Repository navigation
feat(canvas): show and act on a peer's files like local ones - #1880
Conversation
Remote sessions' files now list, preview and download on this device when the daemon advertises features.remote_files (core remote-files-parity): - use-attachment-url / use-waveform: a 504 (owner unreachable or transfer pending) retries on 2/4/8/16 s for five attempts, then a click-to-retry `retryable` state (RetryableAttachment, viewer Retry); 404/409 stay `missing`; no failure is cached. - use-session-files: gated on remote_files; an older daemon keeps today's sentence without a request; peer_update_required names the owner. Remote rows ride the same index/classification/sections/sort/filter path. - Files rows: a small device mark, and RemoteFileActionsMenu driven by the pure remoteFileActions table (core §3): open/preview via digest or the new sessions.files.content media op, "Download to this device" via the new long-budget sessions.files.download op with a receipt toast (Open file / Open folder / Copy path on the LANDED copy), unreadable rows disabled with the owner's reason. I3-1: the owner's display path never reaches openFile, showItemInFolder, readFileBytes or a save. - Library: remote cards carry the mark; reveal/copy-path never offered. - i18n: new keys in all eight locales (ledger audit pending).
i18n audit — round 1Reviewer: reviewer (i18n auditor), session Verdict: changes requested. 3 MAJOR (glossary violations in hi, ru, ur), 4 MINOR, 4 NIT. FindingsI1-1 · MAJOR · ur (
I1-2 · MAJOR · ru ( I1-3 · MAJOR · hi ( I1-4 · MINOR · es, fr, vi, zh-CN (
I1-5 · MINOR · fr (6 I1-6 · MINOR · zh-CN (11 I1-7 · MINOR · hi ( I1-8 · NIT · hi, ur ( I1-9 · NIT · vi ( I1-10 · NIT · zh-CN ( I1-11 · NIT · source/code, not counted in the receipts. No findings: es/ru/hi/zh-CN Receipts and bindingReceipts are in
Exact
Recommended remediation: apply I1-1…I1-7 (and the NITs, at your option) in the ONE remediation commit. Then:
The worktree was left unmodified ( |
Agent review — round 1Reviewer: reviewer (lopdev team, session 20d867221243) · model anthropic/claude-opus-5-5 (no fallback) Verdict: not clean — 1 BLOCKER, 1 MAJOR. The wire handling is solid and well tested; the two gating findings are a remaining I3-1 hole in the Files panel's audio band, and a renderer blob leak on remote opens. Verified (with evidence)
FindingsR1-1 BLOCKER — I3-1 broken: the Files panel reads a remote audio row's owner path from this disk.
R1-2 MAJOR —
R1-3 MINOR — a reloaded remote canvas tab is blank (the coder's self-declared gap).
R1-4 MINOR — remote files over 32 MiB still try to open in the canvas.
R1-5 MINOR — test gaps, and they are how R1-1 slipped through.
R1-6 MINOR — the Library preview drawer has no remote guard.
R1-7 NIT — The coder's self-declared gaps, judged against acceptance
Remediation for the gate: R1-1 and R1-2, each with a regression test. Minors and nits can go in the same commit or be deferred. |
QA report — round 1Verdict: FAIL. Two MAJOR findings (Q1-1, Q1-2), two MINOR, two NIT. Everything else in the matrix passed on the real built app. Context
Build provenance
Rig
Frames
Matrix (AFTER =
|
| # | Surface | Action | Actual | Result |
|---|---|---|---|---|
| M1 | Transcript, 1-image and 5-image folds | Open the remote session | 4 visible tiles each make one GET by digest → 200 at 378–380 ms. Real PNG bytes paint in both themes. img1 is shared by both folds and fetched once. BEFORE (legacy): the same 4 GETs → 409, broken-image tiles. | PASS |
| M2 | 504×2, then 200 | failNext(img5,2), then expand the 5-image fold |
504 @22 · 504 @2035 · 200 @6038. The tagged tile node is still mounted afterwards (sameTileNodeStillMounted:true); only its inner content node is added. Repeated on img3 in the collapsed strip: 504, 504, 200 at 0.38, 2.40 and 6.40 s. |
PASS |
| M3 | 504×99 → retryable → recovery | failNext(img4,99); after the retryable state, failNext(img4,0) and click |
5 attempts at 0.38, 2.40, 6.40, 14.41 and 30.41 s, then stop. The tile reads "Image 4 couldn't be fetched from the device that holds it. Retry". The click makes 1 GET → 200 (+4 ms). The button is swapped for the image inside the same parent (MutationObserver saw −BUTTON, +DIV). |
PASS |
| M4 | Canvas viewer 504 → Retry | failNext(mp4,99), then open the clip.mp4 tab |
5 attempts at 0.02, 2.03, 6.03, 14.06 and 30.06 s, then "Couldn't reach the device that holds this file. Try again in a moment. clip.mp4 [Retry]". Retry after reset makes 1 GET → 200; <video> reaches readyState 4 with duration 3.2 s. |
PASS |
| M5 | 409 is missing, not retried | opts.legacy=true, then expand the sound fold |
attachments/<wav> 409 and …/waveform 409, one each, with no retries over 38 s. Text: "tone.wav could not be played. Its stored copy is not available to this reader." |
PASS |
| M6 | 409 not cached forever | legacy=false, switch away and back |
Bytes are re-requested → 200 and the audio plays. The waveform is not re-requested, so the scrubber stays flat (Q1-4). | PASS (bytes) / see Q1-4 |
| M7 | No request loop | 60 s idle with every surface open | 0 /files or /attachments requests. The 39 background polls were: asks 12, sessions 12, health 6, mcp 4, team-bridge 2, capabilities 2, events 1. |
PASS |
| M8 | Capability absent | opts.remoteFiles=false + page reload |
Today's sentence, "Remote files cannot be opened on this device.", in both themes. Reopening the session with the catalogue already loaded made 0 /files requests. A cold deep-link reload made 1 request, 4 ms before the catalogue answered (Q1-6, pre-existing). BEFORE requests /files every time. |
PASS |
| M9 | peer_update_required |
opts.peerUpdate=true |
"Update lop-remote-node-2 to see its files here." + Retry, in both themes |
PASS |
| M10 | Files panel sections | Open canvas → files | 13 files. Highlights 1; Working files 12 = Images 5, Videos 1, Sounds 1, Documents 3, Spreadsheets 1, Other 1. Every row has [data-remote-file-mark] (12×12, "On lop-remote-node-2"), and the tooltip carries the owner reason. Opening the panel made 0 extra requests (thumbnails reuse the transcript's bytes). The unreadable compacted.png was never requested. |
PASS |
| M11 | Search, filter, sort | Type "render"; Filter → Sounds; Sort → Name | Search: 5 render rows, then back to 13 when cleared. Filter: ["tone.wav"], label Filter: Sounds. Sort: rows reorder within their sections, label Sort by: Name, Ascending. |
PASS |
| M12 | Row menu, readable | render-2.png ⋯ |
Show in canvas · Expand image · Download to this device · Save as… · Copy image. notes.md has Show in canvas · Download · Copy path on lop-remote-node-2 · Save as… |
PASS |
| M13 | Row menu, unreadable | .env, main.py, old.csv, dataset.tar, compacted.png |
Show in canvas / Open file / Open folder are disabled with one reason each: "keeps this file private", "Lives in …'s workspace. Ask the session to copy it to its scratchpad.", "No longer on …", "Too large to fetch from …", "no longer shares this file". Copy path on the device stays enabled (it is not offered for a digest row). | PASS |
| M14 | Download to this device | render-2.png → Download |
POST …/files/0000000000000002/download → 200. Toast: Downloaded render-2.png · Open file · Open folder · Copy file path. The file on disk has sha256 6961745e…ca84b964 = source image-2.png. The row menu then offers Open file / Open folder / Copy file path. |
PASS |
| M15 | Open folder / Open file on the landed copy | Row menu | main shell.showItemInFolder("<config>/downloads/render-2.png") and shell.openPath(same). Both are the landed local path, never the owner's. |
PASS |
| M16 | Download collision | Plant a different downloads/tone.wav, then download tone.wav |
Toast Downloaded tone (1).wav. tone.wav is untouched (sha dbf21b2e…); tone (1).wav sha 215be2d4… = source. |
PASS |
| M17 | Download while the owner is unreachable | failNext(file 03,1) → Download |
1 POST → 504. Toast "Still fetching from lop-remote-node-2. Try again in a moment.", with no automatic retry. |
PASS |
| M18 | Canvas, remote doc | notes.md → Show in canvas |
GET …/files/0000000000000008/content → 200. The doc renders "Remote notes / Written on lop-remote-node-2." The strip also says "No longer on disk…" (Q1-3). |
PASS / Q1-3 |
| M19 | Canvas, remote image, video and audio | Show in canvas | All open by digest with 0 extra requests. Video readyState 4, duration 3.2 s. The audio waveform plays. | PASS |
| M20 | App reload with remote canvas tabs | CDP Page.reload |
All 6 tabs remain. notes.md keeps its content (it did not go blank). The digest tabs re-fetch (200) and render. |
PASS |
| M21 | Header "Open folder" on a remote session | Press it | Toast "Could not open loading. The file may have been moved, renamed, or deleted.", which means the owner's cwd was stat'd on this disk (Q1-1). BEFORE does not render the header for remote sessions. |
FAIL |
| M22 | Library grid | Navigate to /library |
5 cards; the 3 remote cards carry the mark; the remote-render-3 thumbnail 404 shows the glyph. Without any interaction: GET /library/assets/la_0000000000000004/content → 200 for the remote audio card (Q1-2). |
FAIL |
| M23 | Library, open a remote card | Click remote-render-2 | Detail view paints the image (content 200 ×2, a duplicate fetch also seen for local cards). Open → content?download=1 (the browser download path, pre-existing for path-less assets; no visible receipt headless). |
PASS |
| M24 | Transcript lightbox | Expand Image 2 | Dialog with all 5 images; img5 fetched on demand → 200. | PASS |
| M25 | Audio row and waveform | Play | …/waveform 200; the player runs to 0:01/0:01 and returns to Play, in the transcript, Files and canvas. |
PASS |
| M26 | Video | Real H.264 clip | Transcript generate_video card, Files "Open to play" and canvas tab: readyState 4, 320×240, 3.2 s; played to 1.6 s (frame). BEFORE: "could not be played". |
PASS |
| M27 | Renderer memory | Before and after opening every remote asset | Largest renderer RSS 325.3 → 415.6 MB after opening all (7 canvas tabs including video), 469.7 MB after reload and lightbox. App process tree 1337.5 → 1539.1 → 1605.9 MB. JS heap at the end (Runtime.getHeapUsage) 71.6 MB. BEFORE, same flow minus the unopenable assets: 327.5 → 410.1 MB, heap 59.7 → 68.0 MB. performance.memory is bucketed in this Electron (constant 73.1 MB), so I did not use it. The assets total about 0.4 MB, so the growth is UI surfaces, not asset buffers. |
observed (no anomaly; see gaps) |
Findings
Q1-1 — MAJOR — Files header "Open folder" sends the owner's cwd to this device's reveal IPC
- Repro: in AFTER, open the remote session → Canvas → files, then press the header's Open folder.
