Repository navigation
feat(j5): platform notices and prompts in the timeline on web - #916
Conversation
Platform notices are now one grey line that opens to its detail, and platform prompts an outlined pill, both drawn from the shared classifier. The raw A2A card is gone: a notice J5 can't read is upstream's row. Closes #698 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
|
No actionable comments were generated in the recent review. Walkthrough
Merge Risk:
|
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Upstream's row offers copy on these messages, and long reports fold under "Background task completed", so they stay copyable. D38 now describes what's built, and crews.md points to the look and loses stray conflict markers. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Upstream draws PR-watch updates, restarts, finished tasks, delegated questions and idle compaction on its own notification and compaction rows, so they never reach the user row. Group 4 is J5's notices, and group 5 is automation runs, playbook steps and native questions. A test now takes each kept kind through the timeline's own derivation. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Jacksondr5
left a comment
There was a problem hiding this comment.
Combined review at 5e2aab3042 (Opus lead and Sol 6.1)
No blocking code findings. One visual fix to make, and one evidence gap to close or accept. Both reviewers read the source independently and agreed these findings.
Should fix
A folded pill cuts off its own label. The folded playbook capture reads "Playbook step 1 ..." and "Playbook step 2 o..." with room to spare, and opened it reads "Playbook step 1 of 2". The likely cause is that FoldedMessageHeader gives the label max-w-1/2 shrink-0 truncate, and a pill sized to its content has a short half when the title is short. "Automation" fits, so upstream never hits this.
It hides the step count, which is what the label is for. By source reading, "Question from <title>" would lose the sender's name the same way, though no capture shows that.
It's purely visual, so new captures are the proof and no test is needed. The smallest fix we can see stays in the existing case 7 hunk, for example (expanded || j5Prompt) && "w-full" on the pill. Any fix that adds no new upstream hunk is fine. Replace the affected playbook captures afterwards.
Evidence gap
The captures cover four kinds: silence, declined roster, automation and playbook step, each folded and open. None of the three notices that motivated #698 (not delivered, Exchange dropped, ask withdrawn) has an image, and neither does the fallback for an unrecognised notice. The PR says this honestly, and the brief allows tests-only for a kind that can't be reached, so this isn't a code finding.
Two of these look cheap to capture:
- Exchange dropped: have an agent ask a peer, then archive the peer.
LifecycleServicesends the asker the notice. - Unrecognised notice: the live-test home calls
runDevDeliverySeed, which seeds atemplate-v999notice.
We'd include both in the extra tester pass. Whatever stays uncaptured is Jackson's call when reviewing the UI.
Checked and fine
- Design and scope. Groups 4 and 5 match the narrowed scope: J5's notices are lines, and Automation, playbook step and native question are pills. The PR body flags where this differs from #698 and the design. Leaving upstream's rows alone narrows Jackson's 2026-10-24 decision, so that's for Jackson to confirm.
- Upstream's rows. A
notificationturn item draws as a work row (session-logic.ts:608), including delegated completions and delegated-task questions. A/compactmessage with arunIddraws as a compaction row. Neither reachesUserMessageRow, andtimelineRows.test.tscovers both. - #750's rule.
isHiddenCrewNoticeandwithoutHiddenCrewNoticeshave only import changes.crewNoticeLineskeeps its seat and tone logic and only splitsreasonfromdetail, so the same notices stay hidden. - Nothing raw shows. Participant ids in notices go through the identity read, and an unknown one reads "Unnamed participant". The playbook step's tagged block is stripped from the body, copy and prompt link. Crew context still opens to the text the model read, as designed.
- Fallback. A notice the classifier can't read returns null and draws as upstream's row, and "Raw A2A delivery" is gone from web.
- Real data. An earlier version of the classifier ran read-only over all 9,807 user-role messages in a live database with no raw fallback. That shows what the classifier reads, not what gets drawn, and it wasn't re-run after the rework.
timelineRows.test.tscovers drawing, from stored turn item through the timeline derivation. - Find. A hidden detail stays mounted while find is active and opens on a match. The tester confirmed this live on a silence detail and a playbook prompt.
- Animation. No continuously repainting animation; the only transition is the chevron rotating.
- Mobile. No file under
apps/mobilechanged,timelinePresentation.test.tspasses unedited, andtsconapps/mobilereports no errors. The PR lists what mobile should follow. - Upstream edits. 5 hunks in
TimelineUserMessageRow.tsx, one of them the import, all recorded in FORK case 7. That's over the design's estimate, and reusingAutomationMessageRowis a good reason: it keeps fold, find, copy, the steer marker, status and Edit from here without exports or an import cycle. - Captures after the rework. The commits since
82f0c24079only removeupstreamRowand unreachable readers, so the captures still hold for how those rows draw. The label fix will change the playbook ones. - Docs.
thread-a2a-rendering.mdand D38 are rewritten to what's built,crews.mdpoints to the rendering doc, and the leftover conflict markers on trunk are gone.
Tests run on this head: 117 passed across the six focused files. CI was still pending when this was written.
Reviewed by Claude Opus 5.5 and GPT Sol 6.1.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Jacksondr5
left a comment
There was a problem hiding this comment.
Follow-up at 51afeadb84 (Opus lead)
No open code findings. CI is green. This is my own check of the changes since the combined review and of seven new captures. Sol 6.1 reviewed the code up to 5e2aab3042, not these captures.
- Label fix: confirmed. Folded pills now read "Playbook step 1 of 2" in full. A very long "Question from …" title still truncates inside the pill, which is reasonable, and the full title stays reachable.
