The Agent Skills Directory

DATA_EXFILTRATION (MEDIUM): The skill contains logic to crawl the file system starting from the user's home directory. Specifically, the collapse-all-skills function accesses the HOME environment variable and uses listFiles() to enumerate directories and locate SKILL.md files.
Evidence: (str/replace dir "~" (System/getenv "HOME")) and .listFiles (java.io.File. expanded).
COMMAND_EXECUTION (LOW): The skill instructions include commands that execute local files using Babashka (bb) and the just task runner. This relies on the presence of a local script lib/wev_26_worlds.clj which is not provided in the snippet.
Evidence: bb lib/wev_26_worlds.clj, just wev-collapse.
INDIRECT_PROMPT_INJECTION (LOW): The skill is designed to ingest and process data from other SKILL.md files across the system. This creates a surface where a malicious skill file elsewhere on the machine could influence the agent's behavior during the 'collapse' protocol.
Ingestion points: SKILL.md files found in skill-dirs.
Boundary markers: Absent; the skill reads the file contents directly.
Capability inventory: File system enumeration, local script execution via bb.
Sanitization: None detected for the ingested file paths or metadata.

wev-tesseract