fix(web): file view comments stay inline after switching threads#33472

Problem

In the file view (right panel → Files → a file), a line comment shows inline under the commented lines, but EditableFileSurface kept those inline comments only in local state that started empty. After switching to another thread and back, reloading, or closing and reopening the file tab, the inline comment was gone while its composer chip (math.ts L2) and the draft's reviewComments still had it. When two threads had the same file open, the still-mounted editor also showed the previous thread's comments, and an edit or Undo there could copy them into the other thread's draft.

Reproduced on main (b8c2cfad6b) in the web dev server with a fresh database: comment on math.ts L2 in thread A, switch to thread B and back → inline comment missing, chip still present. Same after a reload and after closing/reopening the file tab.

Change

• EditableFileSurface initialises its inline annotations from the composer draft's reviewComments for this file (sectionId === "file:<path>", line range from startIndex/endIndex), the same source the diff view restores from.
• The composer draft decides which comments exist; the editor decides where they sit. The surface subscribes to the draft and drops an inline comment as soon as its id leaves the draft (chip removed in the composer, or the draft cleared by a send). An edit only writes back moved lines for comments the draft still holds, so an edit or Undo can no longer bring a removed comment back into the draft.
• The surface's key now includes the composer target, so a thread change (or starting/ending the edit of a queued message, which swaps the composer target) remounts it with that target's comments and a fresh edit history; the previous target's comments can no longer be shown or replayed by Undo into the new one.
• The line-link reveal resets its handled request when the file view's scroll container changes, so a remounted view scrolls back to the linked line instead of staying at the top.
• The "append entry under its end line" grouping moved into fileCommentAnnotations.ts (appendFileCommentEntry) and is shared by new and restored comments.

What else changes for a user: a thread switch resets the file editor's Undo history and scroll position (to the top, or to the linked line). Removing a comment's chip, or sending the message, now also removes the comment from the file view; on main it stayed inline and the next edit wrote it back into the draft. Deleting a commented line hides the comment inline and keeps it in the draft with its original quote, as on main; Undo brings it back inline.

Scope and approval

Fixes pingdotgg/t3code#33369

Verification

• vp test run src/components/files/FilePreviewPanel.browser.test.tsx src/components/files/FilePreviewPanel.test.ts (apps/web): 43/43 pass at a55fcc906c; vp run check:changed passes. New tests, each failing before the change it covers:
  • "shows a comment already in the composer draft when the file view mounts" (fails on main: expected [] to deeply equal [['annotation-2', …]]);
  • "shows only the current thread's comments when the thread changes under a mounted file view";
  • "does not replay another thread's comment when undoing after a thread change" (edit in A, switch to B, Undo: without the remount, B's draft received A's comment);
  • "keeps a linked line revealed when the thread changes under the file view" (without the reveal reset the remounted view stays at the top);
  • a comment removed from the composer is no longer shown inline, and the next edit does not add it back to the draft;
  • "deleting a commented line neither moves nor rewrites its comment, and Undo restores it";
  • "undoing an earlier edit keeps a comment added after it";
  • unit tests for grouping, the per-file filter (diff-section comments excluded), and the prune helper.
• Three adversarial review rounds on this head's predecessors; their findings (removed comments returning on edit, a deleted line re-anchoring and re-quoting its comment, Undo losing comments) are fixed by the prune-only model above and covered by the last three tests.
• Real web app, before (main) and after (this branch), fresh database per side, synthetic fixture repo: thread switch, reload, and file-tab close/reopen keep the inline comment after the fix (rows M1–M3 below). These captures predate the @pierre/diffs 1.7 upgrade on main (d655e2cca6) and the composer-sync commits above; they are not yet recaptured at this head.
• Not captured: line-link reveal after a remount, composer chip removal with the file open, and line deletion followed by Undo (covered by the browser tests above).
• Not checked: the Electron desktop shell (same web code path); mobile has no file-view comments.