- Toast:
"Could not open loading. The file may have been moved, renamed, or deleted."(frame).
- Toast:
- Code path:
finder-files-view.tsx:718-726:{cwd && <Button onClick={() => void revealLocalTarget(cwd)}>}is not gated onremote.- That reaches
main/index.ts:4084-4103:statSync(resolveUserPath(cwd))and thenshell.showItemInFolder(resolved).
- Why this PR: BEFORE renders only the sentence for a remote session (no header), so the PR makes this button newly reachable.
- Impact: when the owner's cwd also exists on the viewer (the same home layout on both Macs, e.g.
/Users/<me>/project), the button reveals an unrelated local folder. This is the I3-1 class the PR guards for rows, and the acceptance asks for "disabled or explained". - Suggested fix: hide or disable the header Open folder when
remote, with "On " as its reason. Add a hook or test pinning that.
Q1-2 — MAJOR — Library audio cards pull a remote asset's full bytes on view
- Repro: navigate to
/libraryand do nothing else. The fixture logsGET /v1/desktop/library/assets/la_0000000000000004/content 200forremote-tone.wav(remote:{device…}) about 1 s after the grid paints. - Code path:
library-card.tsx:184-189CardAudioWell:wantsBytes = asset.exists && near→useLibraryBytes(asset.id, "content", wantsBytes), with noasset.remotegate. - Why it matters: core design §5 routes
…/contentof a remote-occurrence asset through §1 fetch-on-miss, and its "No per-view mesh fetch" rule reserves that for explicit open, lightbox and download. So every remote sound card scrolled into view pulls the whole file over the mesh. That contradicts the acceptance "Library stays fast — no per-view byte fetch over the mesh" (music files make it worse). - BEFORE control: the same card code makes the same read when given the same rows (before-dark-e-library-grid-same-remote-rows). The read is old behaviour for local assets, but the remote rows are what this feature adds.
- Suggested fix: for
asset.remote, keep the glyph or flat scrubber and fetchcontenton the play press. Add a test asserting nolibrary.contentcall for a remote card on mount.
Q1-3 — MINOR — A remote document in Canvas shows local-disk freshness chrome
notes.mdopened from the remote row shows"Not read from disk yet · No longer on disk — showing the version we last read."and a Re-read from disk control that does nothing when pressed. The strip persists after reload. Frames: doc, after reload.- The doc path is a
blob:URL (remoteOpenableDocument), and the freshness layer treats it as a missing disk file. No local path is read, so this is a copy and honesty issue, not I3-1. - Fix: skip the freshness strip for
blob:orremotedocuments, or replace it with "From · read-only".
Q1-4 — MINOR — A waveform 409 is cached for the app's lifetime while bytes are not
- Repro: M5 then M6. After the owner answers 200 again, reopening the session re-fetches the bytes, but
/waveformis never asked again, so the scrubber stays flat (frame). - Code:
use-waveform.tsrequestPeakssettles every non-504 asnullincacheandsettledpermanently. - §1 makes 409 permanent, so this is defensible. But the operator-derived acceptance says the UI "must not cache the 409 forever", and the bytes path already re-asks on remount.
- Fix: drop the cached
nullwhen the answer was 409, the same way 504 is handled.
Q1-5 — NIT — A download onto an identical existing file reads like a fresh one
- With
created:false(same bytes already in Downloads), the toast still saysDownloaded notes.md. - Suggested copy: "Already on this device · Open file".
Q1-6 — NIT — out of scope, pre-existing — capability-absent race on a cold deep-link reload
- The hook issues one
/filesrequest 4 ms before the catalogue marks the session remote. - BEFORE requests every time, the output is unchanged, and the warm path makes 0 requests as the PR claims.
Coder's self-declared gaps, judged
| Gap | Verdict |
|---|---|
| Remote user-attachments keep their local path actions | Untested. The fixture's user-attachments returns items: []. It is the same I3-1 class as Q1-1, so it should be either closed or explicitly deferred in the thread. |
| Library download over 32 MiB | Untested. There is no large asset in the fixture. |
| Canvas tab goes blank after reload | Not reproduced for markdown: content persists. Digest-backed tabs re-fetch and render. Binary blob docs (PDF) are untested. |
| No MP4 in the fixture | Resolved in QA with a MediaRecorder H.264 clip. Video works on all three surfaces (M26). |
| i18n ledger (42 findings) | Out of scope here; owned by the i18n stream. |
Gaps I could not cover
- The real core and mesh. The fixture stands in for the core wire, so download landing and naming are the fixture's, and the UI half is what was verified. No live peer was used, per the isolation rules.
- RAM under large assets. The fixture assets are about 0.4 MB, so "RAM flat" for big media is unmeasured.
- Clipboard actions. Copy image, Copy path on the device and Copy file path were not read back.
- Save as… It opens an OS dialog, which this headless run cannot drive.
- A real Finder or file manager. Shell calls were recorded, not executed.
- Remote user-attachment rows — see the table above.
Design review — round 1Reviewer: designer (lopdev). Served model Verdict: changes requested. One MAJOR (D1-1), five MINOR and two NIT. Overall the transcript, canvas and Library now show remote media the way they show local media. The remote indicator is the item that still has to change. What works (no change needed)
FindingsD1-1 — MAJOR — The remote mark is the wrong glyph, and it sits next to the metadata instead of the name.
D1-2 — MINOR — The compact retryable tile is a bare arrow.
D1-3 — MINOR — The pending state during the owner backoff is a blank, wrongly sized well.
D1-4 — MINOR — The menu on an unreadable row has a different shape from the readable one, and its disabled labels are barely legible.
D1-5 — MINOR — Unreadable rows don't show why at row level; the blank tile looks like a failed thumbnail.
D1-6 — MINOR — Library detail of a remote asset doesn't say it is remote.
D1-7 — NIT — The canvas unreachable copy doesn't name the device it already knows.
D1-8 — NIT — The download toast doesn't say where the file landed.
