The Agent Skills Directory

[PROMPT_INJECTION] (HIGH): High susceptibility to Indirect Prompt Injection during the template rendering process.
Ingestion points: scripts/render_template.py ingests data from CLI arguments (--var), JSON files (--vars-file), and environment variables (DOCKER_ARCH_*).
Boundary markers: None. Variables are directly interpolated into Dockerfile and Compose templates using the {{VAR}} syntax.
Capability inventory: The skill is designed to execute docker build, docker run, and docker compose up --build on the resulting files via scripts/smoke_test_container.sh and scripts/smoke_test_compose.sh.
Sanitization: Absent. There is no escaping or validation of variable values. An attacker-controlled variable (e.g., DEV_COMMAND) could inject arbitrary shell commands into the final container configuration.
[COMMAND_EXECUTION] (HIGH): The skill automates the execution of powerful container management commands using configurations generated from untrusted inputs.
Evidence: scripts/smoke_test_container.sh and scripts/smoke_test_compose.sh wrap docker and docker compose binaries, allowing the agent to launch containers and mount volumes (e.g., -v ./:/app in docker-compose.dev.yml) based on files created during the session.

docker-architect