Vignette Catalog Compose Notebook

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

1. Read the repository's AGENTS.md and catalog.toml. Read auth requirements, documented browser access, and [getting_started].first_notebook; skip caveats and notebook code that launching does not require.

2. Choose the user-facing host, then run one command; it reuses a healthy session for that exact notebook or starts one, ensures the cells ran, and prints the result:

   <skill-dir>/scripts/catalog-session.py open [--host <documented-address>] [notebooks/<name>.py]
   

   Omitting the notebook opens [getting_started].first_notebook. For interactive launches, select the access route documented by the user, system, or repository without requiring the user to repeat it. Do not assume the user's browser runs on the execution machine. For direct access, bind to the documented Tailscale or LAN address or hostname; use --url-host <hostname> when the printed hostname should differ from the bind address. The helper disables marimo's access token, so use --host 0.0.0.0 --url-host <hostname> only when the documented environment calls for all-interface binding, such as simultaneous localhost and remote access. Omit host options for local-only access or when no direct route is known, and identify the resulting URL as local to the execution machine; use SSH forwarding when the requested remote access requires a tunnel.

   Host options affect only a newly started server. If open reuses a session whose reported URL is not reachable at the selected hostname, report the mismatch and do not claim that the requested launch succeeded. Preserve that session's live work; replace it only when the user's existing instructions authorize a restart, or after obtaining that authorization.

3. Report url= and session= and the cells= line verbatim. cell_error= lines mean the notebook ran but has failing cells: report them; the session stays live for iteration. Answer any follow-up about session state with <skill-dir>/scripts/catalog-session.py status --cells, rerun cells with run <port>, and stop a session you own with stop <port>.

Compose workflow

Use marimo-notebook for general notebook authoring and marimo-pair for every live-kernel action. Repository instructions, catalog.toml, and this skill's notebook contract override generic marimo-notebook advice when they are more specific. If either project skill is absent, stop and give the user the repository's documented install command rather than installing it implicitly.

1. Read the repository's AGENTS.md, catalog.toml, and any path named by [data].caveats. Use the manifest to find likely notebooks, then read their actual code and docstrings. The manifest is a curated routing table, not necessarily an inventory of every notebook or helper.

2. Take the shortest path that answers the question. Change inputs in an existing notebook when its workflow already fits. Otherwise create a composed notebook and import the closest helpers instead of recreating their requests, parsing, joins, or plots. Read references/notebook-contract.md when authoring or changing a notebook.

3. For a new composition, create its smallest valid marimo scaffold, including the initial setup imports. Open the exact notebook you are adapting or composing with catalog-session.py open notebooks/<name>.py, adding host options according to the fast-path access rule. Use the reported URL and session id with marimo-pair; the active runtime is authoritative until you validate the saved file.

4. Work through the live kernel. Use marimo-pair code mode for durable cell edits, run each changed cell, and inspect the returned tables and rendered figures before interpreting them. Start remote or REST exploration with a bounded query, then widen deliberately. Keep every dependency of the answer in the notebook, not only in scratch state.

5. Prove the saved notebook from a clean state:

   bash <skill-dir>/scripts/validate-notebook.sh notebooks/<name>.py
   

   The validator runs pinned marimo checks, a stable Ruff rule set, cold execution, and an explicit scan for failed cells because marimo can report failure while exiting zero. It restores source and snapshots by default; pass --write only when the catalog policy calls for formatting and a refreshed snapshot. Follow the catalog's own policy for generated session snapshots and analysis outputs.

6. Report what ran, what you inspected, the answer and its limits, and the live URL if the session remains useful. Stop a session you no longer need with <skill-dir>/scripts/catalog-session.py stop <port>. Promote a composed notebook into catalog.toml only when the user wants it curated as a reusable vignette.

When asked to verify the whole catalog, enumerate the actual notebook files rather than assuming catalog.toml is exhaustive, and validate each in a disposable worktree or archive if the catalog tracks generated snapshots.