The Agent Skills Directory

[SAFE]: The skill enforces secure git practices by explicitly prohibiting the use of global additions (git add . or git add -A) and requiring selective indexing of files, which prevents the accidental inclusion of untrusted or sensitive data.\n- [SAFE]: The instructions include specific rules to identify and exclude files that typically contain sensitive information, such as .env files, credentials, and private keys (*.pem, id_rsa*).\n- [SAFE]: A mandatory human-in-the-loop validation step is required; the agent must present a detailed commit plan and obtain explicit user approval before performing any indexing or commit operations, significantly reducing the risk of unintended actions.\n- [PROMPT_INJECTION]: The skill has a surface for indirect prompt injection as it processes untrusted content from the repository's file system through commands like git diff and git status. However, this is assessed as safe due to the following evidence chain:\n
Ingestion points: Untrusted data enters the agent context via the output of git status, git diff, and git log as specified in the Step 1 (Inspecter) of SKILL.md.\n
Boundary markers: The skill explicitly mandates a stop and wait for human validation ("S'ARRÊTER et attendre une validation explicite") after the planning phase, which acts as a robust boundary against automated exploitation.\n
Capability inventory: The agent is authorized to use git add and git commit to modify the project's history.\n
Sanitization: The agent is instructed to filter out sensitive file patterns and must present all proposed commit messages and file lists for manual review, allowing the user to detect any injected content before it is committed.

git-commit