The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill generates GDScript files for scene building and runtime behavior based on project requirements and executes them using the Godot engine's headless mode.
Evidence: The workflow described in SKILL.md and the templates provided in scene-generation.md and script-generation.md involve writing programmatic instructions to .gd files and executing them via godot --headless --script.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection via the Visual QA system.
Ingestion points: Task goals, requirements, and verification criteria are read from PLAN.md (or provided as arguments) and interpolated into the prompts for the external LLM in visual_qa.py.
Boundary markers: The prompt templates in static_prompt.md and dynamic_prompt.md lack explicit delimiters or instructions to ignore potential instructions embedded within the task context.
Capability inventory: The agent has the ability to execute shell commands (godot, ffmpeg, git) and write to the local filesystem.
Sanitization: No validation or sanitization is performed on the input from the plan before it is processed by the Visual QA agent.
[COMMAND_EXECUTION]: Extensive use of system-level commands to facilitate the development environment.
Evidence: Utilizes godot, ffmpeg, xvfb-run, glxinfo, and git. The capture.md file also includes a fallback timeout mechanism using Perl.
[EXTERNAL_DOWNLOADS]: Fetches documentation and project dependencies from external sources.
Evidence: The ensure_doc_api.sh script clones the official Godot engine repository from GitHub to bootstrap local API references. This targets a well-known and trusted organization.

godot-task