TUI

mecha-graph tui — keystroke-speed surfaces for the jobs a one-shot CLI is bad at: review-queue triage, merge review, a search REPL with provenance drill-down, quick capture, an entity browser with fact supersede, a task board, and health stats.

Global keys

Tab / Shift-Tab   cycle screens
1-7               jump straight to a screen (when nothing is being typed)
Esc               back out / empty the input buffer
q                 quit (when nothing is being typed; Esc-then-q works from anywhere)
Ctrl-Q            quit, even mid-typing

The seven screens

1 · Review

The pending fact-candidate queue, opened on clusters — candidates grouped by (proposer, predicate), each showing its class's acceptance history ("✓ 41 / ✗ 3, 93% accepted on this class"), which is exactly the evidence the autonomy ladder promotes on.

j/k        move            a      accept the cluster
Enter      inspect items   A      accept, creating new topic entities
p          proposer view   r/R    reject (R records a reason for the whole cluster)
s          shadow view     c / Esc  back

s opens the surfaced-verdict view (review-on-use): live shadow facts that are about to matter — contradicting a reviewed fact (⚡), actually served in a context pack, or spot-checked by a sampled class — at most ten at a time. y confirms (tier → reviewed), r refutes as never true, R refutes with a typed reason (it feeds rejection memory). Since extraction mints shadow facts instead of queueing, this view is where most review now happens; the candidate queue keeps only what cannot become a fact without a human (commitments, flags, unresolvable subjects).

g on a cluster opens its semantic groups — the class's near-repeats at the shared 0.83 floor, one verdict per group: a accepts, r/R rejects, and in every case the leader is your verdict while members cascade machine-labeled (one keystroke is one human verdict). Within-class only, by measurement: same-class pairs carried the same human verdict ~89% of the time; cross-class only ~63%, so crossing stays off this surface (mecha-graph calibrate-groups reproduces the numbers).

p rolls the queue up one level further — by proposing mechanism, with each one's human accept rate and how much evidence it rests on (unjudged / thin / some / solid). The pipeline's own dedup rejections are shown beside the rate as "auto-dropped", never inside it, and a mechanism nobody has judged shows a dash rather than 0%: "never reviewed" and "always rejected" are opposite findings.

Inside a cluster: a/r/Space per item, e edits a candidate before accepting (↑/↓ move field · Enter save · →/Ctrl-F complete an entity name). Commitments materialize tasks — Enter reviews them individually.

2 · Merge

Duplicate-entity candidates, side by side.

j/k   move        ←/→   swap which side is kept
m     merge       s     skip

3 · Search

The REPL. Type to search; results carry provenance you can drill into.

#tag        filter to a tag          @source     browse one source
Enter       open the hit             Ctrl-E      semantic (embedding) search
Ctrl-P      toggle private tiers     /           lookup from a fact pane

An opened episode offers: t tag · n note · m link an entity · p cycle sensitivity tier · e edit · d delete — and a deleted episode says so: Ctrl-Z (or mecha-graph undo) restores. Source-owned episodes are not editable, and the screen says which.

4 · Capture

A quick note that saves as an episode with entities auto-linked (Enter); Ctrl-T switches to capturing a fact instead.

5 · Entity

Lookup with suggestions (type · ↑/↓ pick · Enter opens). An entity shows its facts and timeline:

j/k    move          s      supersede a fact (bi-temporal close + replacement)
Enter  follow        h/l    switch pane
/      lookup        —      timeline entries open their episode
y / u  confirm / refute an unreviewed (◌) fact in place

Facts nobody has vetted are marked ◌ — opening an entity is itself a review trigger, so the verdict happens where the context is: y stands behind the fact (tier → reviewed), u says it was never true. Reviewed facts answer to s (supersede), not u: stopped-being-true and never-was-true are different retractions.

6 · Tasks

The GTD board.

a      add            e      edit schedule
n/i/w/s/d/x           set status (next/inbox/waiting/scheduled/done/dropped)
Space  cycle status   z      show closed        Enter   page

7 · Stats

The same health numbers as mecha-graph stats, live.

Where it fits

The TUI is the human half of the autonomy ladder: precheck drains the mechanical part of the queue first, the Review screen's cluster verdicts are what the per-class acceptance history is built from, and everything it writes goes through the same store functions the CLI uses — there is nothing the TUI can do that the CLI cannot, only faster hands.