The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on several local bash scripts (report.sh, verify.sh) and the GitHub CLI (gh) to automate git operations. These scripts use standard git commands (log, diff, merge-tree, ls-tree) to manage the repository state. The execution is scoped to the local repository environment and includes verification steps to ensure fork-only files remain byte-identical after merging.
[PROMPT_INJECTION]: The skill processes untrusted data from the upstream repository, specifically commit messages and file diffs, during the 'Recommend' and 'Resolve' phases. This presents an indirect prompt injection surface.
Ingestion points: Commit logs from git log and commit details from git show are analyzed in Step 2; file contents are processed during conflict resolution in Step 5.
Boundary markers: No explicit delimiters or warnings are used to prevent the agent from being influenced by instructions embedded within commit messages or diffs.
Capability inventory: The agent can execute shell scripts, perform git merges, and create pull requests via the GitHub CLI.
Sanitization: No sanitization or filtering of external commit metadata or file content is performed before the agent evaluates it.
[REMOTE_CODE_EXECUTION]: In Step 6, the skill uses git archive and tar to extract and execute a version of its own scripts from a different branch. While this involves dynamic execution, the source is the repository's own git history rather than an arbitrary remote URL. This technique ensures that the verification logic matches the code in the sync branch.

check-upstream