The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a wide range of shell commands (git, node, python, curl, sqlite3, grep, rg, ast-grep) and language-specific test runners (pytest, npm test, jest, cargo test, go test) to perform baseline checks and verify findings.
[REMOTE_CODE_EXECUTION]: The agent is instructed to dynamically create and execute reproduction scripts or one-liners using node -e and python -c. Additionally, the skill attempts to identify and run validation commands found within the repository's own Makefile, package.json, or CI configurations, which involves executing untrusted code from the repository environment.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. It ingests untrusted data (git diffs and repository source code) and uses that information to drive command execution and script generation. There are no explicit boundary markers or instructions to ignore commands or malicious directives embedded in the source code being reviewed.
Ingestion points: Git diffs (git diff), branch history, and repository files.
Boundary markers: Absent; no delimiters are used to separate code-under-review from agent instructions.
Capability inventory: Extensive shell execution, dynamic script generation (node -e, python -c), and network access via curl.
Sanitization: Absent; the skill does not specify any validation or filtering of the content extracted from the files before use in reproduction scripts.

ultrareview