The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill reads and processes local repository documentation and historical pull request data to learn project conventions, which constitutes a potential injection surface.
Ingestion points: The skill reads files such as AGENTS.md, CONTRIBUTING.md, .github/copilot-instructions.md, and recent pull request descriptions as specified in the 'Detect conventions' section of SKILL.md.
Capability inventory: The agent is authorized to perform git operations (push, rebase, merge) and execute GitHub (gh) and GitLab (glab) CLI commands to create or update pull requests.
Sanitization: The skill uses a dedicated Python script (scripts/lint_body.py) to validate the drafted PR body against specific rules, providing a automated layer of oversight before the content is presented to the user.
Boundary markers: While no explicit delimiters are used for ingested content, the process enforces a strict user confirmation step ('Go ahead with this?') before any pushes or host writes occur.
[COMMAND_EXECUTION]: The skill utilizes local shell commands and sub-processes to perform repository introspection.
Evidence: scripts/lint_body.py executes git diff --name-only via subprocess.run to ensure that files referenced in the PR description are actually included in the diff.
Best Practices: The instructions in references/github.md and references/gitlab.md explicitly mandate the use of --body-file and --description-file flags rather than passing content via shell arguments, effectively mitigating common shell injection vectors.

submit-changes