Limits¶

This is the honest list of what Scelo TUI does not do, kept here rather than left for you to discover.

Scope¶

Eight analyses. Scelo's real catalog is around thirty. The reasoning is that model fits belong where their diagnostics fit, which is the IDE.
No swarm. The council and population simulation are IDE features.
One dataset at a time.

No session persistence. The model choice and API keys survive; the loaded dataset, the chat history and the analysis you switched to do not. Exporting the .sce at the end is the workaround, and arguably the feature.
Chat history is per-pane and per-session, and is not written anywhere.
Reopening the picker re-runs the pipeline rather than keeping the loaded state. A new model reading the data itself is defensible, but this is what falls out of remounting rather than a decision that was taken.

The .sce export carries the session one way. The TUI does not open .sce files.
Inside the Scelo IDE the export lands in the workspace, but the session does not auto-load. That needs an IDE-side terminal listener: read the escape sequence, parse the .sce, restore. The detection contract, SCELO_IDE and SCELO_IDE_WORKSPACE, already ships in the IDE, which is the first half.
RStudio cannot be told to open a file in its editor. No CLI exists for it (rstudio#1850, #14226), so files in your open project plus a ready source() line is the whole of what can be automated there.

No scrollback in the panes. You see the tail.
Needs 140 columns, or a portrait window at least 56 × 32 for the stacked layout. Below both it refuses rather than rendering something unreadable.
Clicking picks a pane by column, not by the pane's exact rectangle, so a click below the panes still moves focus.
The header's ⇩ export click target assumes the app is drawn from the terminal's top row. That holds everywhere tested, but it is the layout falling our way rather than a guarantee.

The Excel data sheet caps at 10,000 rows and says so on the summary sheet. data.csv always has everything.
TUI runs map to the catalog's descriptive model and the runs record in the .sce stays empty. The IDE wants a numeric KPI headline for a result card, and inventing one to fill the slot would put a fake number on it.

No release binary. It installs from source, and needs the scelo repo checked out as a sibling because @scelo/core is consumed over a file: path. That is deliberate: profiling, typing, coercion and filtering have one definition shared with the IDE, so the two cannot drift.