You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Concierge does not respond when ask: expenses broken down by category in offline mode #99060
Version Number: 9.4.57-0 Reproducible in staging?: Yes Reproducible in production?: Yes If this was caught during regression testing, add the test name, ID and link from BrowserStack:#98621 Email or phone of affected tester (no customers):emad12@gmail.com Issue reported by: Applause Internal Team Bug source: Pull Request QA execution Device used: Windows 11 / Chrome App Component: Other
Action Performed:
Open staging.new.expensify
Go offline
Open Concierge chat in a workspace with expenses in VND.
Ask "expenses broken down by category".
Expected Result:
As per PR instructions : chart rendering doesn't depend on network state so it should show the results.
Actual Result:
Concierge does not respond when ask:expenses broken down by category in offline mode
Root Cause: This is expected behavior of the offline-first architecture, not a regression from PR #98621 (which only added a font-glyph fallback for the ₫ symbol in the chart renderer — it changed nothing about network handling). Concierge is a server-side AI: asking a brand-new question sends an ADD_COMMENT write that is persisted to the offline queue and held until you're back online, so the server never receives the question and never generates the reply. The donut chart is server-authored HTML (<victorychart> tags) that the client only renders; with no server reply to render, nothing appears.
The confusion comes from PR #98621's offline test note — "chart rendering doesn't depend on network state so it should show the results." That's true for re-rendering an already-received chart (e.g. scrolling back to it offline), but not for generating a new one. QA tested the latter, which can never work offline.
Recommended Fix: No rendering bug to fix. The only actionable improvement is a product/UX decision: whether Concierge should surface explicit "you're offline" feedback when a question is queued, instead of silently showing a "thinking" placeholder. That is a design call, not a defect — recommend closing this as expected behavior or converting it to a UX-improvement request.
Investigation details (evidence)
A freshly typed question goes through addComment → API.write(ADD_COMMENT), which builds only the user's optimistic comment (no client-side answer): src/libs/actions/Report/index.ts:1262.
The write is persisted and handed to the SequentialQueue, which returns early while offline — the request is never transmitted: src/libs/Network/SequentialQueue.ts:274.
Alternative explanations considered: (1) A regression from #98621 — ruled out; that PR only touches getChartSkiaTypeface / label rendering, never the network or comment paths. (2) A client-side chart computation that should work offline — ruled out; the chart data is embedded in server HTML, not derived on-device.
Related context:#97341 is the original ₫-glyph exploratory issue that PR #98621 fixed — related, but a different bug (rendering vs. offline generation).
Next Steps for Contributor+ team: To accept:@MelvinBot implement [this](https://github.com/Expensify/App/issues/99060) to create a draft PR. To refine:@MelvinBot <your feedback> To reject: Explain why you are rejecting Melvin's proposal.
changed the title [-]Concierge does not respond when ask:expenses broken down by category in offline mode[/-][+]Concierge does not respond when ask: expenses broken down by category in offline mode[/+]on Aug 20, 2026
@brunovjk before this gets closed as expected behavior, I want to flag one gap in the automated analysis above.
The RCA is correct on the mechanism: Concierge is server-side, the donut chart is server-authored <victorychart> HTML the client only renders, and offline the ADD_COMMENT write never leaves the SequentialQueue, so no reply and no chart can be produced. The QA step tested new-chart generation offline (which can't work), not re-rendering an already-received chart, so the PR #98621 offline note was misread. All of that holds.
Where I disagree is the conclusion. The analysis dismisses the only real defect here as "a product/UX decision," but there is a concrete, deterministic, fully client-side fix that needs no new design call: right now, asking Concierge something while offline gives zero feedback. The message is silently queued and the chat stays blank, which is exactly why this read as a broken response to the tester. Every other server-dependent action in the app surfaces offline feedback; Concierge is the outlier that shows nothing.
