The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses a mandatory human-approval step (Workflow Step 5) to mitigate risks from indirect prompt injection. This ensures the user reviews and approves all generated test scenarios and plans.
Ingestion points: User-provided functional specifications (Workflow Step 1).
Boundary markers: The prompt uses Gherkin structure and documentation reference files to constrain the agent's output.
Capability inventory: Subprocess execution for build tools (Maven/Gradle) and local file-system writes for feature files.
Sanitization: A mandatory review checkpoint (Step 5) requires the user to approve the full plan before the agent implements any files.
[DATA_EXFILTRATION]: The referenced library source code (ObjectSteps.java) contains utilities to access environment variables and system properties. These are intended for integration testing and are part of the framework's legitimate functionality.
The file-writing utility in the framework (output_in) includes a path traversal check to ensure files are only written within the project's resource directory.
[COMMAND_EXECUTION]: The skill uses standard local build wrappers (./mvnw, ./gradlew) to confirm that the project environment is stable and that tests are correctly discovered.
[SAFE]: The skill utilizes localized reference files (references/steps-*.md) as a 'source of truth' for legal step patterns, which prevents the agent from generating arbitrary or unsafe execution steps.

add-cucumber-tests