workflow-test-fix
Pass
Audited by Gen Agent Trust Hub on Apr 4, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes various shell commands using the Bash tool, including git for version control (commits and rollbacks), npm for test execution, and a custom CLI tool called 'ccw' for interacting with LLM models and loading project specifications.
- [PROMPT_INJECTION]: The skill uses specific instructions labeled as 'COMPACT DIRECTIVE' and 'COMPACT SENTINEL' to influence how the agent platform manages its internal context window. These instructions direct the agent to preserve certain content during summarization or compression to maintain operational state.
- [DATA_EXPOSURE]: The workflow involves reading and analyzing project source code, test reports, and session metadata. This information is passed to sub-agents and external AI services (Gemini, Qwen, Codex) via the 'ccw' tool to perform its core functions.
- [INDIRECT_PROMPT_INJECTION]: The skill presents an indirect injection surface by processing external codebase content (Phase 2) and user-provided descriptions (Phase 1) which are then used as context for AI-driven task generation and code fixing. While this is a common characteristic of development-focused agents, the skill lacks explicit sanitization or boundary markers for this ingested content.
- [DYNAMIC_EXECUTION]: The skill automatically generates and executes new code in the form of tests and project fixes based on AI analysis. This behavior is the primary intended function of the test-fix cycle and is controlled via task-specific logic and success criteria thresholds.
Audit Metadata