Evidence

Revision note: every capture below was recorded at 0148848580 or 34e0e58371, before the @pierre/diffs 1.7 upgrade from main (d655e2cca6), the reveal-after-remount commit 309f4c73a5, and the composer-sync commits 70e5c05904–dabd2a3f5e. They show the restoration behaviour at those revisions; current-head captures are pending.

Behavior beyond stored-comment restoration: switching threads now resets the file editor's Undo history. In O3, Undo after B→A changes nothing: the saved first blank line and A's restored L3 comment remain. This is the intended remount behavior and should be included in review scope.

Capture artifacts: the continuous after recording retains a cached relative-age label after reload; the normal-reload supplement supplies the comparable M2/M3 stills. Main O3 recordings have different synthetic sidebar contents; normalized O3 stills hide that sidebar. Reused O1/O2 retain their documented three 1/255 project-icon pixels.

For #33369: before b8c2cfad6b; recaptured after 0148848580. 4/9 matrix rows recaptured at the new head (M1, M2, M3, O3), plus the 1440×1000 light key still; 5/9 optional rows reused at 34e0e58371. All 9 rows retain paired before/after evidence. The before originals and their public links are unchanged. No product source changes, component harness, or real database seed.

Environment for every recapture: T3 Browser preview; headless Chromium 156.0.0.0, Linux x86_64; no device emulation; DPR 2; 100% zoom; default font scaling, 16px root. M1–M3 and O3 use 1600×1000 CSS px dark; the key still uses 1440×1000 light. File panel 540 CSS px; source scroll at top. Native video 3200×2000. Full-surface PNG exports are tool-scaled to 1280×800 or 1280×889, with no crop. Server and page wall clocks freeze at 2026-11-09T16:00:00Z; timers and monotonic clocks remain native. The preview tool has no reduced-motion override, so the same capture CSS suppresses motion, transitions, and caret. Composer fixture: Replay / Multi turn. The new run uses an incognito browser profile with the same defaults.

Each retained after still matches a second screenshot 500ms later pixel for pixel, after fonts, idle resources, and thread synchronization. All retained captures, diffs, sheets, and recording milestone frames were opened and reviewed. Diff0: zero-threshold absolute RGB difference amplified ×32. DiffT16: maximum RGB channel difference >16/255, magenta on dimmed before context. No masks, alignment, cropping, or normalization.

The action driver and seed are reused. The driver now waits for navigation/thread synchronization before reapplying the clock freeze and refuses stills with loading/syncing status. A preliminary capture was rejected before upload because of sidebar/synchronization mismatches. The full recording was repeated after recreating synthetic A/B threads until their equal-time ID ordering matched the unchanged before frames.

M1 — A→B empty→B populated→A→B→A; 1600×1000 dark

Approximate times in both recordings: 0–3s select L2, type and submit A; 12–24s switch to initially empty B and open math.ts; 25–29s submit B at L6 and return to A; 49–53s B→A again. Inspect the final A frame around 53–56s. Baseline shows B’s L6 block in A while A’s L2 chip remains. Fixed shows only A’s L2 block; B has only its own L6 block. Initial empty B has no A block/chip. Editor instance identity was not instrumented; these are real-app navigation captures.

M2 — normal reload; 1600×1000 dark

At 66s the unchanged Ctrl+R attempt does not navigate this headless browser. At 84.5s the current page is reloaded by navigating to its same URL; the driver waits for synchronized messages before reapplying the clock freeze. Inspect 103–112s; the actual close action follows at 113s. Baseline loses the inline block but retains its chip; the new head restores exactly one A L2 block and chip without resubmission.

Normal-reload supplement (same A state, no resubmission): after-M2-M3-supplement.mp4
0148848580

0–6s normal location.reload and synchronization; 10–18s restored L2 / M2; 21–25s actual file-tab close/reopen; 25–27s M3. The retained M2/M3 stills come from this supplement.

M3 — actual math.ts tab close/reopen; 1600×1000 dark

Before: around 107s; after: at 113s click Close math.ts; the right panel closes and the chip remains. Around 120–123s toggle the right panel, choose Files and math.ts. Inspect 124–126s. Baseline has no inline block; fixed restores one A L2 block. This closes the actual file tab, not only panel visibility.

