[moiri-gamboni]

Problem

A reply stored in YAML endmatter is written to the file and then never shown. The reviewer types a reply, it appears while the composer is open, and it is gone on the next load — so from the user's side the reply simply vanished.

This is reproducible with Roughdraft's own output, which is what makes it more than a formatting nit:

1. Open any document with an anchored comment {#c1}.
2. Click the reply button on that comment, type, save. Roughdraft writes c2: {body, by, at, re: c1} into the endmatter, with no inline marker — the shape the README and the RFM spec both document for replies.
3. Reload the page. The reply is gone from the rail, though it is still in the file.

The server agrees the reply is real: GET /api/review-index returns it with kind: "reply" and the correct parentId. Only the rail drops it.

Root cause

buildCommentThreadRailItems builds each rail group from group.commentIds, which comes from the anchor element's data-comment-ids. That attribute only lists comments carrying an inline marker.

Endmatter replies have no inline marker by design, so although addEndmatterFeedback parses them into the comments map with the right parentCommentId, they are filtered out of visibleComments before buildCommentThreads ever runs. The thread is built from the root alone and the reply is discarded.

This looks like an incomplete part of #91, whose description says the review rail "now handles unanchored comments, replies, and suggestion metadata in the endmatter-backed format".

Likely the same root cause as #87, which reports it as a bullet-list rendering problem: "I added a comment which was rendering fine, but then when the model made a reply to that comment, it stopped rendering as a comment in the sidebar." The list is incidental; the reply is the trigger.

Fix

Before building threads, pull each anchored root's descendants back in from the comments map, using the existing getCommentDescendantIds helper. Comments already present in the group are skipped, so inline replies are not duplicated, and a group with no anchored roots is unaffected.

Tests

packages/app/src/document-comments.test.ts (new): parses a document whose only reply lives in the endmatter and asserts the rail item contains both c1 and c2, plus a second case asserting no duplication when the reply is also anchored inline. The first test fails on main and passes with the fix.

pnpm test — 377 passed. pnpm lint clean.

Scope

App-side only, and independent of the open runWatch timeout PRs (#121, #122, #126, #139, #144) — no overlapping files.

[vercel]

@moiri-gamboni is attempting to deploy a commit to the Lex Team on Vercel.

A member of the Team first needs to authorize it.

[moiri-gamboni]

Pushed two follow-up commits after finding the same root cause on two more surfaces. The original commit fixed the comment rail only, which left the bug half-fixed.

Selection banner. activeCommentIds comes from getSelectionCommentIds, i.e. the marks under the selection, so it drops endmatter replies exactly like the rail did. Rather than repeat the filter, the descendant walk is now collectAnchoredThreadComments in document-comments.ts and both surfaces call it.

Replying to a reply did nothing. replyToComment anchors the new reply via addCommentIdsToAnchor(editor, commentId, ...), which returns null when the target has no inline mark (PageCard.tsx:296). For an endmatter reply that is always, so the function returned early: clicking Reply on a nested card produced no composer and no error. It now walks up to the nearest anchored ancestor for the anchor while keeping the clicked card as the parent, so the reply nests under it instead of being flattened onto the root.

Reproduction for the second one, using only the app's own output: reply to any anchored comment through the UI (Roughdraft writes re: into the endmatter), reload, then try to reply to that reply. Nothing happens.

New test in packages/app/test/page-card.test.tsx covers the composer opening from an unanchored reply; it fails on main. pnpm test 381 passed, pnpm lint clean.