diff --git a/tools/re-capture/bin/screenshot b/tools/re-capture/bin/screenshot
index c1eeb1bb..abccb1dc 100755
--- a/tools/re-capture/bin/screenshot
+++ b/tools/re-capture/bin/screenshot
@@ -25,9 +25,26 @@ RAW="$(mktemp /tmp/shot-raw-XXXXXX.png)"
 trap 'rm -f "$RAW"' EXIT
 "$REAL" "$RAW" >/dev/null || exit 1
 
-# "…": ("xenia_canary" "Xenia_canary")  1280x745+0+0  +0+0
-geo=$(xwininfo -root -children 2>/dev/null \
-      | grep '"xenia_canary"' | grep -oE '[0-9]+x[0-9]+\+-?[0-9]+\+-?[0-9]+' | head -1)
+# "…": ("xenia_canary" "Xenia_canary")  1280x745+1+20  +1+45
+#
+# TWO traps here, both measured (2026-08-18):
+#   * `-children` only lists DIRECT children of the root, and openbox reparents
+#     the game window, so the real one is not there at all;
+#   * this app owns more than one window with the class "xenia_canary" — an
+#     off-screen 10x10 helper as well as the 1280x745 game window — so `head -1`
+#     picked the 10x10 one and every grab came back a 10-pixel sliver. Nothing
+#     errored: the pixel oracles simply classified the sliver, and a whole
+#     session's screen ids were noise.
+# So walk the full tree and take the LARGEST xenia window, by area. Use the
+# trailing ABSOLUTE geometry (the second +X+Y), because a reparented window's
+# own +X+Y is relative to its frame.
+geo=$(xwininfo -root -tree 2>/dev/null \
+      | grep '"xenia_canary"' \
+      | grep -oE '[0-9]+x[0-9]+\+-?[0-9]+\+-?[0-9]+[[:space:]]+\+-?[0-9]+\+-?[0-9]+' \
+      | awk '{ split($1, g, /[x+]/); a = g[1] * g[2];
+               if (a > best) { best = a; split($2, p, /\+/);
+                               out = g[1] "x" g[2] "+" p[2] "+" p[3] } }
+             END { if (best) print out }')
 if [ -z "$geo" ]; then           # no window found — hand back the raw grab
   cp "$RAW" "$OUT"; echo "$OUT"; exit 0
 fi