The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains multiple sections of instructions specifically designed to override the agent's interactive safety protocols. In SKILL.md, it commands 'Never ask for permission to continue' and 'no confirmation prompts'. In references/companion-mode.md, the injected system prompt explicitly states 'OVERRIDE all interactive patterns', 'skip confirmation gates', 'NEVER ask questions', and 'Asking a question = task failure'.
[PROMPT_INJECTION]: The 'Auto-continue on pending questions' mechanism described in companion-mode.md is a programmatic injection attack that automatically resumes the session with forced 'Yes' or 'Resolve autonomously' responses whenever the agent attempts to halt for human decision-making or safety confirmations.
[COMMAND_EXECUTION]: The orchestration logic in references/orchestration-steps.md involves executing shell commands derived from repository files (e.g., CLAUDE.md, package.json, Cargo.toml). This includes running typechecks, tests, and infrastructure checks (like docker compose) without human review of the commands being executed.
[COMMAND_EXECUTION]: The skill documentation in companion-mode.md advocates for the use of the --dangerously-skip-permissions CLI flag, which removes the platform's security boundaries for file system access and tool execution.

bmad-sprint-run