Not judged / out of scope
|
UX review — round 1Verdict: NOT TERMINAL. Three MAJOR findings (U1-1, U1-2, U1-3), five MINOR, two NIT. They don't duplicate R1-, Q1- or I1-*. The Files panel's sticky sections, sort, filter, row marks and the receipt toast all feel like the local ones. The friction is in the waiting and failure states, and in a remote document that can be edited but never saved. Context
Flows walked
FindingsU1-1 — MAJOR — A remote picture waits 30 s or more as an empty, unlabelled box, and Retry turns back into that box
U1-2 — MAJOR — "Download to this device" is silent while it runs, and a 504 always says "Still fetching", even when the owner is offline
U1-3 — MAJOR — A remote text file opens in a live editor whose edits are silently never saved
U1-4 — MINOR — An unreadable row looks readable, and keyboard focus doesn't show the reason
U1-5 — MINOR — A missing remote row says "No longer on disk", and its own menu says "No longer on lop-remote-node-2"
U1-6 — MINOR — Owner-offline copy hides the device name and promises "a moment"
U1-7 — MINOR — "Update lop-remote-node-2 to see its files here." says what, not how
U1-8 — MINOR — Library Open/Download of a remote card gives no progress, and on 504 a generic failure
U1-9 — NIT — The download receipt toast and its three actions dismiss after Sonner's default 4 s
U1-10 — NIT — The Library detail view for a remote asset carries no remote mark
Retained, no finding
For the remediation wave
|
Answers the round-1 ledger of PR #1880 (review, QA, design, UX, i18n): - R1-1: a remote sound row never reaches FinderPathSound; unreadable rows state the reason, readable ones play from the content route on a press. - R1-2/R1-3: remote canvas documents are `remote-file:` identifiers whose viewers fetch over the content route under one refcounted blob entry, released with the tab; nothing dead persists across a reload. - Q1-1: the Files header never reveals a remote session's (owner's) cwd. - Q1-2: remote Library sound cards read nothing on view; a press arms it. - D1-1: one Monitor mark on the Files count line; the per-row glyph is gone. - U1-1: a slow owner shows "Fetching from <device>…" (transcript + canvas); the compact retry tile is captioned; Retry starts in the waiting state. - U1-2: Download shows a loading toast, re-POSTs a still-running transfer on the backoff, names the offline device, the landing folder, and an identical file already there; the receipt toast lasts 8 s. - U1-3: remote text opens read-only ("From <device> · read-only") with no local freshness strip. - Folds in D1-4..D1-7, U1-4/5/7/8/10, Q1-3/4, R1-4, R1-6..R1-9 and the I1 string corrections across the 7 locales.
…the translator
Remediation — round 1Implementer: coder (lopdev), served model MUST FIX
FOLD IN
Also, found on the way (same I3-1 class, small): DEFERRED
Not addressed (noted)
i18nEdited pairs (round-2 delta audit, COMMANDS.md §C): es/hi/ru/ur/vi/zh-CN/fr Targeted test evidenceHead
|
Agent review — round 2Reviewer: reviewer (lopdev, session 581b171b8866), served model Verdict: 1 MAJOR open (R2-1). Everything else is MINOR/NIT. Round-1 findings: verified on the head
Delta review: what you asked me to probe
FindingsR2-1 — MAJOR — a remote workbook is editable, and its saves fail silently (the U1-3 class, still open for xlsx).
R2-2 — MINOR — a holder that unmounts mid-fetch releases another holder's reference (inherited pattern, now on an always-shared key).
R2-3 — MINOR — the inline remote sound has no size gate (R1-4 gap).
R2-4 — MINOR — two presses for a remote path sound (coder's not-addressed item 2).
R2-5 — NIT — library-model.ts:208-209: the moved Coder's not-addressed items
Tests run (the delta's own files only, bounded)I did not re-run the "fails on a4462e7 src" counts. They are the coder's, from the remediation reply. Out of scope — pre-existing: a remote workbook's base64 |
i18n audit — round 2Reviewer: reviewer (i18n auditor), session Method. For each changed value I read the en value with its
Read-only gates at head: Verdict: clean. No BLOCKER or MAJOR. 2 MINOR + 1 NIT are counted in receipts, and 2 MINOR are source-level and not counted. All are deferred; none needs a new commit. I1 verificationEvery requested string is byte-exact at head:
I1-1/I1-5 on FindingsI2-1 · MINOR · zh-CN ( I2-2 · NIT · hi ( I2-3 · MINOR · source/code, not counted. In I2-4 · MINOR · source/code, not counted · ur RTL. Out of scope for i18n: Receipts and bindingReceipts, one per locale, are in
The exact bind lines and translator-provenance lines are in
Bind against the bytes at 79ae385. Applying I2-1/I2-2 would change those bytes and need another delta audit, so they stay deferred. The worktree was not written to ( |
QA report — round 2Verdict: PASS. All round-1 FAIL rows (Q1-1, Q1-2) and every MUST FIX row I could observe now pass on the built app at Context
Build provenance: AFTER is
Rig
Frames (64 PNGs, light + dark): https://github.com/damianvtran/local-operator-ui/tree/evidence/1880-qa-r2/frames
In the table, request lines are fixture hit-log entries ( Matrix (AFTER =
|
| # | Ledger | Surface | Action | Actual | Result |
|---|---|---|---|---|---|
| N1 | R1-1 | Files row, remote path sound song.wav |
Idle, then press Play from lop-remote-node-2, then Play | Idle: no request. Press 1: +10 · GET …/files/000000000000000e/content → 200 and IPC read-file-bytes ×0. Press 2: play() on a blob: holding 66 194 B audio/wav (the content bytes), which runs 1.5 s to 0:01/0:01. The decoy at the owner's path on this disk was never read. |
PASS |
| N2 | R1-1 | Canvas, remote path sound | Row menu → Show in canvas → Play from → Play | 1 content GET → 200, read-file-bytes ×0, and it plays. After a prior row play the canvas can say "This sound could not be played" (Q2-4, pre-existing). |
PASS |
| N3 | R1-2 | Canvas open/close ×10, remote doc notes.md |
Show in canvas → close tab, 10 times, then GC | 10 content GETs. 0 blobs created for the doc. The 10 createObjectURL calls were all audio/wav from the tone row remounting as the panel returned, and 10 were revoked; the one left live is the mounted tone row. JS heap after GC went 48.8 → 50.8 MB, and the largest renderer RSS went 325.2 → 332.5 MB. |
PASS |
| N4 | R1-2 | Same, song.wav with load and play each cycle |
10 cycles | 20 creates and 20 revokes (10 song + 10 tone remount), live = the mounted tone row only. Heap 51.9 MB, renderer 336.4 MB. | PASS |
| N5 | R1-2 | Same, budget.xlsx |
10 cycles | 0 xlsx blobs (it uses base64 content). 9 content GETs for 10 opens. Heap 56.8 MB, renderer 345.6 MB, app tree 1441.6 MB. IPC read-file-bytes ×0 across all 30 cycles. |
PASS |
| N6 | R1-3 | App reload with 4 remote canvas tabs | CDP Page.reload with notes.md, budget.xlsx, render-2.png, song.wav open |
All 4 tabs were restored. notes.md re-fetched (content → 200) and rendered with its read-only caption. render-2.png painted. song.wav was idle at "Play from…". budget.xlsx kept its in-memory buffer, matching the close toast's "kept until the app quits". There were no blob: paths and 0 read-file-bytes. |
PASS |
| N7 | Q1-1 | Files header, remote session | List the header buttons | Documents view, files 16, Variables view, New file, Open file from disk, Close canvas, Mentioned, Filter has no Open folder. The count line reads 16 files · [Monitor] On lop-remote-node-2. The shell log was empty. Control: a local session's header on the same build shows Open folder. |
PASS (round-1 M21 FAIL fixed) |
| N8 | Q1-2 | /library |
Navigate there and idle 26 s, including two theme captures | 1 list request and 4 thumbnails (la_…3 → 404 glyph). 0 content/waveform requests. The first press on the remote tone's Play from lop-remote-node-2 sends +5 · …/la_0000000000000004/content → 200 and …/waveform → 404 (the fixture has no library waveform). The control then becomes Play remote-tone.wav. |
PASS (round-1 M22 FAIL fixed) |
| N9 | U1-1/D1-2/D1-3 | Transcript image tile pending | failNext(img2, 99) → open the session |
The tile keeps its 117×78 box with a centred spinner. "Fetching from lop-remote-node-2…" appears only as aria-label/title (Q2-2). |
PASS / Q2-2 |
| N10 | U1-1 | Transcript retryable → Retry | Wait for the backoff → press | The tile shows a ↻ glyph with a visible Retry caption. Its name is "Image 2 lop-remote-node-2 isn't answering. Try again when it's online. Retry". The press shows the pending spinner box, not an empty box (+11 · 504, +2015 · 504). After failNext(…,0) and Retry, all 5 tiles decode. |
PASS |
| N11 | U1-1 | Transcript audio and video pending | failNext(tone/mp4, 99) |
Sunken wells read "Fetching from lop-remote-node-2…" next to the spinner, in both themes. | PASS |
| N12 | U1-1/D1-2 | Canvas video and image pending → retryable → recovered | failNext(mp4/img4, 99) → Show in canvas |
Pending reads "Fetching from lop-remote-node-2…" (no "Opening…"). Retryable reads "lop-remote-node-2 isn't answering. Try again when it's online. · clip.mp4 · [Retry]". After the reset, <video> reaches readyState 4, 3.2 s, 320 px wide. |
PASS |
| N13 | U1-2 | Download, pending ×3 | failNext(file04, 3, fetch_pending) → Download |
