The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill uses uvx to download and run specific versions of marimo (0.23.16) and ruff (0.16.2) from public package registries. These are well-known development tools and their use is consistent with the skill's purpose.
[COMMAND_EXECUTION]: The catalog-session.py script executes system commands via subprocess to manage marimo server processes and verify their identity using the ps utility. The validate-notebook.sh script executes marimo and ruff commands for code quality and execution checks. All command arguments are strictly controlled by the scripts.
[DYNAMIC_EXECUTION]: The catalog-session.py script performs dynamic execution by sending Python code snippets (e.g., for status reporting and running all cells) to the marimo kernel's API. The executed code is hardcoded within the script and is necessary for managing the notebook state.
[INDIRECT_PROMPT_INJECTION]: The skill handles untrusted data through ingestion of catalog.toml, AGENTS.md, and marimo notebook files. Ingestion points: Files are read in SKILL.md and processed by scripts catalog-session.py and validate-notebook.sh. Boundary markers: The skill uses marimo's built-in sandbox (--sandbox flag) during execution and validation. Capability inventory: Includes file read/write access and network operations to the local marimo server. Sanitization: Notebooks undergo static analysis with ruff and marimo check, and cold execution validation in validate-notebook.sh.

vignette-catalog-compose-notebook