The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local bash script (scripts/agent-eval/audit.sh) using arguments such as <VERSION> and <question> that are directly collected from user input via the AskUserQuestion tool. The interpolation of these unvalidated strings into the shell command creates a potential risk for command injection.
[EXTERNAL_DOWNLOADS]: The workflow clones a large variety of external git repositories from corpus.json to the /tmp/codegraph-corpus directory. While many targets are well-known repositories, the skill facilitates the broad downloading and indexing of external third-party code.
[COMMAND_EXECUTION]: The audit process involves significant system-level interaction, specifically mutating the global codegraph installation and later restoring it, which may have unintended side effects on the stability or security of the host environment.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it ingests content from external repositories and reads tool-generated log outputs (parse-run.mjs, parse-session.mjs) to report benchmarking results. Evidence chain:
Ingestion points: External repository code and benchmark log outputs (Step 6).
Boundary markers: None identified in instructions.
Capability inventory: Shell script execution, git operations, and local file system modifications.
Sanitization: No sanitization or validation of external or user-supplied content is documented.

agent-eval