POSTs 504, 504, 504, 200 at intervals of +2025, +4003, +8002 ms, with no re-click. The loading toast "Downloading render-4.png from lop-remote-node-2…" is replaced in place by "Downloaded render-4.png to downloads · Open file · Open folder · Copy file path". The landed sha256 equals the source. |
PASS |
| N14 | U1-9 | Success toast duration | Sample at 200 ms | Visible from 6.53 s to 14.66 s, which is 8.1 s. | PASS |
| N15 | U1-2/U1-6 | Download, owner offline | failNext(file02, 1, owner_unreachable) |
1 POST → 504, with no re-POST. Toast: "lop-remote-node-2 isn't answering. Try again when it's online." | PASS |
| N16 | U1-2 | Download, pending exhausted | failNext(file01, 9, fetch_pending) |
5 POSTs (+0, +2.1, +6.2, +14.2, +30.2 s), then "Still fetching from lop-remote-node-2. Try again in a moment." |
PASS |
| N17 | Q1-5 | Download onto an identical file | Reload, then download render-5 / render-4 again | 1 POST → 200 created:false. Toast: "render-5.png is already in downloads on this device · Open file · Open folder · Copy file path". No new file was written. Within one app session the row offers Open file instead of Download, so the case needs a reload to reach. |
PASS |
| N18 | M15 (neighbour) | Landed actions | Row → Open folder / Open file | showItemInFolder and openPath were each called with <config>/downloads/render-5.png, the landed path. |
PASS |
| N19 | U1-3/Q1-3 | Canvas, remote notes.md |
Show in canvas, then click and type "XYZ" | Caption "From lop-remote-node-2 · read-only". No disk freshness strip (/disk/ matched nothing). CodeMirror contenteditable=false aria-readonly=true, and the text was unchanged after typing. The content came over the route; the decoy was not shown. |
PASS |
| N20 | not addressed | Canvas, remote budget.xlsx |
Double-click a cell, type 99, Enter, Ctrl+S, close the tab | The edit is accepted and Ctrl+S gives no feedback: no toast, no write IPC, no POST. Only on close does the app say "Closed budget.xlsx. Your edits could not be saved, and are kept until the app quits — opening it again brings them back." The owner file is never touched. | Q2-1 (MINOR) |
| N21 | not addressed | Two-press sound | Files row, canvas, Library card | Each surface needs Play from (load) and then Play. No surface fetches on view. | Q2-3 (NIT) |
| N22 | D1-1 | Remote mark | DOM and frames | One [data-remote-file-mark] per panel, on the count line: lucide-monitor size-3 shrink-0 text-ink-dim inside text-meta text-ink-muted. No per-row glyphs. Library: one Monitor glyph on each of the 3 remote cards and none on the 2 local ones. |
PASS |
| N23 | D1-4/D1-5/U1-4 | Unreadable rows and menus | .env, main.py, old.csv, dataset.tar, compacted.png |
Same entries as a readable row, with transfers [DIS] and Copy path enabled. The reason appears once as a label at the top (e.g. "lop-remote-node-2 keeps this file private"). Each row shows its reason in the status slot, and compacted.png has a type glyph in its tile. |
PASS |
| N24 | U1-5 | Missing row | old.csv |
Status: "No longer on lop-remote-node-2". | PASS |
| N25 | R1-4 | Remote doc over 32 MiB | big-notes.md (40 MiB) menu |
"Show in canvas" is [DIS] with "Too large to open from lop-remote-node-2. Download it to this device instead."; Download is enabled. |
PASS |
| N26 | U1-7 | Older peer | opts.peerUpdate=true |
"Update lop-remote-node-2 to see its files here." / "Run lop network update lop-remote-node-2 on this device, or update Local Operator on lop-remote-node-2, then retry." / Retry. lop network update [peer] exists (lop network update --help). |
PASS |
| N27 | Q1-4 | Waveform 409 not cached | legacy=true → open → legacy=false → switch away and back |
waveform 409 + bytes 409, then on remount waveform 200 + bytes 200. |
PASS |
| N28 | D1-6/U1-10/R1-6 | Library remote detail | Open remote-render-2 | Meta reads "… · [Monitor] On lop-remote-node-2", with buttons Open · Download to this device. The asset menu has no Open file or Show in folder entries. The shell log was empty. | PASS |
| N29 | U1-8 | Library remote Download | Press Download to this device | Loading toast "Downloading remote-render-2.png from lop-remote-node-2…", then "Saved a copy"; content?download=1 → 200. |
PASS (offline untested) |
| N30 | beyond ledger | Canvas viewer menus, remote image and video | ⋯ |
Image: Save as…, Copy image. Video: Save as…. Neither offers Open file, Open folder or Open in default app. | PASS |
| N31 | M7 (neighbour) | No request loop | 60 s idle with the transcript fold, Files panel and 4 remote canvas tabs open (fresh launch) | 34 background polls (sessions 12, asks 10, health 6, events 2, team-bridge 2, capabilities 2). 0 /files, /attachments, /content or /waveform. |
PASS |
| — | R1-5 | Regression tests | Targeted run | 82/82. | PASS |
| — | R1-7..R1-9, I1-1..I1-10 | Comments and translations | — | Not a QA surface (I1 belongs to the i18n audit stream). The English copy above was read on screen. | n/a |
| M1, M5, M8, M11, M16, M19, M24–M26 | — | Round-1 rows that the delta does not reach | — | — | retained (a4462e7) |
| M2–M4, M6, M7, M10, M12–M15, M17, M18, M20–M23, M27 | — | Superseded by N-rows above | — | — | re-run |
Findings
Q2-1 — MINOR — A remote .xlsx accepts edits, and the refused save is only explained when the tab closes
- Repro (N20): Files →
budget.xlsx→ Show in canvas. Double-click42and type99, then Enter, then Ctrl+S.- The cell shows 99. There is no toast, dirty mark or write. The owner file never changes, and there is no IPC write or POST.
- Closing the tab shows "Your edits could not be saved, and are kept until the app quits".
- Frames: edited, close toast.
- Why only MINOR: nothing is lost or overwritten, and the refusal is explained, just late.
- It is the same class as U1-3. The acceptance reads "disabled or explained", and explaining at close meets it in a weak form.
- Suggested fix:
- Give the workbook viewer the same "From · read-only" caption and a read-only grid (
editable:false), as markdown now has. - Alternatively, defer it in the thread as the coder proposed.
- Give the workbook viewer the same "From · read-only" caption and a read-only grid (
Q2-2 — MINOR (design to judge) — The compact pending tile shows only a spinner
- What I saw (N9): "Fetching from …" is the tile's
aria-label/titleonly. The 117×78 transcript tile shows a spinner without text. The wells (audio, video) and canvas do show the sentence. - The ledger's U1-1 asked for "a centred Spinner plus 'Fetching from …'". The coder's comment says there is no room in the tile.
- Frame: zoom.
- Whether this deviation meets D1-2 is the design stream's call. It does not gate on QA.
Q2-3 — NIT — Remote path sounds take two presses (N21; the coder already listed this)
- Load, then Play, on the Files row, the canvas and the Library card. This is the deliberate trade for "no per-view mesh fetch".
- Suggested fix: start playback automatically once the bytes land after the first press. The fetch would still only happen on the press.
Q2-4 — out of scope, pre-existing — After a row play, the same sound in a canvas tab can fail with "This sound could not be played"
- Repro:
- Files row: Play from → Play, and let it end.
- Show in canvas, then wait about 30 s.
- Play from → Play.
- Result:
MediaError 2and "This sound could not be played" (frame). With a gap of about 6 s it happens to play from the element's buffer.
- Cause:
- The row and the canvas share
ownerId(sound:remote-file:…).audio-engine-store.toggletreatsendedasresume()on the old elementsrc, which is the row'sblob:that was revoked when the row unmounted. The canvas's freshly fetched URL is never used.__qaPlayshowsplay()on the revoked…a4c3131c118ewhile the new…a9ffe3cdbb66sits unused.
- The row and the canvas share
- Why it is pre-existing:
- Same-head control: a LOCAL session's
local-song.wav(FinderPathSound,ownerId sound:path:…) reproduces it identically (frame). - Base
83a18c632has the sameownerId, the per-holdercacheKeyand theended → resumepath. - The new remote player inherits this behaviour; it did not introduce it.
- Same-head control: a LOCAL session's
- Suggested follow-up (not this PR):
toggleshould callplay()whenurl !== state.url.
Gaps I could not cover
- Library offline sentence (U1-8 504). The fixture's library route has no 504 gate.
RemoteAssetUnreachableErroris covered only by the coder's unit tests. - Where Library "Saved a copy" lands. It uses the browser-anchor download path (pre-existing), and nothing was observable headless.
- Real core and mesh, and RAM for large assets. As in round 1, the fixture stands in for the core wire, and the assets are under 120 KB.
- Save as… OS dialog and clipboard read-back. Not driven headless.
Design review — round 2Reviewer: designer (lopdev). Served model Verdict: TERMINAL. All of D1-1..D1-8 are resolved. No BLOCKER or MAJOR is open. Two new MINORs and one NIT are in the delta, and none of them gates. Round-1 findings
Ruling on Q2-2 (spinner-only small pending tile): meets D1-3, no change from designThe D1-3 spec for the 117×78 tile was to keep the box and centre the shared Spinner. I specified the sentence only for the canvas, and the wells now carry it too. A caption would not fit:
The sentence stays in New in the delta (non-gating)D2-1 — MINOR — In tile layout, the unreadable reason truncates to the device name.
