The Agent Skills Directory

[DATA_EXFILTRATION]: The skill programmatically reads the contents of modified and untracked files to attach them to GitHub issues. This process uses git diff and cat on all untracked files not excluded by .gitignore. This poses a significant risk of exposing sensitive information, such as environment variables, local secrets, or private configuration, if they are present in the workspace but not yet ignored.
[COMMAND_EXECUTION]: The skill relies on extensive shell command execution to perform its primary functions.
Evidence: In references/workspace-changes.md, the skill executes git ls-files --others --exclude-standard -z | xargs -0 cat to read the full content of untracked files.
Evidence: The skill implements a cleanup procedure that executes rm on a list of files identified as part of the workspace changes, which could lead to data loss if not carefully reviewed by the user.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted data from the local filesystem.
Ingestion points: references/workspace-changes.md reads the content of modified and untracked files.
Boundary markers: While file contents are wrapped in markdown code blocks, there are no explicit instructions for the agent to ignore potentially malicious instructions embedded within the file content itself.
Capability inventory: The agent has access to gh for remote issue creation, git push for branch creation, and rm for file deletion.
Sanitization: No sanitization or filtering of file content is performed before interpolation into the issue draft.

create-issue