The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes specialized scripts to launch and supervise processes.
scripts/peer-job-runner.py is a detached process manager that executes arbitrary command-line arguments (argv) provided by the agent.
scripts/elevation-dispatch.sh executes the claude CLI to perform reasoning-heavy tasks, although it incorporates security flags such as --safe-mode and restricted --tools.
[REMOTE_CODE_EXECUTION]: The skill can fetch and process external content.
scripts/packs-resolve.py identifies and clones external Git repositories (referred to as 'Compound Packs') using git clone or git fetch based on URLs provided in repository configuration files (config.yaml).
[INDIRECT_PROMPT_INJECTION]: The planning workflow is designed to ingest and process data from multiple external surfaces, which constitutes an attack surface for indirect injection.
Ingestion points: Research findings are aggregated from the web (via web-researcher.md), Slack conversations (via slack-researcher.md), and external Git repositories (via learnings-researcher.md).
Boundary markers: The references/reasoning-elevation.md file contains explicit instructions for the agent to treat research dossiers and dialogue history as 'untrusted data' that must be interpreted rather than followed as instructions.
Capability inventory: The skill can execute shell commands through peer-job-runner.py and manage files across the project workspace.
Sanitization: Instructions in references/reasoning-elevation.md require the orchestrating model to validate the output returned from elevated reasoning steps to ensure it is a valid plan and not malicious instructions.

ce-plan