The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest and process data from git logs, diffs, and conflict markers, which can contain content controlled by external actors.
Ingestion points: The skill reads from git log, git show, git status, and conflict files (e.g., rates.ts.conflicted) across multiple workflows (resolve-conflicts, integrate-upstream-changes, recover-lost-work).
Boundary markers: No specific delimiters or "ignore embedded instructions" warnings are prescribed when reading untrusted git data.
Capability inventory: The skill has access to high-impact commands including git push --force-with-lease, git rebase, git branch -D, git reset --hard, and git worktree remove.
Sanitization: No sanitization or validation of git metadata (author names, commit messages) is implemented before processing.
[DYNAMIC_EXECUTION]: The find-the-commit-that-broke-it workflow utilizes git bisect run, which executes an external test script at runtime.
Evidence: references/bisect-and-archaeology.md provides instructions for writing and executing test scripts (e.g., git bisect run <script>).
Mitigation: The skill includes a safety recommendation to run the script by hand on known-good and known-bad commits before automating the bisection.
[COMMAND_EXECUTION]: The skill is heavily dependent on the Git CLI, performing numerous shell operations to manage repository state, history, and worktrees.
[EXTERNAL_DOWNLOADS]: The skill references the external tool git-filter-repo for history rewriting.
Evidence: references/history-rewrite-guardrails.md suggests using git-filter-repo and notes it is a separate install not present by default. No automated installation or remote script piping is suggested.

git-workflow