The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes several internal scripts to perform its operations. These include packs-resolve.py for configuration management, peer-job-runner.py for process supervision, and light-webserver.js for serving visual mockups. These scripts are executed as local sub-processes by the agent.\n- [REMOTE_CODE_EXECUTION]: To handle computationally expensive reasoning tasks, the skill dispatches work to the Claude CLI using the elevation-dispatch.sh script. This process is designed with security in mind, explicitly restricting the model's access to a read-only set of tools and a specific handoff directory.\n- [PERSISTENCE]: The skill includes a background task manager (peer-job-runner.py) that uses platform-specific techniques, such as double-forking on POSIX and Job Objects on Windows, to allow long-running operations to continue independently of the agent's main session lifecycle.\n- [INDIRECT_PROMPT_INJECTION]: The skill explores data from the local repository and Slack, which potentially introduces a surface for indirect prompt injection. However, the skill explicitly instructs the agent to treat this external context as untrusted data and applies strict tool restrictions to model-elevation calls to mitigate risk.

ce-brainstorm