loop-engineering
Pass
Audited by Gen Agent Trust Hub on Oct 16, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill utilizes a local Python script (
scripts/validate_loop_contract.py) to validate loop contracts before execution. It also provides examples of running package-specific commands (e.g.,pnpm test) within a bounded loop environment.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external data such as CI logs and repository files within autonomous loops, creating a potential ingestion surface for indirect prompt injection.\n - Ingestion points: Contract inputs defined in
SKILL.mdand failure logs mentioned inreferences/ci-repair-loop-example.md.\n - Boundary markers: The skill explicitly requires a 'Permission boundary' in the contract and utilizes
scripts/validate_loop_contract.pyto enforce its presence.\n - Capability inventory: The agent is authorized to perform bounded actions (file writes, test execution) as specified in the validated contract.\n
- Sanitization: The framework requires an 'independent verifier' (e.g., deterministic test results) to confirm success, preventing the agent from self-certifying completion based on potentially poisoned output.
Audit Metadata