guardian
Pass
Audited by Gen Agent Trust Hub on Oct 22, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONPRIVILEGE_ESCALATIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted repository data, creating a potential attack surface for indirect prompt injection.\n
- Ingestion points: The skill reads Git diffs, commit logs, and PR descriptions as part of its pr and audit recipes, as seen in SKILL.md and reference/history-audit.md.\n
- Boundary markers: SKILL.md includes warnings that AI-assisted code review requires enhanced human review of intent and tradeoffs, and that AI review is a filter, not a replacement for human judgment.\n
- Capability inventory: Extensive repository management capabilities including git merge, git push --force-with-lease, git branch -D, and gh pr merge are documented in reference/git-recipes.md and reference/pr-ship-flow.md.\n
- Sanitization: SKILL.md mandates stripping session and tool metadata from commits and forbids auto-merging without explicit user consent.\n- [COMMAND_EXECUTION]: The skill makes extensive use of the git and gh (GitHub CLI) tools to manage code repositories.\n
- Evidence: Recipes in reference/git-recipes.md and reference/pr-ship-flow.md detail the use of shell commands for creating PRs, merging branches, and reshaping history. While these are standard development tasks, the skill's ability to execute these commands provides a powerful capability set that relies on user-verified guardrails like Ask First gates.\n- [PRIVILEGE_ESCALATION]: The skill documentation acknowledges the use of high-privilege flags in specific CLI tools.\n
- Evidence: reference/pr-ship-flow.md mentions the --admin flag for gh pr merge as an Ask First gate, noting the compliance and audit risks associated with bypassing branch protection rules.\n- [DYNAMIC_EXECUTION]: The skill generates Git command scripts for complex operations like history reshaping.\n
- Evidence: reference/history-reshape.md provides templates for multi-step Git scripts that the agent is expected to propose and execute after obtaining user consent. This is categorized as low risk as it uses known templates for standard developer tools.
Audit Metadata