The Agent Skills Directory

[COMMAND_EXECUTION]: The skill orchestrates complex tasks by executing the run.sh entry points of other local skills using the subprocess.run function.
Evidence is present in converge.py and pipeline.py, where the orchestrator calls scripts for create-music, review-music, prompt-lab, memory, and create-stems.
Command arguments are dynamically built from project parameters, file paths, and user-provided configuration values.
[PROMPT_INJECTION]: The skill's pipeline is exposed to indirect prompt injection risks because it ingests and processes data from external sources and includes it in prompts for downstream LLM tasks.
Ingestion points: In pipeline.py, lore fragments are fetched from the memory skill and musical reference data is retrieved from the consume-music skill.
Boundary markers: The skill does not employ specific delimiters or "ignore instructions" guardrails when interpolating this external data into the prompts sent to other tools like create-story and prompt-lab.
Capability inventory: The skill possesses broad capabilities, including the ability to execute shell commands, read and write to the local filesystem, and communicate with local network services for UI updates.
Sanitization: External inputs are truncated to fixed character counts (e.g., 500 or 300 characters) but are not sanitized or escaped to neutralize potential instruction-based injection attacks.

music-lab