git-workflow
Pass
Audited by Gen Agent Trust Hub on Oct 29, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill analyzes local git repository state, including commit messages, diffs, and logs, which serve as ingestion points for potentially untrusted data from external sources or collaborators.
- Ingestion points: Workflow steps in
SKILL.mdandreferences/workflow.mdrequire the agent to inspect repository status, diffs, and commit history (e.g.,git log,git diff), which may contain malicious instructions embedded by third parties. - Boundary markers: While the output follows a strict structure defined in
assets/templates/output.md, the instructions do not explicitly mandate delimiters or escaping for the raw git data ingested during the analysis phase. - Capability inventory: The skill produces plans for shell command execution (Git) and utilizes local Python scripts for validation logic. It is capable of generating destructive commands like
git reset --hardandgit push --force. - Sanitization: The workflow includes manual diff review and automated scans for secrets (e.g., operation O-04 in
assets/examples/happy-path.md), but lacks specific automated sanitization to prevent the agent from obeying instructions found within the repository content. - [DYNAMIC_EXECUTION]: The skill includes and utilizes local Python scripts (
scripts/validate-output.pyandtests/run-validation.py) to validate that agent-generated plans meet safety requirements. - Evidence: The
tests/run-validation.pyscript employsimportlib.utilto dynamically load the validation logic fromscripts/validate-output.pyusing a static relative path.
Audit Metadata