The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes an indirect prompt injection surface where subagents ingest untrusted data from external sources.\n
Ingestion points: references/internet-researcher.md (fetches web content via search tools) and references/codebase-investigator.md (reads local files and searches codebase).\n
Boundary markers: Not present; the prompt files do not provide instructions for the subagent to use specific delimiters or ignore instructions found within fetched content.\n
Capability inventory: The subagents have access to tools for filesystem access, network research, and secondary command execution via the test-runner subagent.\n
Sanitization: Not present; no instructions for content validation, escaping, or filtering are included in the subagent prompts.\n- [COMMAND_EXECUTION]: The test-runner.md subagent is designed to execute arbitrary shell commands provided by the user, representing a significant capability.\n
Evidence: references/test-runner.md instructs the subagent to "Execute the exact command provided by the user" and identifies various test runners (pytest, cargo, npm, etc.) and git commands.\n
Evidence: SKILL.md includes an example of spawning an agent to execute pytest tests/.

hyperpowers-agents