D2-2 — MINOR — The pending video well is a 38 px pill that grows by about 204 px when the video resolves.
D2-3 — NIT — The device name breaks at its hyphen when a line wraps.
Not judged / out of scope
|
UX review — round 2Verdict: TERMINAL. U1-1 through U1-10 are all fixed at Context
Round-1 findings
Flows not moved by the delta (sticky sections, sort and filter, and row-menu focus return) are retained (a4462e7). Ruling on Q2-3 (two-press remote sound)Not acceptable as the end state. Fine to ship as non-gating MINOR now; fix it in the next wave beside R2-1.
New findings (round 2)U2-1 — MINOR — The new remote text preview and remote path sound treat every 504 as "offline", have no backoff and no Retry, and the text preview still says "Opening…"
U2-2 — MINOR — Pressing Retry on a transcript tile drops keyboard focus
U2-3 — NIT — Long unreadable reasons are cut, and the actionable part is cut first
U2-4 — NIT — The older-peer command is plain prose
Noted, outside the delta (not filed): in the Files panel, a thumbnail that is pending or offline (render-2 in this frame) shows the same File glyph as an unreadable row, with the status "In session". The only change to For the next wave (non-gating)
|
- R2-1: a peer-held workbook opens read-only. No editable cell, no autosave, no Ctrl+S, and the "From <device> · read-only" caption the remote text viewer uses. Its save could never reach the owner and failed silently. - R2-2: useFileBlobUrl releases only a reference this holder took. A holder that unmounted mid-fetch spent another holder's reference on the shared remote key and revoked the URL that holder was rendering. - R2-3: the inline remote sound row refuses rows over the relay window with the existing "Too large to open from <device>" sentence, not a failing press. - U2-1: the remote text viewer and the remote path sound now use the transcript's 504 backoff, show "Fetching from <device>…" while waiting, and offer Retry. Main forwards the content read's 504 code, so attachment_fetch_pending reads "still fetching" and only an unreachable owner is called offline. - U2-2: Retry on a transcript tile or well keeps keyboard focus through the wait and back. - D2-1: tiles no longer paint the truncated unreadable reason; it stays in the tooltip and the accessible name. - D2-2: a pending transcript video reserves a 16:9 box at the player's 240 px ceiling. - D2-3: device names stay on one line in the download toast and menu reasons. - I2-3: the "lop network update" line shows only with a real device name. - R2-5: JSDoc blank line. No catalogue bytes change; every string is an existing key.
Remediation — round 2Implementer: coder (lopdev), served model Dispositions
i18n binding (
|
| file | pass | fail | run |
|---|---|---|---|
| scripts/remote-viewers-round2.test.mjs | 5 | 0 | rc=0 wall=2.7s |
| scripts/file-blob-url.test.mjs | 22 | 0 | rc=0 wall=1.5s |
| scripts/remote-sound-row.test.mjs | 5 | 0 | rc=0 wall=1.6s |
| scripts/remote-files.test.mjs | 22 | 0 | rc=0 wall=1.5s |
| scripts/notification-sound-main.test.mjs | 8 | 0 | rc=0 wall=3.8s |
| scripts/attachment-url.test.mjs | 25 | 0 | rc=0 wall=0.6s |
| scripts/chat-image-expand.test.mjs | 45 | 0 | rc=0 wall=9.1s |
| scripts/canvas-attachment-locale.test.mjs | 1 | 0 | rc=0 wall=2.4s |
| scripts/finder-reveal-scroll.test.mjs | 7 | 0 | rc=0 wall=1.2s |
| scripts/sound-player.test.mjs | 12 | 0 | rc=0 wall=2.4s |
| scripts/image-preview-decode.test.mjs | 4 | 0 | rc=0 wall=2.1s |
| scripts/library-model.test.mjs | 14 | 0 | rc=0 wall=1.6s |
| scripts/library-card.test.mjs | 7 | 0 | rc=0 wall=2.1s |
| scripts/session-files-hook.test.mjs | 10 | 0 | rc=0 wall=2.6s |
| scripts/canvas-file-freshness.test.mjs | 40 | 0 | rc=0 wall=1.8s |
| scripts/desktop-contract.test.mjs | 56 | 0 | rc=0 wall=44.9s |
Against 79ae38539's src (this head's tests, git stash push -- src):
remote-viewers-round2: 4 of 5 FAIL. Only the local control passes.file-blob-url: 3 of 22 FAIL.remote-sound-row: 2 of 5 FAIL.remote-files: 4 of 22 FAIL.notification-sound-main: 1 of 8 FAIL.
Other gates:
tsc --noEmitpasses ontsconfig.app.jsonandtsconfig.node.json(rc=0).- biome on the 20 changed files: 0 errors. The warnings are the test files' existing
useTopLevelRegexpattern. - The pre-push gate (scripts lint, src lint, types) passed against origin/main
5033f64.
Not run here: pnpm test:affected and the full suite (CI's), and Xvfb captures.
For the round-3 streams: D2-1 and D2-2 need a rendered frame. U2-1 and U2-2 are covered in JSDOM, with the real resolver and timers.
Agent review — round 3Reviewer: reviewer (lopdev), served model What I reviewed: the delta round for R2-1 (MAJOR), plus the fold-ins and the new Verification of prior findings
|
| file | tests | pass | fail | rc | wall |
|---|---|---|---|---|---|
| remote-viewers-round2.test.mjs | 5 | 5 | 0 | 0 | 2.7 s |
| file-blob-url.test.mjs | 22 | 22 | 0 | 0 | 1.7 s |
| remote-sound-row.test.mjs | 5 | 5 | 0 | 0 | 1.8 s |
| remote-files.test.mjs | 22 | 22 | 0 | 0 | 1.9 s |
| notification-sound-main.test.mjs | 8 | 8 | 0 | 0 | 3.8 s |
I did not re-run the "fails on 79ae385" claims, because that would need a second checkout. For R2-1, R2-2 and R2-3 I derived from the diff that the old code fails each new assertion (hard-coded editable: true, unconditional release, no size gate).
Verdict: clean. No BLOCKER or MAJOR remains, and R2-1 is fixed. This round is terminal: MINOR-1, MINOR-2 and NIT-1 are deferred in this thread and need no new commit.
QA report — round 3Verdict: PASS.
