From 56e9ca2dbefabe648037ab469462fde617858be0 Mon Sep 17 00:00:00 2001
From: Michael Grant <mgrant@grant.org>
Date: Fri, 21 Aug 2026 10:02:03 +0100
Subject: [PATCH] screen-redraw: erase only cells a pane owns, not its raw
 rectangle

image_redraw_start() was called with a pane's full nominal geometry
rather than the cells it actually owns in the current scene. When a
floating pane occluded part of that rectangle, the erase blanked the
floating pane's on-screen content, and the following text redraw
correctly skipped those cells (not owned by this pane), leaving them
blank with nothing to restore them.
---
 regress/image-floating-content-wipe.sh | 71 ++++++++++++++++++++++++++
 screen-redraw.c                        | 35 +++++++------
 2 files changed, 89 insertions(+), 17 deletions(-)
 create mode 100644 regress/image-floating-content-wipe.sh

diff --git a/regress/image-floating-content-wipe.sh b/regress/image-floating-content-wipe.sh
new file mode 100644
index 000000000..59891bca2
--- /dev/null
+++ b/regress/image-floating-content-wipe.sh
@@ -0,0 +1,71 @@
+#!/bin/sh
+
+# Regression test for a fixed bug: on a client detected as SIXEL-capable, a
+# pane-content redraw (PANE_REDRAW) erased the image backend's rectangle
+# using the redrawing pane's raw nominal geometry (wp->xoff/yoff/sx/sy)
+# instead of the cells it actually owns in the current scene. When a
+# floating pane occluded part of that rectangle, the erase blanked the
+# floating pane's on-screen area, and since the subsequent text redraw
+# correctly only refills cells the redrawing pane owns, nothing repainted
+# it back in - the floating pane's content stayed blank.
+#
+# Fixed in screen-redraw.c:redraw_draw_pane_lines() by erasing only the
+# cell ranges the pane actually owns per line (its REDRAW_SPAN_PANE spans),
+# not its raw rectangle. See tmux-sixel-erase-ignores-occlusion.md and
+# IMAGE-REDRAW-DISCUSSION.md for the full write-up.
+#
+# Confirmed by direct A/B test against this exact sequence: reliably wipes
+# the floating pane's content on the unfixed code, reliably leaves it
+# intact on the fixed code.
+
+PATH=/bin:/usr/bin
+TERM=screen
+LC_ALL=C.UTF-8
+export TERM LC_ALL
+
+[ -z "$TEST_TMUX" ] && TEST_TMUX=$(readlink -f ../tmux)
+TMUX="$TEST_TMUX -Lifcw$$ -f/dev/null"
+TMUX2="$TEST_TMUX -Lifcw-inner$$ -f/dev/null"
+$TMUX kill-server 2>/dev/null
+$TMUX2 kill-server 2>/dev/null
+
+TMP=$(mktemp)
+trap "$TMUX kill-server 2>/dev/null; $TMUX2 kill-server 2>/dev/null; rm -f $TMP" 0 1 15
+
+$TMUX2 new-session -d -x 60 -y 20 "sh -c 'i=0; while [ \$i -lt 30 ]; do printf \"TILE%02d\\n\" \$i; i=\$((i + 1)); done; exec sleep 100'" \
+	|| exit 1
+$TMUX2 new-pane -x20 -y6 -X8 -Y3 -T FLOATTITLE "sh -c 'printf FLOATCONTENT; exec sleep 100'" \
+	|| exit 1
+[ "$($TMUX2 display-message -p '#{image_support}')" = 0 ] && exit 0
+
+# The attaching client must be detected as SIXEL-capable - this bug does
+# not reproduce on a fallback or Kitty client.
+$TMUX2 set -as terminal-features ',*:sixel' || exit 1
+
+$TMUX new-session -d -x 60 -y 20 || exit 1
+$TMUX send-keys -l "$TMUX2 attach-session" || exit 1
+$TMUX send-keys Enter || exit 1
+sleep 1
+
+# Sanity check: the floating pane's content is visible before anything
+# disturbs it.
+$TMUX capturep -p >$TMP || exit 1
+grep -q 'FLOATCONTENT' $TMP || exit 1
+
+# Click the tiled pane (making it active), enter copy-mode scrolled up
+# (matching a PageUp binding), move up one line (matching an Up binding),
+# then exit copy-mode - the exit is what triggers window_pane_reset_mode(),
+# which is what actually caused the erase.
+$TMUX2 select-pane -t%0 || exit 1
+sleep 1
+$TMUX2 copy-mode -u -t%0 || exit 1
+sleep 1
+$TMUX2 send-keys -t%0 -X cursor-up || exit 1
+sleep 1
+$TMUX2 send-keys -t%0 -X cancel || exit 1
+sleep 1
+
+$TMUX capturep -p >$TMP || exit 1
+grep -q 'FLOATCONTENT' $TMP || exit 1
+
+exit 0
diff --git a/screen-redraw.c b/screen-redraw.c
index e262e1bd5..4728f240b 100644
--- a/screen-redraw.c
+++ b/screen-redraw.c
@@ -1597,25 +1597,26 @@ redraw_draw_pane_lines(struct redraw_draw_ctx *dctx, struct window_pane *wp,
 		bottom = scene->sy;
 
 #ifdef ENABLE_IMAGES
+	/*
+	 * Only erase cells this pane currently owns in the scene, not its raw
+	 * geometry - a floating pane may be occluding part of this pane's
+	 * rectangle, and erasing under it would leave those cells blank with
+	 * nothing to redraw them back in.
+	 */
 	if (flags & REDRAW_PANE) {
-		int	left, right, image_y;
-
-		left = wp->xoff - (int)scene->ox;
-		if (left < 0)
-			left = 0;
-		if (left > (int)scene->sx)
-			left = scene->sx;
-		right = wp->xoff + (int)wp->sx - (int)scene->ox;
-		if (right < 0)
-			right = 0;
-		if (right > (int)scene->sx)
-			right = scene->sx;
-		if (left < right && top < bottom) {
-			image_y = top;
+		for (y = top; y < bottom; y++) {
+			line = &scene->lines[y];
 			if (dctx->flags & REDRAW_STATUS_TOP)
-				image_y += dctx->status_lines;
-			image_redraw_start(&scene->c->tty, left, image_y,
-			    right - left, bottom - top);
+				cy = dctx->status_lines + y;
+			else
+				cy = y;
+			spans = &line->spans[REDRAW_SPAN_PANE];
+			TAILQ_FOREACH(span, spans, entry) {
+				if (span->data.p.wp == wp) {
+					image_redraw_start(&scene->c->tty,
+					    span->x, cy, span->width, 1);
+				}
+			}
 		}
 	}
 #endif