Skip to content

The three panes

One screen, three columns, the same shape as the Scelo IDE's pipeline: soft data becomes tools becomes hard data, one way only.

Each pane has a chat box at the bottom. tab moves between them, and the focused pane is marked ◂ focus in its title and given extra rows at the expense of the other two.

SOFT · data

What was loaded, what was done to it, and what the agent made of it.

section shows
file(s) the file name, and the shape after cleaning
summary how many clean steps over how many passes, and which columns were dropped
the reading the agent's prose account of what the rows represent, which columns carry the analytical weight, and what might block analysis

The reading is the part worth arguing with. It is the model's interpretation rather than a measurement, and the chat box under it is how you correct it.

Row counts differ on purpose

The TOOLS pane says read file · 900 rows x 8 cols and SOFT says 900 rows × 7 cols. The first is what came off disk, the second is what survived cleaning: here an empty notes column was dropped, which the dropped: notes line records.

TOOLS · models

What the agent decided, and why.

The pipeline list is the run itself, one line per stage, each carrying its own result:

pipeline
✓ read file · 900 rows x 8 cols
✓ auto-clean · 1 step over 1 pass
✓ understand
✓ choose analysis · Value by segment
✓ run · `sum_insured` across `line` (5 segments)

Below it, analyses shows the menu as a node diagram: the dataset at the top, the chosen analysis marked and connected with a solid arrow, the runners-up marked · on dotted ones, and +5 more · /list for the tail.

Under that is the agent's stated rationale, in its own words. If you disagree with the choice, this is the pane whose chat box changes it, or use /run.

HARD · output

The result, three ways.

table is the analysis output. ctrl-t expands it when there are more rows than fit; press it again to shut it.

plot is the same numbers as a bar chart, sized to the pane. The header line names the plot and points at /charts for the full-screen gallery.

flow is a small node diagram of the provenance: which analysis produced this result, how many runs have happened, and the reminder that ctrl-e exports.

/graph off turns both diagrams off and gives the rows back to the tables and the chat. /graph on brings them back.

The chat boxes

Three bots, one per pane, each scoped to that pane's job. They exist to change what the agent decided, not to start work.

Anything you type goes to the pane's bot, except slash commands, which never reach the model.

When the window is too small

Below the minimum the app refuses rather than rendering something unreadable:

Terminal too small for the three panes.
100x30 — needs 140x24 side by side, or 56x32 stacked. Make it taller.

There are two viable layouts:

layout needs when
side by side 140 × 24 the default
stacked, soft above tools above hard 56 × 32 when the window is portrait, meaning narrower than twice its height in character cells

The portrait rule uses a cell aspect of 2, since terminal cells are about twice as tall as they are wide. A tall narrow window therefore gets a usable stacked layout instead of a refusal.