Context
Frames: 44 PNGs in light and dark, on orphan branch
Matrix (AFTER =
|
| # | Ledger | Surface | Action | Actual | Result |
|---|---|---|---|---|---|
| Q1 | R2-1 | Canvas, remote budget.xlsx |
Row menu → Show in canvas. Then: double-click the 42 cell, real key events 9,9, Enter (plus, in an earlier pass, single-click + key 8 and F2), Ctrl+S, close the tab. |
Open: 1 GET …/files/000000000000000f/content → 200. Caption reads "From lop-remote-node-2 · read-only". The root has data-spreadsheet-readonly="true".Typing: double-click opens no editor (no .ag-cell-inline-editing/.ag-cell-editor). Single-click + 8 and F2 opened none either. The cells stay Relay 42 Storage 17, and with real key events the composer stays ''.Ctrl+S: no toast. Main-process IPC after Ctrl+S held only background polls ( sessions.list, asks.list, mcp.list, backend-status): no write and no POST.Close: no toast within 2.5 s. The toast recorder's negative control did catch the "Too large…" toast on the same page. CONTROL 79ae, same driver code: the editor opens and the cell becomes 99. Ctrl+S gives no feedback. Close shows "Closed budget.xlsx. Your edits could not be saved, and are kept until the app quits — opening it again brings them back." |
PASS (MAJOR fixed) |
| Q2 | R2-2 | Same remote sound in two holders (Files row → canvas tab) | failNext(song, 3, fetch_pending). Press Play from on the row, which then waits in the backoff. Show in canvas, so the row unmounts mid-fetch. Press Play from in the canvas. |
Content GETs: +0 504, +2.0 504 (row), +5.6 504, +7.6 200 (canvas). The row's third attempt, due at +6 s, never went out: its unmount stopped asking. The blob log shows 1 create (the canvas's audio/wav) and 0 for the abandoned row, and the canvas URL stayed live. The concurrent two-holder case is not reachable in this UI (see Gaps), so I ran the coder's R2-2: a holder that leaves mid-fetch never releases another holder's reference test, which passes. |
PASS (real path: sequential handoff; concurrent case: unit test) |
| Q3 | R2-3 | Files row, big-song.wav (40 MiB, path-backed) |
Idle. Press the name. Open the row menu. | Row: "41.9 MB · Too large to open from lop-remote-node-2. Download it to this device instead." The row has no Play from control and no <audio>.Name press: a toast with the same sentence and 0 requests to …/0000000000000011/.Menu: Show in canvas [DIS] with that reason. Download is enabled.CONTROL 79ae: the row offers Play from lop-remote-node-2. |
PASS |
| Q4 | U2-1 | Canvas, remote notes.md |
failNext(notes, 2, fetch_pending) → Show in canvas. |
"Fetching from lop-remote-node-2…" (not "Opening…"). GETs 504, +2.0 s 504, +4.0 s 200, then the content renders with the read-only caption. |
PASS |
| Q5 | U2-1 | Same | failNext(notes, 9, fetch_pending) |
"Fetching from lop-remote-node-2…" for about 30 s over 5 GETs, at gaps of 2.0, 4.0, 8.0 and 16.0 s, all 504. Then: "Still fetching from lop-remote-node-2. Try again in a moment." · notes.md · [Retry] (data-remote-text-retry). |
PASS |
| Q6 | U2-1 | Same | failNext(notes, 0) → keyboard Enter on Retry |
1 GET → 200. "Fetching…" shows, then the content within 0.2 s. | PASS |
| Q7 | U2-1 | Same | failNext(notes, 9, attachment_owner_unreachable) |
The same 5-GET backoff, then "lop-remote-node-2 isn't answering. Try again when it's online." · notes.md · [Retry]. CONTROL 79ae under fetch_pending ×9: 1 GET, then at once the offline sentence with no Retry. |
PASS |
| Q8 | U2-1 | Files row, remote path sound song.wav |
failNext(song, 2) → Play from |
"Fetching from lop-remote-node-2…" (spinner), then the player (Play song.wav, 0:00). GETs at +0 504, +2.0 504, +6.0 200. |
PASS |
| Q9 | U2-1 | Same | failNext(song, 9, fetch_pending) → Play from. Then failNext 0 → keyboard Enter on Retry → Play. |
5 GETs (0, 2.0, 6.0, 14.0, 30.0 s), then "Still fetching from lop-remote-node-2. Try again in a moment. · Retry" (data-remote-sound-retry). Retry sends 1 GET → 200, and the player loads and plays to 0:01 / 0:01. |
PASS |
| Q10 | U2-1 | Same | failNext(song, 9, attachment_owner_unreachable) |
5 GETs, then "lop-remote-node-2 isn't answering. Try again when it's online. · Retry". | PASS |
| Q11 | U2-1 (main) | desktop-media.ts forwards the 504 code for sessions.files.content |
Rows Q5/Q7/Q9/Q10 | The same HTTP 504 renders different sentences depending on detail.code. The renderer sees that code only through refusal, which main now forwards for this op. CONTROL: fetch_pending was told "offline". |
PASS |
| Q12 | U2-2 | Transcript image tile (Image 2) | failNext(img2, 99) → wait for Retry → .focus() the tile → Enter |
Focus moves BUTTON (Retry) → OUTPUT ("Fetching from lop-remote-node-2…", tabindex=-1, data-attachment-pending) → after 5 GETs, BUTTON ("Image 2 lop-remote-node-2 isn't answering…"). Focus is never on <body> during the swaps. |
PASS — see Q3-1 for the success case |
| Q13 | U2-2 | Transcript image tile and audio well, Retry that succeeds | failNext(img2, 0) → Enter. Audio: failNext(tone, 1) → Enter. |
Image: 1 GET 200, the tile becomes Expand Image 2, and focus goes to BODY. Audio: BUTTON → OUTPUT → BODY once the player loads (504, +2.0 s 200). |
Q3-1 (MINOR) |
| Q14 | D2-1 | Files tiles band, unreadable compacted.png |
DOM + zoom frames | The only visible text is "compact…" (the name), plus the type glyph. The reason sits in span.sr-only (1×1 px, clip-path: inset(50%)). The button's accessible text is "compacted.png / lop-remote-node-2 no longer shares this file", and the title also carries the reason. CONTROL 79ae: the reason is in a visible <bdi>. |
PASS |
| Q15 | D2-2 | Transcript video well (clip.mp4, a 4:3 clip) |
failNext(clip, 3) → open the fold. Sample the box and the next message on every rAF, with a PerformanceObserver('layout-shift'). |
Pending: 424×239 (16:9) with spinner and "Fetching from lop-remote-node-2…". After +14.1 s 200: <video> readyState 4 at 320×240. The next message's top moved 961 → 965 → 964 px (≤ 4 px). 0 layout-shift entries during the load; the only entry, 0.0216, came from opening the fold, before any bytes. |
PASS (≈0); width note in Q3-2 |
| Q16 | D2-3 | Row menu reasons and disabled-action reason | Open menus for .env, main.py, old.csv, dataset.tar, big-notes.md |
Each label wraps over 2–4 lines at 208/240 px, but lop-remote-node-2 sits in a span.whitespace-nowrap with 1 client rect every time. |
PASS |
| Q17 | D2-3 | Download loading toast | failNext(render-4, 3, fetch_pending) → Download |
The toast wraps as "Downloading render-4.png from" / "lop-remote-node-2…". The name is a single rect (nowrap). POSTs: 504, 504, 504, 200 (retained N13 behaviour). |
PASS |
| Q18 | I2-3 | Files panel, older peer | opts.peerUpdate=true, with device_name: null and then with the name |
No name: only "Update the other device to see its files here." There is no lop network update line. Named: "Update lop-remote-node-2 to see its files here." / "Run lop network update lop-remote-node-2 on this device, or update Local Operator on lop-remote-node-2, then retry." |
PASS |
| — | R2-1…I2-3 | Delta tests | Targeted run | 62/62 | PASS |
| N1–N19, N21–N31 | — | Round-2 rows outside this delta's retest list | — | — | retained (79ae385) |
| N20 | — | Old xlsx row | — | Superseded by Q1 | — |
Findings
Q3-1: MINOR. A Retry that succeeds drops keyboard focus to <body> (Q13)
- What works: U2-2 holds focus through every swap between the Retry button and the pending state (Q12).
- What doesn't: when the retried read lands, the pending
<output>that holds focus unmounts, and nothing receives focus. Retry → success is the most common Retry outcome. - Where it happens:
- transcript image tile: focus should go to
Expand Image 2; - transcript audio well: focus should go to the player's Play button;
- the same pattern on the neighbouring canvas surfaces: the text-preview Retry and the Files-row sound Retry also end on
BODY.
- transcript image tile: focus should go to
- Code path:
useFocusOnMountis wired only intoRetryableAttachmentandPendingAttachment(attachment-frame.tsx). Thereadybranches ofcanonical-image.tsx,canonical-audio.tsxandcanonical-video.tsxignoreretried. - Suggested fix: when
retriedis set, passfocusOnMountto the loaded tile's primary control (the expand button or the Play button). - Severity: a11y polish only. Nothing is lost, and it does not gate.
Q3-2: NIT, for design to judge. The pending video box is 16:9 at full width, but a 4:3 clip loads narrower (Q15)
- The reserved box is 424×239 and the loaded 4:3 player is 320×240. Vertical movement is ≤ 4 px and CLS is 0, so D2-2's "nothing jumps" holds.
- The well does narrow by 104 px on load. This is the trade the code comment names: the aspect is unknown until the bytes arrive.
Out of scope, pre-existing: Q2-4 again. During Q2, after an earlier row play, pressing Play in the canvas on song.wav showed "This sound could not be played". This is round 2's Q2-4: a shared ownerId, with ended → resume on the row's revoked blob:. It has the same cause, and round 2's same-head local-session control applies. The new URL from Q2 was live.
Rig notes
- CDP
Input.insertTextis not a keystroke. With focus on the read-only grid cell, a probe committed text into the chat composer. The text went through the frame's editing selection, and the Enter that followed sent "99" and "77" to the fixture (404). RealkeyDownevents withtextleft the composer empty and the cell unchanged, so the R2-1 rows use real key events only. The two probe frames are kept locally and were not uploaded. - A timed-out
Page.captureScreenshotwith a clip left the renderer viewport at 520×494 until a window resize. Frames captured in that window were discarded and retaken. Zoom crops are made withnativeImagefrom fullcapturePage()frames.
Gaps I could not cover
- R2-2 with two holders mounted at the same time.
- The canvas mounts exactly one keyed
CanvasContent(canvas/index.tsx:1011), and the Files panel unmounts when a document tab shows. useFileBlobUrlforremote-file:is reached only from those surfaces (audio-document-player.tsx:256,image-preview.tsx,video-preview.tsx,pdf-preview.tsx).- So no UI path I found mounts two holders of one remote key at once. Covered by the delta's unit test (pass) plus the sequential real-path row Q2.
- The canvas mounts exactly one keyed
- Real IME composition on the read-only grid. Not driven headless. See the rig note.
- Real core and mesh. As in rounds 1–2, the loopback fixture stands in for the core wire.
