The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes repository files and user-provided inputs to drive automated workflows, creating an attack surface for indirect instructions.\n
Ingestion points: Workflow instructions are read from brief.md and repository configuration files as described in references/runbook-review-loop/prompts/common.md.\n
Boundary markers: The orchestration framework uses --- message --- delimiters for subagent isolation and provides instructions for agents to focus on specific runbook files.\n
Capability inventory: The orchestration engine in references/agent_runbooks.py can execute shell commands, write to the filesystem, and manage subagent sessions.\n
Sanitization: The engine uses shlex.quote for proper escaping of command-line arguments.\n- [COMMAND_EXECUTION]: The skill executes shell commands as part of its core workflow automation logic.\n
Evidence: Subagents are instructed to run project 'checks' (shell commands) provided via user input or repository metadata in references/runbook-review-loop/prompts/common.md.\n- [DYNAMIC_EXECUTION]: The engine supports executing dynamic Python logic defined within runbook flow scripts.\n
Evidence: The CodeStep implementation in references/agent_runbooks.py enables the execution of Python functions defined in flow.py at runtime.

agent-runbook-authoring