The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill reconstructs session state using a "state-source ladder" that reads from local files and git history. \n
Ingestion points: The agent uses view_file to read from .agents/artifacts/HANDOFF.md and .agents/context/current.md, and bash to parse the output of git commands. \n
Boundary markers: The instructions define a brief summary format but lack explicit delimiters or safety instructions for the ingested text to prevent the agent from following instructions embedded within the state data. \n
Capability inventory: The agent utilizes the bash, view_file, and write_to_file tools, providing a wide scope of action if the agent is influenced by malicious state. \n
Sanitization: There is no evidence of sanitization or validation performed on the content retrieved from state-source files before identifying the "Next Step". \n- [COMMAND_EXECUTION]: The skill uses the bash tool to execute repository forensics commands (e.g., git status, git log, git diff) to identify the current work state. \n- [EXTERNAL_DOWNLOADS]: The documentation references installation via npx from a vendor-owned repository (harshsinghmp/muse-skills).

relay