The Agent Skills Directory

[COMMAND_EXECUTION]: The skill frequently executes the git and gh (GitHub CLI) command-line interfaces to perform repository operations, read configurations, and interact with the GitHub API.
The implementation uses a robust Invoke-Tool wrapper in scripts/pr-common.ps1 that passes arguments as arrays (& $command.Source @Arguments), effectively preventing shell injection vulnerabilities.
[EXTERNAL_DOWNLOADS]: The skill interacts with GitHub, a well-known service, to fetch repository metadata, pull request details, and template files.
These operations are performed via authenticated GitHub CLI calls (gh api, gh pr) and standard Git commands (git fetch, git ls-remote).
These interactions are transparently documented in the preparation phase and are essential to the skill's primary function.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a processing surface for untrusted data by ingesting Git commit messages and file diffs to generate pull request summaries.
Ingestion points: scripts/prepare-pr.ps1 reads the output of git log and git diff to collect evidence for the PR body.
Boundary markers: The skill enforces a strict Markdown structure for the generated body using Assert-PrBodyStructure in scripts/pr-common.ps1, which helps maintain the integrity of the agent's output.
Capability inventory: The skill possesses significant capabilities, including the ability to push commits (git push) and modify PR metadata (gh pr create/edit).
Sanitization: While the skill performs structural validation, it relies on the underlying LLM's guardrails to handle potentially malicious instructions embedded in commit text. The multi-step 'Prepare -> Plan -> Approve -> Execute' workflow provides a strong human-in-the-loop defense against accidental execution of injected instructions.

git-remote-pr