The Agent Skills Directory

[Command Execution]: The skill relies extensively on executing local shell scripts (async-review.sh and check-async-review.sh) to manage background tasks, git worktrees, and project builds. This involves spawning multiple background processes and running project-specific commands like npm run build or npm test.
[Indirect Prompt Injection Surface]: The skill fetches and processes external data from GitHub, specifically PR diffs and CI logs, which are then passed to an LLM (gemini executable) for review and test execution. This creates a surface for indirect prompt injection if the PR content contains malicious instructions intended to influence the AI's review or command generation.
[Automated External Tools]: The skill uses the GitHub CLI (gh) to fetch data and check PR status. It also dynamically invokes a gemini executable. The integrity of these external tools and the paths from which they are loaded (${HOME}/.gcli/...) should be verified by the user to ensure expected behavior.
[Direct Command Flow]: The system processes the output of shell commands to synthesize a final assessment. Errors or malicious output from these commands could theoretically influence the final recommendation provided to the user.
[Ephemeral Worktree Management]: The skill creates and removes git worktrees in the .gemini/tmp directory. While this is done to prevent workspace pollution, it involves commands that modify the git state and local file system based on PR numbers.

async-pr-review