Normal-reload supplement (same A state, no resubmission): after-M2-M3-supplement.mp4
0148848580

0–6s normal location.reload and synchronization; 10–18s restored L2 / M2; 21–25s actual file-tab close/reopen; 25–27s M3. The retained M2/M3 stills come from this supplement.

Key moment — A after B→A, math.ts; 1440×1000 light

Supplemental settled light still after B→A with the same stored A L2 comment. No new comment is submitted. Required interaction proof is in the dark recordings above.

O1 — grouped ranges and other-file filtering; 1600×1000 dark

Captured at 34e0e58371, reused unchanged: the production diff to 0148848580 only changes thread-switch handling (remount); mount restoration is unchanged.

Approximately 0–16s select L2–L3 and submit A: range, add A: same end at L3 and A: multiplication at L10; 17–21s open other.ts and submit Other file only at L2, then return to math.ts; 21–35s close/reopen math.ts. Inspect 36–39s. Fixed restores two ordered entries at L3 and one at L10, without the other-file text. Baseline restores none, although all four chips remain. other.ts is shown with its own submitted comment during the sequence. The optional diff-section variant is unverified.

O2 — restored delete/cancel/keyboard submit; 1600×1000 dark — baseline partial

Captured at 34e0e58371, reused unchanged: the production diff to 0148848580 only changes thread-switch handling (remount); mount restoration is unchanged.

Around 63–66s delete the first restored range, enter whitespace at L8, replace it with Canceled draft and Escape, then submit Temporary using Ctrl+Enter. Around 89–93s delete remaining blocks and close/reopen math.ts; inspect 95–98s. Fixed keeps the sibling after the first deletion, adds no canceled chip, and ends without empty gaps or resurrected math blocks; only other.ts L2 remains. Baseline restored controls are absent. Temporary is successfully submitted and deleted around 89s, but the original math comments cannot be deleted through their missing inline controls. Baseline delete-to-empty is unverified for those stored comments, whose chips remain.

O3 — edit L2→L3, save and restore; 1600×1000 dark

Main paired recording: 0–3s add A's L2 comment and insert one blank first line in the actual file editor; allow save to settle; 32–35s select B then return to A with sidebar clicks. At 38s focus the source and press Ctrl+Z. On the new head Undo changes nothing after returning: the blank first line remains, and the restored comment and chip stay at L3. No B comment leaks into A. Baseline loses the restored block. The existing baseline recording also shows no visible Undo change; preserving old Undo history is not claimed.

The normalized stills use the same fresh Remap proof setup with the main sidebar hidden. Supplemental paired recordings cover the same L2→L3 edit and B→A full-page navigation; the new supplemental recording also presses Undo after returning. Before supplement: edit 0–4s, full B→A navigation 5–9s. After supplement: edit at the start, full B→A navigation during the loading interval, then inspect 26–29s for restored L3 and Undo. Actual tool completion times are 2s edit, 14.5s B navigation dispatch, 20s return dispatch and 26.4s Undo; recording idle gaps vary, so the contact sheet identifies the visible milestones. These are real-app inputs, with no editor-state injection.

Supplemental paired remap/navigation recordings:

before-O3-supplement.mp4

after-O3-supplement.mp4

O4 — theme remount and multiline text; 1600×1000 light

Captured at 34e0e58371, reused unchanged: the production diff to 0148848580 only changes thread-switch handling (remount); mount restoration is unchanged.

About 0–3s add the multiline L2 comment and switch dark→light. Inspect 3–16s. Around 17–24s resize to 900×900, open the explorer and toggle source word wrap off/on. Fixed retains the wrapped block; baseline loses it after theme remount while keeping its chip. The fixed recording has an extra idle interval during a failed scroll lookup and ends with an outside click closing the narrow view. The short file needs no vertical scroll. Final stills hide the main sidebar on both sides and are taken after notes.md was also seeded: its chip remains but its block does not leak into math.ts.

