The Agent Skills Directory

[DATA_EXFILTRATION]: The script git-commit-context.sh reads and prints the content of added or modified files from the local filesystem to the agent's context. When combined with the git-commit-push.sh script, which stages all changes (git add -A) and pushes them to a remote repository, there is a risk that sensitive information (such as API keys, environment variables, or private configuration) could be unintentionally committed and exfiltrated to a remote server.
[PRIVILEGE_ESCALATION]: SKILL.md instructs the agent to run scripts in "escalated execution mode" to ensure write access to the .git directory. This prompts the agent to seek higher privileges than typically required for standard tasks, which may bypass security boundaries depending on the execution environment.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted content from file diffs and new files in the working tree. This content is used as the primary source for generating commit messages, creating a vulnerability where malicious instructions within those files could manipulate the agent's output or subsequent actions.
Ingestion points: scripts/git-commit-context.sh (reads file content and git diffs).
Boundary markers: None; file content is printed directly into the context.
Capability inventory: scripts/git-commit-push.sh (performs git add, git commit, and git push).
Sanitization: None; raw data is passed into the agent context.
[COMMAND_EXECUTION]: The skill relies on shell scripts to execute powerful system commands, including git add -A, which stages all changes in the working tree without discrimination. This automated staging increases the likelihood of including unwanted or sensitive files in a commit.

pflow-commit