The fix fits an existing slot. ConciergeThinkingMessage already owns the "waiting for Concierge" area below the last message. When we're offline in a Concierge conversation and the newest action is the user's own turn (a reply is pending), it renders a small offline hint instead of nothing:
offline + the report is the Concierge DM or a thread off it (reportID === conciergeReportID || parentReportID === conciergeReportID, the same check the thinking indicator uses) + the newest action's actor is the current user → show a translated line like "You're offline. Concierge will respond once you're back online."
the newest-action-is-the-user check matters because Concierge answers in a thread: there the user's question lives on the parent as an ADDCOMMENT while the thread's own newest action is the CREATED action that opens it — both are authored by the user, so keying on the actor (not the action type) covers the DM and the thread.
on reconnect the queued question flushes, the server responds, and the normal thinking indicator + chart take over unchanged; once the reply lands it's the newest action so the hint clears on its own.
This is passive (never blocks the offline-first send path), scoped to Concierge, and removes the confusing silent no-op without any product/design dependency.
I've implemented this locally and verified it offline with a before/after recording (before: silent no-op; after: the hint appears). Happy to post the full proposal with the diff and test cases if you think it's worth pursuing, or hold off if you'd rather just close it as expected behavior. Let me know which way you'd like to go.
Here's the fix working locally — same offline "expenses broken down by category" flow, before and after.
@mukhrr I don't think a BE change can address this one, and it's worth being precise about why.
While the user is offline the request never reaches the server — ADD_COMMENT sits in the client SequentialQueue and is never transmitted. So the backend has no signal to act on and no channel to reach the device. Any "you're offline, this will be answered once you reconnect" feedback therefore has to be rendered client-side; there's nothing the server can do for a device it isn't talking to.
The chart-generation half is genuinely server-side, but that's exactly the part that is expected not to work offline (it's a server-authored <victorychart>), so there's no BE bug to fix there either — it already behaves correctly once you're back online.
Happy to be proven wrong if you have a specific BE path in mind, but as far as I can trace it the only actionable defect (the silent no-op) is a frontend one.
Activity
applause-bot commentedon Aug 20, 2026
While Applause doesn't embed attachments for this bug, here are the links for quick view:
https://platform.applause.com/services/links/v1/external/50617fa30006965862d6fb4691c4db53ed3948542a6507eaa84abfea6f620656
MelvinBot commentedon Aug 20, 2026
Issue Analysis
Root Cause: This is expected behavior of the offline-first architecture, not a regression from PR #98621 (which only added a font-glyph fallback for the ₫ symbol in the chart renderer — it changed nothing about network handling). Concierge is a server-side AI: asking a brand-new question sends an
ADD_COMMENTwrite that is persisted to the offline queue and held until you're back online, so the server never receives the question and never generates the reply. The donut chart is server-authored HTML (<victorychart>tags) that the client only renders; with no server reply to render, nothing appears.The confusion comes from PR #98621's offline test note — "chart rendering doesn't depend on network state so it should show the results." That's true for re-rendering an already-received chart (e.g. scrolling back to it offline), but not for generating a new one. QA tested the latter, which can never work offline.
Recommended Fix: No rendering bug to fix. The only actionable improvement is a product/UX decision: whether Concierge should surface explicit "you're offline" feedback when a question is queued, instead of silently showing a "thinking" placeholder. That is a design call, not a defect — recommend closing this as expected behavior or converting it to a UX-improvement request.
Investigation details (evidence)
addComment→API.write(ADD_COMMENT), which builds only the user's optimistic comment (no client-side answer):src/libs/actions/Report/index.ts:1262.SequentialQueue, which returns early while offline — the request is never transmitted:src/libs/Network/SequentialQueue.ts:274.src/components/HTMLEngineProvider/HTMLRenderers/index.ts:73.src/libs/actions/Report/SuggestedFollowup.ts:75-126.Alternative explanations considered: (1) A regression from #98621 — ruled out; that PR only touches
getChartSkiaTypeface/ label rendering, never the network or comment paths. (2) A client-side chart computation that should work offline — ruled out; the chart data is embedded in server HTML, not derived on-device.Related context: #97341 is the original ₫-glyph exploratory issue that PR #98621 fixed — related, but a different bug (rendering vs. offline generation).
