The terminal app

poisson-prove with no verb, on a terminal, opens the terminal app: your runs, your comparisons and their verdicts, a sweep of the spec in the folder, and every condition and episode behind a verdict, reached by keys alone. It computes nothing of its own: every number, verdict and sentence on its screens is the one the commands print.

How it opens

poisson-prove
poisson-prove --store path/to/prove-runs

It opens on the store the commands use, ./prove-runs (or $POISSON_PROVE_STORE), or the one --store names, and s sweeps the folder’s prove.toml. It opens only when it can be used: stdin and stdout are both a terminal, CI is not set (set to anything, it is CI), and TERM is not dumb. Otherwise poisson-prove with no verb exits 2 and points at poisson-prove --help, so a script never finds itself inside an app. With NO_COLOR set to anything but empty, the colours go and bold, dim and reverse stay. The app needs a window of at least 80 by 24; in a smaller one it says the size it needs and waits.

It opens on Prove’s banner, the mark of the Poisson Labs tools with Prove’s name and version, while it reads the store, then shows home after a moment; any key shows home sooner, and q quits.

The screens

Every screen has the same frame: the tool and the screen’s name at the top with what it is about, the store at the top right, the body, and the keys at the bottom. The focused row is in reverse with › beside it. Home, at 80 by 24, with the example’s two runs and their saved comparison (a command too long for its line starts on the next one, whole, so it can be copied):

poisson-prove · home                                                  prove-runs
────────────────────────────────────────────────────────────────────────────────
  ok  2 runs and 1 saved comparison in prove-runs

  latest     mock-agent 2026-10-26 b (complete)
  compared   mock-agent 2026-10-26 a → b: trade
  next
poisson-prove report "mock-agent 2026-10-26 a" "mock-agent 2026-10-26 b" --x
task_id --y tool_fail_rate
  dashboard  index.html

  kind        name                         result     rows  created
› comparison  mock-agent 2026-10-26 a → b  trade            2026-10-26 14:53Z
  run         mock-agent 2026-10-26 b      complete  1,746  2026-10-26 14:46Z
  run         mock-agent 2026-10-26 a      complete  1,746  2026-10-26 14:46Z

────────────────────────────────────────────────────────────────────────────────
  ↑↓   enter   tab links   s sweep   c compare   r runs   y   ? help   q quit

home: the store at a glance, the latest run and comparison, the next command, and every run and saved comparison, newest first.
runs: every run with its number, status, backend and row counts, and the saved comparisons.
run: one run: its status, target, outcomes and their kinds, the design it was swept with, and each condition’s share of each outcome with its interval; g shows its event log.
sweep: the plan sweep prints, the same question, then the sweep’s progress; declining writes nothing.
compare: pick two runs, space for before then after, or l for previous and latest, and enter compares them; the comparison is saved, as diff --save saves it.
comparison: the verdict block diff prints, then each event’s count of conditions worse, safer, unchanged and could-not-tell, and links to its charts.
conditions (t): every condition and event with its shares before and after, its change, q, verdict and limit.
condition: one condition: each run’s share with its interval, the change and its interval, the test, the rise and fall it could have detected, and its episodes.
episodes: the episodes behind an event in one condition, with the snippet that reproduces one; e moves to the condition’s next event.
could not detect (d): the question the nothing-moved words answer, and each unchanged condition’s limits against the threshold.

Every screen but home, sweep and compare is also a page of the dashboard. A link on a screen (a chart, a page) opens it in your browser with Enter; the first link shown builds the dashboard if the store has none.

The keys

On every screen:

key	what it does
`↑` `↓`, `PgUp` `PgDn` `Home` `End`	move the focus
`Enter`	open the focused row’s screen, or its link in the browser
`Esc`	back, with the focus where it was
`Tab`	move the focus between the links and the table
`/`	filter the table’s rows by text; `Esc` clears it
`y`	copy the focused link’s `file://` address, or the focused command
`m`	the lines the window has no room for, when a screen says so
`?`	the keys, and the screens
`q`	quit, exit `0`; during a question, no
`Ctrl-C`	quit, exit `130`; in a sweep, stop it and keep what is done

On one screen: s sweep and c compare (home, runs), r runs (home), t conditions and d could not detect (comparison), v back to the verdict (conditions), e the next event (condition, episodes), g the event log (run), space and l (compare), and p (sweep, when the plan is refused for its size) to plan again with the replicates the refusal names.

A sweep stopped with Ctrl-C leaves the run partial, as the command does, and once the app has left the terminal it prints the same stopped line and the next: command that resumes it.

What it never does

It makes no network call, reads no key, and writes only what the commands write: a comparison it shows is saved through the same code diff --save uses, a sweep is the same sweep sweep --yes runs, and like every command it keeps its notes under ~/.cache/poisson-prove/. prove.toml is never written. If Textual, the library it is drawn with, cannot be imported, the app says it cannot start and exits 1; every verb still works.