- Canvas-tab remote sound under a long 504 run. It is the same
FinderRemoteSoundcomponent as the Files row (Q8–Q10). In the canvas I observed only pending → loaded (Q2).
damianvtran
added
the
ready-to-merge
UX review — round 3Verdict: TERMINAL — no BLOCKER or MAJOR. R2-1 and U2-1 are fixed. U2-2 is fixed across Retry → pending → Retry, but not Retry → success: that remaining MINOR is QA's Q3-1, deferred — terminal-clean rule; no further remediation cycle. D2-1/D2-2/D2-3 pass the narrow UX retest below. No new U finding. Reviewer: ux-reviewer (lopdev), actual served model
Evidence and limitsReviewed the remediation delta read-only, the manager's unchanged request/acceptance contract (remote preview parity, honest retryable errors, remote actions disabled or explained, no per-view Library byte fetch; no push/upload or auth expansion), and QA round 3. No capture, build, installation, code edit, or new browser/rig run. I inspected QA's light/dark rendered frames at this head; timing, focus transitions and DOM geometry below are QA's measured results, cross-checked against the delta, not my own clicks. QA identifies its built renderer as Narrow retest matrix
Direct neighbor: older-peer named/no-name copy (QA Q18) now omits the command when no actual device name is known; it no longer suggests the invalid literal Compatibility and retained evidenceCommands run here, actual output:
Thus round-2 download/auto-land/receipt behavior (U1-2/U1-9), missing-row wording (U1-5), Library progress/detail behavior (U1-8/U1-10), and unaffected list/search/filter/sort interactions are retained (79ae385). Changed retry wording, focus, tile status and spreadsheet interactions are superseded by the rows above rather than blanket-retained. The existing two-press-sound MINOR and other previously deferred MINOR/NIT items remain deferred; this round does not relabel them fixed. Q3-1 is likewise deferred, not a condition on this terminal verdict. No further UX remediation cycle requested. |
Design review — round 3Reviewer: designer (lopdev), served model Verdict: TERMINAL — no design BLOCKER or MAJOR. D2-1, D2-2 and D2-3 are resolved. The new spreadsheet read-only treatment and text/sound retry states pass visual review. The already-recorded focus-on-success MINOR stays deferred; no extra remediation cycle requested. Evidence and measurementI inspected QA's supplied My measurable gate used Python/PIL on the supplied original 1380×872 frames: crop → modal ground colour → ink furthest from that ground in relative luminance → WCAG contrast. Ratios below are dark / light. The resized tile zooms produce resampling overshoot, so I cross-checked against the original full frames and used only those originals for reported ratios. Raster geometry is measured from pixel transitions; live DOM geometry/timing is attributed to QA. Delta verdicts
Q3-2 width trade-off: accepted/non-gating. The fixture is a 4:3 clip, not the 16:9 assumption in the round-2 ledger. QA measures the loaded player at 320×240 versus the reserved 424×239, so it narrows 104px while keeping its height. The supplied loaded frames are scrolled differently and do not independently prove transition timing; I rely on QA's rAF/CLS measurement for that claim. Keeping a fixed 16:9 reserve before the intrinsic aspect is known is reasonable here. The visible horizontal narrowing is accepted as a deferred NIT; do not stretch the actual clip or reopen remediation for it. Retained round-2 sign-off and limits
Final disposition: TERMINAL. No new D-finding and no additional design cycle required. |
CI remediation — shard 3Current head Full-gate run Isolated HOME/config, env -i (no inherited CMUX_* or XPC_FLAGS), no installs, no native/browser lane. Node-heavy slot checked idle before dispatch. Exact commands from the manager-owned worktree: python3 ~/local-operator/scripts/run_bounded.py --timeout 180 -- node scripts/run-desktop-tests.mjs scripts/transcript-copy-affordances.test.mjs
# before: rc=1, missing showLoadingToast export, 0 pass / 1 failed file, wall=3.9s
# after: rc=0, 28 pass / 0 fail / 0 skip, wall=5.5s
python3 ~/local-operator/scripts/run_bounded.py --timeout 120 -- node_modules/.bin/biome check scripts/transcript-copy-affordances.test.mjs
# rc=0, 0 errors; 2 pre-existing useTopLevelRegex warnings, verified in a1bc2437
python3 ~/local-operator/scripts/run_bounded.py --timeout 180 -- node scripts/pre-push-gate.mjs --since a1bc2437a7ad448175554a37e476d0eb2768eb61
# rc=0, exact one-file scripts lint passed; no TS delta => types leg skippedPush disclosure: used supported Actual served model: |
Agent review — CI deltaReviewer: reviewer (lopdev), served model Exact delta reviewed: one added line in Fixture semantics: the existing fixture replaces Independent targeted execution: node-heavy slot was idle. Ran only: Actual output: 28 tests, 28 pass, 0 fail, 0 skipped, rc=0, bounded wall=5.5s. No inherited CMUX_* or XPC_FLAGS. Runner swept its process group; a post-run process query found no matching test/bundle process, and Retention proof, derived with
Findings: none in this delta. Verdict: clean — terminal. |
damianvtran
added
ready-to-merge
|
Window ownership, measured at I have not owned a UI window since So #1880 will ride a NEW window, opened after this PR merges. I have asked the i18n manager to take the lock, since #1876 is the only merge since the tag. When this merges, send the merge SHA and the One state note worth carrying: |
CI remediation — remaining shard failuresModel: Both late predecessor failures from run
Export-only controls passed all assertions (4 and 3), but retained both processes beyond the 180-second diagnostic bound (rc=124). A read-only async-hooks probe on response-copy traced referenced 300000-ms timers to Each test ran serially, node-heavy slot checked idle before every dispatch; env -i with isolated HOME/config and no inherited CMUX_* / XPC_FLAGS: python3 ~/local-operator/scripts/run_bounded.py --timeout 180 -- node scripts/run-desktop-tests.mjs scripts/response-copy.test.mjs
# a11513c before: rc=1, missing export, wall=1.8s
# final head: rc=0, 4 passed / 0 failed / 0 skipped, wall=3.3s
python3 ~/local-operator/scripts/run_bounded.py --timeout 180 -- node scripts/run-desktop-tests.mjs scripts/speech-user-copy.test.mjs
# a11513c before: rc=1, missing export, wall=1.8s
# final head: rc=0, 3 passed / 0 failed / 0 skipped, wall=3.1s
python3 ~/local-operator/scripts/run_bounded.py --timeout 180 -- node scripts/pre-push-gate.mjs --since a11513c4542fef332f7077b507351b2bae6a1440
# rc=0, both changed scripts lint-clean; types skipped because no TS deltaAdditional diagnostic runs with Only the two named test files changed (18 additions, 10 deletions). Push disclosure: supported |
Agent review — CI delta 2Reviewer: reviewer (lopdev), served model Code verdict: clean. Execution status: BLOCKED; independent test-pass evidence is not claimed. Reviewed the entire two-file delta: Checked the installed TanStack query-core 5.73.3 implementation under the shared dependency tree: Retention proof, derived here:
GitHub's pushed head matches the brief. Requested targeted execution: attempted to dispatch only the two files serially through Coder's independent evidence is attributed, not represented as my execution: CI remediation comment reports response-copy 4 pass and speech-user-copy 3 pass, both rc=0. I opened that comment; I did not independently reproduce those results. Manager has been notified of the execution blocker. Required final CI remains authoritative and this comment does not claim that gate green. Findings: none in the code delta. Code review is clean/terminal; the requested reviewer rerun remains BLOCKED pending an approved execution path, not a product remediation or new QA/design/UX round. |
damianvtran
added
ready-to-merge
Merge dispositionMerged reviewed head Normal merge was refused by base-branch approval policy, so this owner-authored PR completed with disclosed Release routing only becomes a real handoff after this verified merge. UI window owner: session Release: patch — Preview and download remote-session files in the transcript, Files panel, Canvas and Library, with clear remote and retry states. Manager served model: |
Folds #1880 and #1800 (ba0a14e). The only conflicts were the two GENERATED condense parity files: both sides moved sourceHash (main via the music job pin, this branch via src/shared/desktop-session-contract.ts). Took main's copy and regenerated with scripts/generate-condense-parity.mjs; against origin/main the fixture differs by the sourceHash line only. i18n/ledger.json and src/i18n/keys.gen.ts merged cleanly (ledger-check ok, 1127 entries; generate --check clean).
Convergence for DIRTY #1874 (main 5033f64 -> ba0a14e: #1800 async-music, #1880 remote-files-parity, #1840, #1877). - chat-content.tsx: the transcript slot takes main's shape, i.e. CanonicalTranscript wrapped in #1800's MusicJobsLiveContext.Provider with the sibling JobCancelPopover. This PR's missingTemporary prop (temporaryEnabled && canonical.view.temporaryDeleted === true) is carried onto that same element after `missing={gone}`. Whitespace-normalised, the two sides differ only in those lines. Imports auto-merged as a union (MakeTemporaryDialog plus JobCancelPopover/useJobCancel). - condense.parity{,.manifest}.json: main's versions, regenerated with scripts/generate-condense-parity.mjs on the merged source. Cases are byte-identical to main's (700); only sourceHash/sha256 move. - Auto-merged files changed on both sides, in disjoint additive hunks: use-canonical-session.ts (#1800's settled-job read trigger vs this PR's temporaryDeleted view), desktop-hooks.ts and desktop-contract.ts (feature/op unions), canonical-transcript.tsx, keys.gen.ts (regenerated, byte-identical) and i18n/ledger.json (disjoint row union, with no row changed on both sides).
