Work in a vignette catalog

A catalog is a catalog.toml plus self-contained marimo notebooks that double as worked examples and importable helper modules. Resolve <skill-dir> to this installed skill and execute its Python scripts directly so their uv shebangs apply. <skill-dir>/scripts/catalog-session.py owns the whole session lifecycle: it starts, reuses, inspects, runs, and stops catalog kernels deterministically and survives shell wrappers. Never launch marimo through ad hoc nohup or backgrounding, and never reconstruct session state from pgrep, ss, curl, or /api/sessions by hand. If the helper fails, report its exact output before trying anything else.

Route the request first

• Open, run, or view a specific notebook, or "get this catalog running": follow the fast path below, then stop. Do not read the notebook contract, unrelated notebooks, or data references on this path.
• Change inputs in a notebook that is already live: reuse its session through marimo-pair code mode. Do not start a second kernel unless the current one is irrecoverable.
• Compose a new notebook or answer a data question: follow the compose workflow.
• Open-ended research: use the compose workflow for mechanics; research methodology stays with the user and repository, not this skill.

Fast path: open an exact notebook