Release v0.14.1 · SciQLop/SciQLop

v0.14.1 — 2026-10-06

%%vp --debug now prints a report in the cell output, not only on the debug panel: each input with its shape, unit, columns, NaN share, value range and time span, then the same for the result, then the checks. When the function fails, the report shows the error and its traceback. Before, a failing function printed nothing in the cell.
Tracebacks of a failing virtual product show your own code again. The line in your function that raised was hidden whenever SciQLop's code called it, which is always. In a development checkout, frames from installed libraries such as speasy were hidden too.

Saving right after adding events now keeps them in their catalog. The Save was committed before the step that puts a new event in its catalog, so after a restart the catalog came back empty and the events sat in "Orphan events". This hit the catalog browser's Save too when pressed right after an import, and could also lose the folder of a catalog created just before saving.
catalogs.persist(path) writes a catalog to disk from a notebook or an agent, like the catalog browser's Save. Writing events from Python only changed them in memory, and nothing in Python could save them.
panel.catalog_mode reads and sets a panel's catalog mode ("view", "jump" or "edit"), like the mode combo and Ctrl+Shift+M.
Reading or changing a catalog from a cell right after startup waits for its events to load. catalogs.get() returned an empty catalog while they were still loading, and add_events or save then worked from that empty list, so the catalog showed only the new events.
The event editor accepts ratings from 1 to 10, the range the catalog store uses. It stopped at 5, so editing an event rated higher by another tool lowered its rating to 5.

Agents are told to ask before acting when a request leaves open a choice that changes the result: the product, the time range, the frame, the exact definition of a quantity, or what to deliver. At most three questions, each with the answer they would pick, so a "yes" is enough. Clear requests go ahead without questions.
Agents get skills, published into each workspace's .claude/skills/ folder, where opencode and Claude Code find them: settling what the user wants before a larger task, finding the right data product, designing and testing virtual products, labeling events in catalogs, troubleshooting empty or slow plots, and writing SciQLop plugins. They load only when a task needs them, so they don't lengthen every session. Agents that don't load skills are pointed at the files from AGENTS.md.
sciqlop_set_time_range accepts ISO-8601 times as well as epoch seconds, like sciqlop_fetch and sciqlop_describe_product. It failed with "could not convert string to float" on the ISO times agents pass to the other tools.
The AGENTS.md guidance tells agents how to try a virtual product's maths before writing it: load the inputs with sciqlop_fetch from their product-tree paths, compute on them, then run the product once with --debug. An agent asked for |V_alpha − V_p| spent about ten tool calls hunting Speasy identifiers instead.

The JupyterLab panel no longer stays blank white until you reload it. If the page fails to load, or loads without Lab starting, it is loaded again on its own (jupyqt 0.6.6). This happened mostly on a busy start, such as the first one after an update.