
Boundary markers: The instructions contain directives to "Preserve task references and directives literally" and "Stay within the assigned task", which provide guidance but do not programmatically enforce isolation from adversarial inputs.
Capability inventory: The agent has the authority to modify files, perform git operations, and execute local CLI tools (yy task, yy merge) within the product worktree.
Sanitization: No explicit content sanitization, filtering, or instruction-override protection is documented for the external task data being processed.
[DYNAMIC_EXECUTION]: The scripts/kanban.sh script employs dynamic execution techniques to manage the skill's runtime environment.
Evidence: The script uses eval to execute shell commands and environment variable exports produced by the controller_resolver.py script.
Evidence: The script uses exec to invoke Python-based helper scripts (e2e_sweep_helper.py, e2e_housekeeping.py) to perform validation and auditing logic at runtime.
[COMMAND_EXECUTION]: The skill's primary function involves orchestrating multiple command-line utilities to manage the development lifecycle.
Evidence: references/implement.md and scripts/kanban.sh contain logic to execute git, yylo-ledger, juno-kanban, and various yy subcommands to fulfill task implementation and record-keeping requirements.

ralph-loop-yylo