spec-loop
Pass
Audited by Gen Agent Trust Hub on Oct 18, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data from plan.json, design.md, and code-review outputs, which could potentially contain instructions aimed at influencing agent behavior.
- Ingestion points: Project specifications in design.md, task definitions in plan.json, and findings from the code-review rq tool.
- Boundary markers: The skill explicitly instructs the agent to treat review text as 'untrusted data' and utilizes fresh subagents to isolate individual task implementations.
- Capability inventory: The skill leverages code-subagents, code-commit, code-pull-request, and git operations to automate the development lifecycle.
- Sanitization: Implementation subagents are constrained by rules that prevent unauthorized staging, committing, or software installation, and are required to return a BLOCKED status instead of prompting the user when hurdles are encountered.
Audit Metadata