The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill relies on analyzing untrusted repository contents, logs, and reports to differentiate between disposable artifacts and essential business data.
Ingestion points: The agent reads and interprets local repository files, including capture.md, judge.md, round.json, and various generated logs/reports (SKILL.md, references/history-cleanup.md).
Boundary markers: The instructions explicitly mandate verifying the purpose of files and forbid guessing based on names/sizes. User confirmation is required for history rewriting and remote synchronization.
Capability inventory: The skill utilizes powerful capabilities including index manipulation (git rm --cached), file system deletion, history rewriting (git filter-repo), and remote pushing (git push --force-with-lease).
Sanitization: No formal sanitization of the content of analyzed files is performed; the agent relies on its internal logic to interpret the 'purpose' of files.
[COMMAND_EXECUTION]: The skill executes high-impact Git commands and third-party utilities. Misinterpretation of the repository state or malicious filenames could potentially lead to unintended data loss during cleanup or history filtering.
[SAFE]: The skill implements several security best practices to mitigate command injection and data loss:
Path Safety: Uses NUL-separated path lists (--pathspec-file-nul) to safely handle filenames containing spaces, newlines, or special shell characters.
Isolation: History rewriting is performed in an external mirror clone rather than the active working directory to ensure recoverability.
Integrity: Mandatory creation and verification of git bundle backups are required before any history modification.
Verification: Employs --force-with-lease for remote updates to prevent overwriting concurrent work on the server.

clean-artifacts