The terminal cockpit, tab by tab

triage tui is the terminal cockpit of a triage-pg project. It is built on lynkeus, a shared Textual shell that gives every PostgreSQL-backed project the same six screens, keys and theme; triage-pg supplies three adapters and three screens of its own. Nothing on any tab is computed in the TUI: every number is a SELECT over the views the dashboard and the CLI already read (headless-complete core, ADR-0012), and the one thing it writes — refreshing the leaderboard matview — runs the CLI as a subprocess and shows its exit code.

just tui                          # uv run triage tui
uv run triage tui --poll 10       # refresh the active tab every 10 s (0 = off)

The database is resolved exactly as for every other verb: --dbfile › database.yaml in the current directory › DATABASE_URL › PG*. Credentials never appear in the TUI, its state files or its logs.

Keys

1–5 open the standard tabs, 6–8 the triage ones, ? help. Everywhere: / focuses the current filter, ^p opens the command palette (tabs, refresh, theme, and every action), r refreshes, t flips dark/light, q quits. Each tab shows its own keys in the row above the footer.

Status

Database facts (size, relation counts, the alembic stamp, artifacts by status), row gauges (exact counts for the small tables, planner estimates for predictions and evaluations), the last five runs, runs per day as a sparkline, and pending work. A fortnight with no run reads none in 14 d rather than a flat line — zeros and a steady rate draw the same picture, so the empty case is said in words. Pending work is never a stored flag: a run still started after six hours, an artifact stuck in building with no live run, a leaderboard matview that is missing completed runs, an experiment without a run. triage status --json prints the same thing.

Runs

The list on the left (/ filters by name, id or state); the selected run on the right: its stage table — cohort, labels, matrices, models, predictions, evaluations, each as done / planned from run_progress, run_artifacts and the run’s plan, with notes such as 9 cached when a re-run reused another run’s artifacts — and its live log. A running run streams LISTEN run_progress notifications (ADR-0021); if LISTEN is unavailable the adapter polls the run_progress view every five seconds and says so in the log. A finished run replays its counts. Keys: l focus the log, k kill (triage-pg has no cancel — the key explains what to do instead), o open the run in the dashboard ($TRIAGE_DASHBOARD_URL, default the local just serve), y copy the detail as JSON.

Headless: triage runs list, triage runs show <prefix>, triage runs tail <prefix> — all with --json; runs tail --json prints one JSON object per event, so an agent can follow a run the way the screen does.

Data

Every schema and relation of the project database in a tree with planner row estimates; the selected relation’s size, vacuum/analyze times, columns with their comments, indexes with sizes, and three sample rows (enter shows twenty). 4 opens the relation in Query with a select * ready to run.

Query

A SQL editor (^enter or ^r runs, x runs explain analyze), a results grid and the saved-queries list. The nine triage views come pre-saved — leaderboard, run_progress, run_summary, experiment_summary, audition, audition_distances, model_group_summary, label_base_rate, cohort_profile — and s saves your own under ~/.local/state/triage/tui/<project>/. Every statement runs inside a read only transaction that is rolled back, so the screen cannot write. y copies the rows as JSON, e exports CSV. Headless: triage query "<sql>" --json.

Actions

The palette: every just recipe of the current directory (with its comment as description) and every verb of the triage CLI, introspected from the typer app — both parsed by lynkeus’s actions module, the same code every lynkeus consumer uses. enter runs the highlighted action as a subprocess; its stdout streams into the right-hand pane and the exit code becomes the state. An action is confirmed before it starts when its name or comment carries a destructive word (drop, downgrade, destroy, truncate, delete, rebuild, purge, reset, clean, prune), and so are gc and archive, which delete artifacts without saying so in their names.

A verb that cannot run without arguments carries them as a hint next to its description (triage run → CONFIG, triage predictlist → MODEL_ID AS_OF_DATE) and enter prompts for them before starting anything; escape or an empty answer cancels. Only required arguments prompt — just test *ARGS and triage gc still run bare on enter, and options are passed headlessly. y copies the command with its arguments. Headless: triage actions list --json, triage actions run "just test", triage actions run triage -- --version.

Experiments (6)

One row per experiment — a prediction problem, ADR-0022 — from experiment_summary: problem type, framing, runs, models, base rate, last run. The detail panel adds the latest run’s base rate and cohort size over its as-of dates as sparklines. enter opens the Runs tab filtered to this experiment; o opens it in the dashboard.

Leaderboard (7)

The triage.leaderboard matview for the selected experiment and one (metric, parameter) pair: one row per model group with mean, min, max, a sparkline over the test as-of dates and the last value, sorted by mean. m cycles through the pairs that were actually evaluated, x through the experiments. R refreshes the matview — through triage leaderboard --refresh, run as a subprocess, its exit code shown in the note line.

Audition (8)

audition_distances and audition for the same experiment and pair: per model group, the average value, the average distance from the best group at each as-of date, the maximum regret, the regret at the next date, and a sparkline of the distance over time — flat and low means steady. Best first. 4 from here opens the distances behind the sparklines in Query.

What stays in the dashboard

Anything richer than a sparkline: the model-group × split heatmap, the threshold curves, feature importances, crosstabs, the derivation graph. The o key on a run or an experiment takes you there.