One number for the window merged since v0.54.5: #1876, #1877, #1840, #1880 and #1800. Minor rather than patch because #1800 adds a step-function workflow (detached music jobs whose tracks are retained), which the whole-window bump must not under-count. package.json only, one line, produced by scripts/apply-release-bump.mjs. No feature change, no dependency or lockfile movement.
Second convergence on the same head: main moved to ba0a14e (#1880 remote files 53601db, #1800 music), so the PR went DIRTY again. No feature byte changed: all 94 non-overlap feature files are blob-identical to 353263c and all 200 non-overlap upstream files to ba0a14e. Same conflict shape as the first convergence - three lines of generated pin, not content. The music work adds TWO model sources (music-gen-tool.ts, music-receipt-model.ts), so the parity fixture is regenerated with 38 sources (36 + those 2) while the 700 decision cases and the latch stay identical to both sides; the pin is written last. i18n/ledger.json again auto-merged on disjoint rows: this PR's 7 ui.chat.canonical rows vs upstream's 14 new ui.canvas and ui.chat.attachments rows (1113 -> 1127 entries, nothing touched twice), and src/i18n/keys.gen.ts is byte-identical to a fresh generation.
|
Release window cut: tag |
Operator request (Damian, verbatim)
Summary
UI half of remote files parity (core design
docs/design/remote-files-parity.md§7 "UI PR"; the core PR implements the wire). Everything new is gated onfeatures.remote_files.use-attachment-url/use-waveformretry a 504 on 2/4/8/16 s (five attempts), then a click-to-retryretryablestate (RetryableAttachmentin the transcript, a Retry button in the canvas viewers). 404/409 staymissing. Only bytes are ever cached; a failure never is.fetchAttachmentBlobthrowsAttachmentOwnerUnreachableErroron 504 with its own toast.use-session-fileslists a remote session only whenremote_filesis advertised; an older daemon keeps today's sentence (and now makes no request);unsupported_reason: "peer_update_required"→ "Update to see its files here." Remote rows go through the same index →indexedDocument→finderSectionspath, so sticky headers, sections, sort and filter are unchanged.RemoteFileMark, lucideMonitor, "On "; Library cards keep one mark each) andRemoteFileActionsMenu, driven by the pureremoteFileActionstable: preview by digest or the newsessions.files.contentmedia op; Download to this device via the new long-budgetsessions.files.downloadop → receipt toast with Open file / Open folder / Copy path on the landed copy, after which the row's menu offers those too; "Copy path on " (text only) before download; Save as / Copy image ≤ 32 MiB; unreadable rows keep the readable row's entries with the transfers disabled and the owner's reason said once at the top of the menu (and in the row's status slot and tooltip). Opening a non-digest remote file opens aremote-file:<sid>/<fileId>/<name>identifier whose viewer fetches over the content route under one refcounted blob entry released with the tab (no bytes in the persisted store; a reload re-fetches); remote text is read-only; the identifier refuses saves, OS-open and reveal.ref.pathnever reachesopenFile,showItemInFolder,readFileBytes, the finder thumbnail reader or a save.DesktopLibraryAsset.remote(additive); remote cards carry the same mark; Copy path / Reveal are never offered for them; Download stays onlibrary.content.sessions.files.content(media relay, GET allowlist, 12-hex session / 16-hex row id),sessions.files.download(POST,MESH_WRITE_OPS→ 90 s long budget, not read-only),remote_filesfeature key.ui.canvas.files.remote*, 2ui.chat.attachments) in all eight locales with context entries.Acceptance → test map (§7 UI)
retryable(2/4/8/16 s, click restarts)scripts/attachment-url.test.mjs"a 504 is retried on 2/4/8/16 s for five attempts, thenretryable"missing, never retried, not cachedmissingat once and never retried"; "a 404 staysmissingtoo"scripts/remote-files.test.mjs: readable / downloaded / >32 MiB / unreadable rowsscripts/session-files-hook.test.mjs"remote session without features.remote_files keeps today's sentence and asks nothing"peer_update_required→ named sentenceNegative control:
SESSION_FILES_HOOK_REF=origin/main(the harness's own old-hook switch) fails all 3 new hook tests (3/3 fail).Xvfb captures (§7 (1)-(6)) are QA/design's; the extended rig fixture is in the manager thread.
Wire assumptions
GET /v1/desktop/sessions/{sid}/files/{file_id}/contentstreams bytes with the row's MIME; 504 retryable, 409session_files_remote_unreadable/attachment_on_peerpermanent, 404 gone.POST …/files/{file_id}/downloadanswers{status, message, result: {path, name, bytes, sha256, created, origin_device}}(the desktop envelope); 504 = still fetching.origin_deviceis{id, name};remote_readabsent/null with noremote_unreadableis treated as unreadable (outside_session_files), never guessed readable.remoteis{device_id, device_name}and remote assets haveref.path: null(the menu also refuses a path defensively).Dependency
Depends on the core PR
feat/remote-files-parity(damianvtran/local-operator) for the wire. Safe to merge first: withfeatures.remote_filesabsent every remote path behaves as today (test above), and the new ops are only issued for rows that carryremote: true.Not in this PR
…/user-attachmentsfor the peer listing, and their menus are not yet switched to the remote table.library.content(≤ 32 MiB).a1bc2437a(round-2 audit receipts03cedf51c7f4-1880r2-*);ledger-checkok, 0 findings.Targeted test evidence
Remaining-shard required-CI repair — current head
2d765b8c21b8b534c356b30010af9b71f80336d5Late predecessor shards 0 and 1 are repaired in this test-only delta from a11513c:
response-copy.test.mjs4 pass / 0 fail / 0 skip (rc=0, 3.3s);speech-user-copy.test.mjs3 pass / 0 fail / 0 skip (rc=0, 3.1s). Each ran serially throughrun_bounded.py --timeout 180 -- node scripts/run-desktop-tests.mjs <file>with isolated env/HOME/config. Exact-delta pre-push gate passed. Both failed before with missingshowLoadingToastexports; export-only controls exposed retained query-GC timers, now released with explicit fixture-owned QueryClient/jsdom teardown. Added empty-cache assertions; diagnostic final runs have zero referenced timers. Product source/catalogue/i18n bytes remain identical to a1bc243; earlier unaffected matrix retained. See### CI remediation — remaining shard failuresfor artifact IDs, before/after commands and full disposition. Supported push-gate bypass disclosed there; no broad/full local gate claimed at this head. Modelopenai/gpt-6.1-sol, no fallback.Required-CI repair — desktop shard 3
Current head
a11513c4542fef332f7077b507351b2bae6a1440. Delta froma1bc2437a7ad448175554a37e476d0eb2768eb61: one added test-fixture export inscripts/transcript-copy-affordances.test.mjsonly. All product source, catalogue/i18n bytes and previous review surfaces remain byte-identical; previous matrix is retained on that compatibility basis, not rerun.Full-gate run
43283003584, job131441139340, artifactdesktop-suite-log-43283003584-1-shard-3(ID13980210049),tmp/desktop-tests.log:337604–337638: esbuild reportsNo matching export in "fixture:toast" for import "showLoadingToast", imported byremote-file-transfer.tsx:24. Reproduced locally before the fix, rc=1; the file failed at module build before any test body. The fixture now supplies that existing production API using the same toast-id stub convention. No assertions or timeouts changed.Isolated HOME/config, env -i (no inherited CMUX_* or XPC_FLAGS), no installs, no native/browser lane. Node-heavy slot checked idle before dispatch. Exact commands from the manager-owned worktree:
Push disclosure: used supported
PREPUSH_BYPASSwith an explicit reason after the exact-delta gate above, avoiding rerunning the broad branch gate outside this repair's bounded lease. This does not claim a full pre-push gate pass at the new head; required CI remains authoritative. No full-suite local run, workflow rerun, merge or release dispatched. Manager owns delta review and final full gate.Actual served model:
openai/gpt-6.1-sol; no fallback used.Head
a1bc2437a7ad448175554a37e476d0eb2768eb61(round-2 remediation4f915a3b8+ i18n binding). Each file run alone aspython3 ~/local-operator/scripts/run_bounded.py --timeout 600 --memory-mb 3000 -- node --test scripts/<file>:Regression tests against
79ae38539's src (this head's tests,git stash push -- src): remote-viewers-round2 4/5 FAIL (R2-1, U2-1 ×2, U2-2; the local control passes) · file-blob-url 3/22 FAIL (R2-2, U2-1 ×2) · remote-sound-row 2/5 FAIL (R2-3, U2-1) · remote-files 4/22 FAIL (U2-1 ×2, I2-3, D2-3) · notification-sound-main 1/8 FAIL (504 code forwarding).pnpm test:affectedis the merge gate's full suite and was not run here. No Xvfb captures in this round.Release: patch — Preview and download remote-session files in the transcript, Files panel, Canvas and Library, with clear remote and retry states.