scema-tui

The Scematica Omni console: the agent loop as a full-screen terminal application. Perceive a world, tick the signals that ground a goal, watch branches compete, read back the records that were sealed and whether their commitments still hold.

$ cargo install scema-tui
$ scema-tui                    # the current directory
$ scema-tui /path/to/project
$ scema-tui --once             # one pass as plain text, pipeable
$ scema-tui --snapshot 120x40  # one frame as text
$ scema-tui --palette          # what colour this terminal can actually carry

If you have scema-cli, scema tui finds this binary next to itself and hands over.

Five tabs

  1·WORLD │ 2·SIMULATE │ 3·RECORDS │ 4·MEMORY │ 5·POLICY │        SCEMA OMNI scema-omni/0.1.0
  which branch wins, and does anything measured support it?                         ○ idle  .
────────────────────────────────────────────────────────────────────────────────────────────
┌ SIMULATION MATRIX  ·  NOT WRITTEN (would seal as c5ac7c5e) ───────────────────────────────┐
│ #   BRANCH                                GAIN   RISK   COST UNCERT REVERS  UTILITY  MEAS │
│▸  1 take: 11 marker(s) in `scema-tools`   0.22   0.40   0.10   0.00   0.70    0.125 ▰▰▰▰▰ │
│   2 give scema-cli tests                     —   0.40   0.10   0.00   0.70    0.093 ▰▰▰▰▱ │
│                                                                                           │
│ measured across the whole matrix: 35/35  (100%)   `—` = not measured; contributed nothing.│
└───────────────────────────────────────────────────────────────────────────────────────────┘

Black and violet, with soft blue for exactly one thing

Every other terminal surface in this repository has its own identity — the sniper dashboard is black and red, mesh-dashboard is indigo over slate, sdk-dashboard is the bond pipeline's green. Omni is black and violet, and that is not decoration: an operator with three of these open must be able to tell at a glance which one is making a claim about their money and which one about a decision record.

Violet specifically, because omni's whole subject is the boundary between measured and unmeasured, and that wants a palette where the interesting distinction is luminance first rather than hue first. A measured value glows; an unmeasured one recedes to a note in the same family. A red/green palette would put the emphasis on good-versus-bad, which is the wrong axis — a measured 0.00 risk and an unmeasured risk are not good and bad, they are known and unknown.

Azure appears in two places only: the branch that was chosen, and a commitment that verifies. Both are claims the agent made. An opportunity signal and a live provenance are observations and wear their own colours, because an accent that appears on every third row is not an accent, it is a second body colour.

Two rules, both mechanised

A renderer names a role, never a colour. theme.rs is the only file with a hex value in it; view.rs is the only file that maps a state to a role; render.rs places rectangles. Same split as alchem_link.theme and lib/mesh/view.ts::toneFor — a rule that encodes a claim about trust gets exactly one implementation.

Colour is decoration, never the message. NO_COLOR, a pipe, --no-color and a 16-colour terminal all produce the same text: — for unmeasured, ▸ for chosen, LIVE/STALE/ABSENT for provenance, RISK/OPP /EST? for signals, EXCLUDED for a forbidden branch. A test walks every role in Depth::Mono and fails if one carries neither a modifier nor a word — a role whose entire identity is a hue disappears on a monochrome terminal, and that is a bug in the role.

--palette prints what the terminal negotiated, so an operator on an unfamiliar one can confirm that "measured" and "unmeasured" are actually distinguishable there before trusting a screen full of them.

The coverage meter is a count, not a bar

▰▰▰▱▱. One cell per term. A proportional bar renders 2/5 and 4/10 identically, and the denominator is the number that matters — a utility computed on two terms out of five and one computed on five out of five are different claims. Above twelve terms it falls back to the 30/90 label rather than to a proportional bar, for the same reason.

An empty coverage draws ∅, never an empty meter. An empty container reads as "measured, and it is zero"; the true statement is "there was nothing here".

enter simulates. D decides.

Two keys, and a confirmation on the second.

simulate and decide compute exactly the same thing and differ only in whether they leave a trace. The only protection against a counterfactual later reading as a decision somebody made is that they are not the same keystroke — so the obvious button is the one that writes nothing, and sealing takes a capital letter and a yes.

Grounding is ticked, never inferred

The signal list is a multi-select, and what it selects is Goal::grounded_in. Nothing reads the goal text looking for a matching signal.

An earlier version of the CLI inferred grounding by keyword overlap and, on its first real run, grounded "add tests to the scema-cli crate" in a marker backlog in a different crate — because scema is a substring of every unit name in this workspace. The branch then inherited a measured expected gain from evidence that had nothing to do with it, which is exactly the laundering the simulator refuses to do on its own. A test pins that a goal naming a signal id verbatim still does not ground it.

--snapshot, and why a TUI needs one

$ scema-tui --snapshot 128x38

Draws one frame into an off-screen buffer and prints it as text. It exists because a TUI is otherwise untestable — its output is a terminal — so layout bugs are found by a human noticing a column is wrong, and ratatui's layout arithmetic underflows on small rectangles. The production failure mode is the console dying on somebody's 80-column terminal with a backtrace instead of a screen.

The crate's own tests draw every tab at four sizes and assert on the symbols, not the styles: everything here has to survive NO_COLOR anyway, so the text is what must be right, and a snapshot carrying escape sequences would churn on every palette tweak until nobody read it.

What it does not do

Act. The whole workspace ends at a decision and a record. A console that quietly gained a write path would invalidate the claim every other crate here makes about being safe to point at a live system.

Talk to a daemon. It holds no token and opens no socket; it drives the loop in-process against a local path. scema-omnid exists for the case where something remote needs to drive the loop, and folding that in would mean putting the whole pairing story on screen for no gain.

Why a separate crate

The same split mesh-dashboard / scematica-mesh makes in the bot workspace, and for the same reason: scema-agent is a library other people embed, and forcing ratatui and crossterm on a consumer that only wants the loop would be the mistake sdk-dashboard exists to avoid. The rendering lives here; the crates next door own the thinking.