diff --git a/apps/web/src/styles/base.css b/apps/web/src/styles/base.css
index 908cd9e..d217318 100644
--- a/apps/web/src/styles/base.css
+++ b/apps/web/src/styles/base.css
@@ -1623,25 +1623,26 @@ ul[data-type='task_list'] li {
 }
 
 /* The checkbox is the top flex child; nudge it onto the first text line and
-   drop the leading/trailing margins of the item's paragraph so the text meets
-   the checkbox instead of dropping a line. Covers both DOM shapes: read-mode
-   `li > input` + `li > p`, and editor `li > label > input` + `li > div > p`.
-   Uses `:first-of-type`/`:last-of-type` (NOT `:first-child`): in the read-mode
-   markup the `<input>` is the first child, so the paragraph is never
-   `:first-child` and a `:first-child` reset would silently miss (issue #137). */
+   drop the leading/trailing margins of the item's paragraphs so the text meets
+   the checkbox instead of dropping a line. DESCENDANT selectors on purpose
+   (issue #137, second regression): the DOM depth differs per surface —
+   public/docToHtml renders `li > input` + `li > p`, but the editor/auth read
+   view (TipTap ReactNodeView) nests the paragraph TWO wrappers deep:
+   `li > div[data-node-view-content] > div > p`. Child-combinator chains keep
+   missing one of the shapes; matching any `p` inside the task item is the
+   robust form. `:first-of-type`/`:last-of-type` (NOT `:first-child`) because
+   the `<input>`/`<label>` precedes the paragraph in the read-mode markup. */
 ul[data-type='task_list'] li > input[type='checkbox'],
 ul[data-type='task_list'] li > label {
   flex: none;
   margin-top: 0.25em;
 }
 
-ul[data-type='task_list'] li > p:first-of-type,
-ul[data-type='task_list'] li > div > p:first-of-type {
+ul[data-type='task_list'] li p:first-of-type {
   margin-top: 0;
 }
 
-ul[data-type='task_list'] li > p:last-of-type,
-ul[data-type='task_list'] li > div > p:last-of-type {
+ul[data-type='task_list'] li p:last-of-type {
   margin-bottom: 0;
 }