O4 supplemental — 900×900 light, explorer hidden, source ≈377px

Captured at 34e0e58371, reused unchanged: the production diff to 0148848580 only changes thread-switch handling (remount); mount restoration is unchanged.

Same O4 comment, with the file view open and main sidebar/explorer hidden. The file view covers part of the composer at this width; judge the chip in the preceding desktop frame and recording. The full viewport is retained and the selected file/comment text is visible. This is resized desktop web, without device emulation.

O5 — Markdown source→rendered→source; 1600×1000 dark

Captured at 34e0e58371, reused unchanged: the production diff to 0148848580 only changes thread-switch handling (remount); mount restoration is unchanged.

About 0–2s add Markdown source note under notes.md L2 and toggle to rendered Markdown. At 3–15s rendered content has no source annotation by design and the chip persists. Around 16s return to source; inspect 17–20s. Baseline loses the source block; fixed restores it under L2. Normalized stills hide the main sidebar and retain both file chips.

O6 — unsent draft, actual file-tab close/reopen; 1600×1000 dark

Captured at 34e0e58371, reused unchanged: the production diff to 0148848580 only changes thread-switch handling (remount); mount restoration is unchanged.

About 0–2s add Draft: check addition at math.ts L2 before the first send. Around 2–4s close the actual file tab and reopen Files/math.ts; inspect 5–7s. Baseline loses the block while retaining the chip; fixed restores it. Normalized stills hide the main sidebar. Queued-message editing and send/promotion variants are unverified.

What else changed

• M1 replaces B's stale L6 block in A with A's L2 block; the intervening code lines shift with their annotation anchors. M2/M3 and the light still restore A's L2 block, moving following code and extending the source background by the block height. Their source text and composer chips otherwise match.
• O3 restores the block at the saved L3 anchor. The first blank source line and L3 chip match the baseline. After the sidebar B→A switch, Undo changes neither the source nor the restored block/chip. This history reset is behavior beyond stored-comment restoration and is called out at the top for review scope.
• At zero threshold all five recaptured still pairs have zero differences outside the file preview. The four M1/M2/M3/light after stills are also pixel-identical to their previous-head after originals. No masks or registration were used.
• The continuous after recording has a cached 49m relative-age label following its reload because the browser clock prelude was reapplied after React had initialized that label. The normal-reload supplement reapplies the same prelude earlier and supplies the final comparable M2/M3 stills. No comment is resubmitted. Main O3 has additional synthetic threads in the sidebar; the normalized paired stills hide that sidebar. Tool latency changes a few action times, recorded in action-plans.json.
• Reused O1/O2 retain the original three project-icon pixels with a 1/255 red-channel difference. O2 chip differences follow from baseline deletions blocked by missing restored Delete controls. Reused O1/O4/O5/O6 differences are restored comment blocks and consequent code movement/wrapping; original optional recordings retain equal-time synthetic thread ordering differences and O4's extra idle tail/outside click. Those captures retain their actual old revision labels.

Unverified coverage

• No requested recapture row is missing. O1, O2, O4, O5, O6 were intentionally not rerun at the new head; their original captures remain linked and labelled with their actual revision.
• Original O2 baseline deletion of restored comments is blocked because its inline Delete controls are absent. Original O1 diff-section filtering, O6 queued edit/promotion, and optional 390px / below-640px variants remain unverified.
• Native mobile, Electron, remote/relay/tunnel, every provider and every rendered file type were not run; this is web evidence for the supplied matrix.

Rebuild scripts and measurements

reproduce.md

seed.sh

capture-driver.mjs

supplement-driver.mjs

action-plans.json

browser-prelude.js

freeze-clock.cjs

make-diffs.py

review-recordings.py

build-evidence.py

upload-artifacts.py

upload-final.py

pixel-stats.json

recording-metadata.json

settled-checks.json

o3-observations.json

Original reused optional reproduction and additional seed: reproduce.md; seed-optional.sh

AI-assisted; verified with the tests and tools listed above.