The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to automate repository management by executing a variety of shell commands. It interacts with the local Git binary, the GitHub CLI (gh), and specialized PR stack tools like Graphite (gt) and Git Town. It also executes an internal Python archive (plate.pyz) to perform resolution and validation tasks. The execution is scoped to the skill's primary purpose of Git workflow management.
[EXTERNAL_DOWNLOADS]: The skill provides instructions for the agent to install or upgrade external dependencies required for its operations. This includes the gh-stack extension for the GitHub CLI and third-party tools like Graphite and Git Town, which are installed via standard package managers such as Homebrew, NPM, or Chocolatey. These tools and sources are well-known and established within the developer community.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the local environment and remote GitHub metadata, creating a potential surface for indirect prompt injection. This is mitigated through explicit architectural constraints.
Ingestion points: The skill reads git status, diffs, log outputs, and PR metadata to generate commit messages and PR bodies.
Boundary markers: Instructions mandate the use of single-quoted heredoc delimiters for commit messages to protect against the interpolation of shell-active characters like backticks and dollar signs.
Capability inventory: The skill possesses the ability to execute git and gh commands that mutate the repository and remote pull requests.
Sanitization: The skill contains explicit rules to detect and reject credentials, .env files, and unexplained large binaries during the staging process, preventing their inclusion in the version control history.

plate