The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to read and process data from local Git repositories, which acts as a surface for indirect prompt injection from untrusted commit messages or diffs.
Ingestion points: The skill uses git status, git diff, git log, and git show to ingest repository content into the agent's context (found in SKILL.md and references/snapshots-and-refs.md).
Boundary markers: No explicit delimiters or instructions are provided to the agent to treat repository-sourced data as untrusted or to ignore embedded instructions.
Capability inventory: The skill enables high-impact operations including git push, git reset --hard, git merge, and git rebase (found in SKILL.md and references/integration-and-recovery.md).
Sanitization: The instructions lack specific validation or sanitization steps for data read from the repository before it influences agent actions.
[METADATA_POISONING]: The skill contains anachronistic or deceptive date information in its documentation and license.
Evidence: references/snapshots-and-refs.md claims a research date of "2026-09-11" and license.txt contains a "Copyright (c) 2026" statement, both of which are future-dated.

manage-git-state