The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements robust defenses against prompt injection by explicitly classifying external repository data (diffs, commit messages, tracker items) as 'passive data' and instructing the agent to ignore any instructions embedded within them. These rules are codified in passive-data-safety.md and integrated into the primary workflow in SKILL.md.\n- [CREDENTIALS_UNSAFE]: The skill enforces a mandatory secret redaction policy, requiring the agent to identify and replace sensitive information like API keys or passwords with descriptive placeholders (e.g., [redacted:api-key]) before inclusion in commits or PR bodies.\n- [COMMAND_EXECUTION]: The skill utilizes standard version control and file system tools (git, cat, test) for its operations and provides a secure shell script template for calculating content digests with explicit file-readiness guards.\n- [INDIRECT_PROMPT_INJECTION]: The skill recognizes the surface area for indirect injection when processing data from git diff, commit history, and trackers. It mitigates this by identifying ingestion points (git logs, trackers), mandating boundary markers ('passive data' rules), inventorying capabilities (git/cat/gh), and requiring sanitization (secret redaction).\n- [EXTERNAL_DOWNLOADS]: No external code, scripts, or packages are downloaded from remote sources; all operations rely on tools and data already present in the local environment.

package-change