The Agent Skills Directory

[COMMAND_EXECUTION]: The orchestrator script scripts/run_arch_epic.py uses subprocess.Popen to invoke external AI agent CLI tools and manage repository-level commands. This behavior is the core functional requirement of the skill and is implemented with security-conscious shell quoting to prevent argument injection.
[PROMPT_INJECTION]: The skill exhibits a standard indirect prompt injection surface by ingesting and processing untrusted data from the repository (such as code files and worklogs) via child agents. However, the system prompts for these agents utilize structured boundary markers and explicit scoping instructions to mitigate the risk of unintended instruction execution.
Ingestion points: Repo files, epic and sub-plan documents, and worklogs are read and passed to child agent prompts.
Boundary markers: The documentation in references/auto-harness-prompts.md defines structured system prompts with headers like ## Mission and ## Boundaries to constrain child agent behavior.
Capability inventory: The skill possesses the capability to execute commands via the claude, codex, and grok CLIs, as documented in the workflow references.
Sanitization: The _shell_quote utility in scripts/run_arch_epic.py ensures all dynamically generated command arguments are properly escaped.

arch-epic