@vincemakes/kiso-tui 0.20.3 → 0.20.4

package/dist/compositor.js

@@ -47,7 +47,7 @@ import { MENU_ITEMS, displayWidth } from "./editor.js";
 import { leadWidth } from "./width.js"; // W23: the ONE width authority (the editor, #inputRow, and editCol share it)
 // KC3.5: the panel-slot reads come from the DISPATCHERS — one source
 // for four reads, so an ask can never render half as an approval.
-import { panelAffordanceOf, panelFrameOf, panelLeadOf, panelRowsOf, panelStatusOf } from "./ask-panel.js";
+import { panelFrameOf, panelLeadOf, panelRowsOf, panelStatusOf } from "./ask-panel.js";
 import { MOUSE_OFF } from "./editor.js";
 import { atPanelRows, bandHeader } from "./at-picker.js";
 // TUI2-R2 ②: the session picker's rows — the band's third occupant.
@@ -1820,8 +1820,25 @@ export class Body {
      *  now). */
     #statusSource() {
         const panel = this.#panelState?.() ?? null;
+        // DC-38: the panel's STATUS replaces the CLI's painting status —
+        // that half of W21 stands. The HINT does not come with it, because
+        // the panel block already ends in its own affordance row and has
+        // since TUI2-R1.5 ("the affordance — the phase's key hint, ONE
+        // row"). Both sites read `panelAffordance`, neither knew about the
+        // other, and an approval at W≥100 printed the same 48-cell sentence
+        // twice, six rows apart.
+        //
+        // The block's copy is the one to keep, on two grounds. It sits next
+        // to the options it names. And it never disappears: the status row
+        // drops its hint when the left text plus the hint will not fit, so
+        // the duplicate was width-dependent — one copy on an ask (long left
+        // text), two on an approval (`⏸ run paused`, twelve cells) — which
+        // is why every fixed-width gate was blind to it. Each flavour's
+        // block carries its own row (approval-panel.ts pushes it for the
+        // approval and the pick, ask-panel.ts for the ask), so nothing is
+        // lost anywhere.
         if (panel !== null)
-            return { status: panelStatusOf(panel), hint: panelAffordanceOf(panel) };
+            return { status: panelStatusOf(panel), hint: undefined };
         // W22: while turns wait in the queue, the right hint shows the
         // count — the chips below carry the lines themselves.
         const queued = this.#queueState?.().length ?? 0;
@@ -2702,24 +2719,37 @@ export class Body {
             // the steady state for the session's whole life. It repairs a
             // killed predecessor straight into the same state. The visible
             // cursor comes back exactly once, in editor.exit().
-            out.push("\x1b[r\x1b[?69l\x1b[?7h\x1b[?25l");
-            // REL-0152-D20 — the mechanism is understood and the obvious fix
-            // is NOT taken here. See the finding.
+            // DC-40 — H line feeds BEFORE the reset, from wherever the shell
+            // left the cursor. The first frame is the full-redraw path and
+            // addresses rows 1..H absolutely; at launch those rows are the
+            // shell's (its prompt, the launch command, the tail of whatever
+            // ran before), and the frame painted over them — gone, not in
+            // the scrollback (REL-0152-D20 established the mechanism; 37/60
+            // shell lines survived on Apple Terminal, every line ON SCREEN
+            // lost). From cursor row r, H feeds move H−r rows and then
+            // scroll exactly r: the shell's rows 1..r enter the scrollback
+            // as CONTENT (at most one blank row — the cursor's own line),
+            // the screen is blank, and the model's assumption "row 0 is the
+            // terminal's row 1" is true by construction. The count does not
+            // depend on r, so nothing is asked of the terminal.
             //
-            // The first frame addresses rows 1..H absolutely and draws over
-            // whatever the terminal was showing. Scrolling a screenful away
-            // first would fix that, and it was built and measured: it pushes
-            // up to H BLANK rows into the scrollback, and TUI2-R2pre's
-            // blank-share gate went from 14/43 to 29/43 — past the "a
-            // healthy session's scrollback is mostly content" invariant that
-            // gate exists to hold. Trading a symptom with a five-second
-            // user-side setting for a broken invariant is a worse deal.
+            // THE ORDER IS THE FIX. `ESC[r` (DECSTBM) HOMES THE CURSOR to
+            // row 1 — VT100 semantics, honoured by Apple Terminal, xterm,
+            // xterm.js and tmux alike. Feeds emitted AFTER the reset start
+            // from row 1, move H−1 rows and scroll ONE: measured on Apple
+            // Terminal, 1/20 shell lines survived with an otherwise
+            // byte-identical frame. Feeds BEFORE the reset: 20/20. The house
+            // emulators did not model the homing and passed the wrong order;
+            // VtScrollback does now, and the DC-40 gate is red on it.
             //
-            // The correct version scrolls only as far as the terminal's
-            // content actually reaches, which needs a cursor-position query
-            // at boot — its own round, with its own risk (this file already
-            // declined a boot-time round-trip once, for racing the editor
-            // for stdin).
+            // The reset still precedes the FRAME (REL-0152-D19: a frame
+            // drawn into an inherited sub-region is the defect); only the
+            // feeds run under whatever region the shell left. A region a
+            // killed foreign program left set makes the feeds scroll that
+            // region alone and the frame paint over the rows outside it —
+            // which is what every frame did before this fix, never worse.
+            out.push("\n".repeat(H));
+            out.push("\x1b[r\x1b[?69l\x1b[?7h\x1b[?25l");
         }
         out.push("\x1b[?7l");
         // REL-0161: ?25l is the STEADY state now, not a frame bracket —

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@vincemakes/kiso-tui",
-	"version": "0.20.3",
+	"version": "0.20.4",
 	"description": "kiso tui — the pure terminal layer (cell renderer, dock, raw editor, diff, palette). Zero runtime dependencies: input is data, output is bytes.",
 	"type": "module",
 	"license": "MIT",
@@ -35,6 +35,6 @@
 	},
 	"homepage": "https://github.com/vincemakes/kiso/tree/main/packages/tui#readme",
 	"dependencies": {
-		"@vincemakes/kiso-tui-cells": "0.20.3"
+		"@vincemakes/kiso-tui-cells": "0.20.4"
 	}
 }