The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill executes a bundled script (select-review-model.mjs) that performs network requests to api.codexradar.com to fetch model benchmarking data. This external data dynamically determines the model selection and timeout parameters for the code review process, introducing a dependency on a non-whitelisted third-party service for core logic configuration.
[COMMAND_EXECUTION]: The skill runs multiple Git and linting commands. It follows secure practices by using null-terminated strings (-z) and the -- separator when adding untrusted filenames to the Git staging area, effectively mitigating common command injection vectors associated with file manipulation.
[DATA_EXFILTRATION]: The skill establishes outbound network connections to codexradar.com domains. Although these appear to be for fetching configuration data, they constitute network operations to an external domain that is not recognized as a trusted vendor, potentially allowing for metadata exfiltration via User-Agent or request headers.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it processes untrusted content from the repository's git diff. This content is used by an LLM to generate CHANGELOG.md entries and perform code reviews, meaning specially crafted content in the codebase could attempt to manipulate the agent's behavior during these tasks.

codex-review