integration-test

Warn

Audited by Gen Agent Trust Hub on Nov 4, 2026

Risk Level: MEDIUMPERSISTENCEPROMPT_INJECTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PERSISTENCE]: The skill contains instructions in references/mode-verify.md to modify system environment variables and shell profiles (such as ~/.bashrc, ~/.zshrc, or the Windows user PATH) to ensure that newly installed tools like uv or specific Python versions remain available across future sessions. While intended for functional readiness, this represents a persistence mechanism on the host environment.
  • [PROMPT_INJECTION]: Instructions in SKILL.md and references/fix-loop.md direct the agent to bypass standard user interaction steps. This includes skipping next-step prompts when running within a workflow and explicitly instructing the agent not to inform the user when internal state-management 'gates' or 'goals' are auto-cleared. These patterns reduce user oversight of the agent's internal operations.
  • [DYNAMIC_EXECUTION]: The skill's behavior is driven by commands extracted from a project configuration file (docs/project-config.json). It uses keys such as systemCheckCommand and quickRunCommand to execute shell operations. Additionally, the skill generates test code based on existing project patterns and subsequently executes that generated code using the configured test runner.
  • [EXTERNAL_DOWNLOADS]: The references/mode-verify.md file describes a protocol for provisioning the uv tool and Python runtimes if they are not detected on the system. This involves downloading software from external sources (e.g., astral.sh or official package managers like Homebrew and WinGet) and executing installers to set up the execution environment.
  • [INDIRECT_PROMPT_INJECTION]: The skill operates on untrusted project data, including configuration files and source code, which are used to determine which shell commands to run. There is a lack of explicit sanitization or validation of command strings read from docs/project-config.json before they are executed in the shell.
  • Ingestion points: docs/project-config.json, project source files, and feature specifications.
  • Boundary markers: The skill does not implement delimiters or 'ignore' warnings when interpolating configuration values into shell commands.
  • Capability inventory: Shell command execution (bash), sub-agent spawning (spawn_agent), file system modification, and git operations.
  • Sanitization: No sanitization or escaping is performed on external strings retrieved from project configuration files prior to execution.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 4, 2026, 10:12 AM