The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute node .gitnexus/run.cjs when found in the repository target. Because the repository being reviewed is potentially untrusted, this allows for the execution of arbitrary code provided by the code author or PR contributor.
[REMOTE_CODE_EXECUTION]: The skill uses npx gitnexus to run the GitNexus CLI tool from the official npm registry as a fallback mechanism.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests untrusted data from GitHub PRs, git diffs, and source files during the review process.
Ingestion points: The workflow involves reading content from PR URLs, local staged/unstaged changes, and untracked files via SKILL.md and associated personas.
Boundary markers: The persona files (e.g., ci-adversarial-lens.md, ci-security-lens.md) contain explicit instructions to treat all data in the checkout and diff as "hostile review data" and to never follow instructions found within.
Capability inventory: The skill utilizes git, gh (GitHub CLI), node, and a suite of GitNexus MCP tools such as query, impact, and explain to analyze the repository.
Sanitization: There is no mention of automated sanitization or filtering of the diff content; security relies on the AI model's adherence to the persona guidelines.

gitnexus-review