The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted source code from a user-defined TARGET_PATH to be refactored by the worker agent. While this code could theoretically contain instructions intended to influence the agent, the risk is mitigated by the highly specific system prompts for both the worker and reviewer agents, which constrain their actions to refactoring tasks and code quality improvements.
Ingestion points: Source code located at the user-specified TARGET_PATH (referenced in SKILL.md and worker-prompt.md).
Boundary markers: No explicit boundary markers or "ignore instructions" warnings are present in the prompts.
Capability inventory: The worker agent has the capability to modify files and execute a user-provided test command. The reviewer agent can view diffs and send messages. A local script, guard-idle-worker.sh, is executed via a platform hook.
Sanitization: No sanitization of the input source code is performed before processing.
[COMMAND_EXECUTION]: The skill intentionally prompts the user for a test command and executes it to verify the safety of refactoring changes. Additionally, it uses a local helper script (guard-idle-worker.sh) via a platform TeammateIdle hook to manage the interaction between the worker and reviewer agents. These executions are transparently described and are essential to the skill's primary function.

refactoring-team