Next Steps for Contributor+ team:
To accept:
@MelvinBot implement [this](https://github.com/Expensify/App/issues/99060)to create a draft PR.To refine:
@MelvinBot <your feedback>To reject: Explain why you are rejecting Melvin's proposal.
view run
melvin-bot commentedon Aug 20, 2026
Triggered auto assignment to Contributor-plus team member for initial proposal review - @brunovjk (
External)melvin-bot commentedon Aug 20, 2026
Job added to Upwork: https://www.upwork.com/jobs/~022090402162801591040
[-]Concierge does not respond when ask:expenses broken down by category in offline mode[/-][+]Concierge does not respond when ask: expenses broken down by category in offline mode[/+]David-2002065 commentedon Aug 20, 2026
@brunovjk before this gets closed as expected behavior, I want to flag one gap in the automated analysis above.
The RCA is correct on the mechanism: Concierge is server-side, the donut chart is server-authored
<victorychart>HTML the client only renders, and offline theADD_COMMENTwrite never leaves theSequentialQueue, so no reply and no chart can be produced. The QA step tested new-chart generation offline (which can't work), not re-rendering an already-received chart, so the PR #98621 offline note was misread. All of that holds.Where I disagree is the conclusion. The analysis dismisses the only real defect here as "a product/UX decision," but there is a concrete, deterministic, fully client-side fix that needs no new design call: right now, asking Concierge something while offline gives zero feedback. The message is silently queued and the chat stays blank, which is exactly why this read as a broken response to the tester. Every other server-dependent action in the app surfaces offline feedback; Concierge is the outlier that shows nothing.
The fix fits an existing slot.
ConciergeThinkingMessagealready owns the "waiting for Concierge" area below the last message. When we're offline in a Concierge conversation and the newest action is the user's own turn (a reply is pending), it renders a small offline hint instead of nothing:reportID === conciergeReportID || parentReportID === conciergeReportID, the same check the thinking indicator uses) + the newest action's actor is the current user → show a translated line like "You're offline. Concierge will respond once you're back online."ADDCOMMENTwhile the thread's own newest action is theCREATEDaction that opens it — both are authored by the user, so keying on the actor (not the action type) covers the DM and the thread.This is passive (never blocks the offline-first send path), scoped to Concierge, and removes the confusing silent no-op without any product/design dependency.
I've implemented this locally and verified it offline with a before/after recording (before: silent no-op; after: the hint appears). Happy to post the full proposal with the diff and test cases if you think it's worth pursuing, or hold off if you'd rather just close it as expected behavior. Let me know which way you'd like to go.
Here's the fix working locally — same offline "expenses broken down by category" flow, before and after.
Before (offline → silent no-op):
99060-before.mp4
After (offline → hint shown):
99060-after.mp4
mukhrr commentedon Aug 20, 2026
i think this need a BE fix
David-2002065 commentedon Aug 20, 2026
@mukhrr I don't think a BE change can address this one, and it's worth being precise about why.
While the user is offline the request never reaches the server —
ADD_COMMENTsits in the clientSequentialQueueand is never transmitted. So the backend has no signal to act on and no channel to reach the device. Any "you're offline, this will be answered once you reconnect" feedback therefore has to be rendered client-side; there's nothing the server can do for a device it isn't talking to.The chart-generation half is genuinely server-side, but that's exactly the part that is expected not to work offline (it's a server-authored
<victorychart>), so there's no BE bug to fix there either — it already behaves correctly once you're back online.Happy to be proven wrong if you have a specific BE path in mind, but as far as I can trace it the only actionable defect (the silent no-op) is a frontend one.
brunovjk commentedon Aug 20, 2026
Reviewing ASAP. Thanks.
luacmartins commentedon Aug 20, 2026
This is expected. Closing.