- CodeRabbit fixes: both fine.
J5NoticeLine.browser.test.tsxuses the default router, and an unreadable platform message now sits on the left. Thetemplate-v999capture shows its text visibly in upstream's bubble. - New captures: native question, Exchange dropped (named, no ids) and the amber "report missing" line with Open seat thread all match the design.
For Jackson's UI review, none of these block merge:
- An open native question shows raw ids. The body reads "Placed agent (threadId thread:…, requestId runtime-request:…)", which fills several lines. The design asks for "the thread and request to read", and these aren't participant ids, so it meets AC12 as written, but it reads badly. Tidying it would be a small change in
nativeQuestionPromptif wanted. - Five kinds are tests-only: queued removed, not delivered, Exchange withdrawn, handoff reminder and the red Crew line. That includes two of the three notices that motivated #698. Withdrawn is sent only to a receiver on a peer server (
SendService), so it needs a peered setup, which is why a localj5_clear_own_askdidn't produce one. - Scope is narrower than the design. Upstream's platform messages stay on upstream's rows, as the PR body says.
Reviewed by Claude Opus 5.5.
…s seat lines (#981) * fix(mobile): platform notices read as named lines, and Crew notices as seat lines Mobile follows #916's classifier (#918, #836 piece 2): each J5 notice is one grey line on the left that opens to its detail, with participants named; a Crew notice shows its approval or decline and one line per seat that needs the person, each opening the seat's thread; playbook steps and native questions take upstream's Automation row under J5's heading. Crew notices with nothing for the person are dropped after the feed is derived, through the filter web now shares. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(j5): FORK's ThreadFeed.tsx row lists case 104 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(mobile): the Automation row's label no longer truncates beside its summary Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Problem
J5's platform notices had no shared look on web. Three of them (not delivered, Exchange dropped, ask withdrawn) showed their raw envelope in a "Raw A2A delivery" card. Closes #698.
What changed
Groups 4 and 5 of the agreed timeline design, for the messages that actually reach a user row, drawn from the shared classifier.
<name>'s turn ended without replyingand its variants<name> never read a queued message: it was removed before it ranMessage to <name> not delivered/may not have been deliveredExchange with <name> dropped/<name> withdrew its ask<seat> couldn't start, red or amber iconCrew approved · n members/Crew <name> · roster declined by <who>Introduction from J5/Roster changed from J5Reminded to write its handoffThis differs from #698 and the design:
notificationturn items (notificationTurnItem), and web draws/compactas a compaction row (MessagesTimeline.logic.ts), so they never reach this row. The live database agrees: 0 PR-watch and 0 restart user rows against 446 and 3 notification rows. Restyling them would be a new edit to upstream's work rows that Jackson never saw, so the Captain chose to follow upstream and is telling Jackson; he can reverse it on reading this.A bug this fixes: trunk's queued-removed parser never matched the server's real
Your message <id> to X, so those notices fell to the raw card. It now reads both forms.Docs: besides the rewrites, this removes leftover merge conflict markers from the History section of
crews.md, with every entry kept as written.UI changes
Matched captures use the same isolated test home on trunk and head. Unchanged silence, declined roster and automation captures were rechecked on
51afeadb84; playbook captures were replaced for the full-width label fix. Native question, Exchange dropped, amber missing-report and fallback captures are from51afeadb84. The base lacks folds for several of these rows, so its single before view is reused for folded/open after states.Opening a playbook step, before:
playbook-expansion-before.mp4
Opening a playbook step, after:
thread-j5-a2a-bdfbb3dc26f227374b81748feacd67912c7847ffc9d77eb81622199e4a2369a0-dfff17a2-1652-4326-af45-f8b6a9c35cd8-mp4.mp4
Uncaptured: queued removed, not delivered, Exchange withdrawn, handoff reminder, red Crew failure. A very long native-question title truncates, by design.
Live testing
Verdict: partial pass; no defect remains demonstrated in the exercised rows. Latest isolated web pass:
51afeadb84; real agent turns used Claude Haiku 5.5. Base was exact trunkb85f2625d8, using the same preserved isolated home. Both test servers were stopped through setup's tracked process groups.Passed live:
template-v999notice uses upstream's bubble on the left, with no Raw A2A delivery card.Still not reached live; keep test-only coverage explicit:
No mobile, Electron, second-device, relay/tunnel or production-volume claim. Full report:
artifacts/test-reports/j5-698-tester/51afeadb84.md.Upstream impact
apps/web/src/components/chat/TimelineUserMessageRow.tsxAutomationMessageRow; that row takes J5's text, title, label and icon, and moves left without its fillThat's 5 hunks, one of them the import, against the design's "1–2 + class edit". J5's pills reuse
AutomationMessageRowrather than copying it, which keeps its fold, find, copy, steer marker, status and Edit from here without exporting them and avoids an import cycle. FORK case 7 records why.Product decision: moving the Automation row left and removing its fill is Jackson's design (2026-10-24, revised 2026-10-31), register D38, rewritten here to match what's built.
What mobile should follow
No file under
apps/mobilechanged, and its tests pass unedited. These are for the mobile orchestrator.detailandparticipantIds; prompts gainlabelandpreview; there are new kinds (not-delivered,exchange-dropped,exchange-withdrawn,handoff-nudge,playbook-step,native-question).nameJ5NoticeParticipantswith the identity read;participantIdsForJ5UserMessagenow returns those ids.was rolled_backnow readswas rolled back.crewNoticeLinesandcrewApprovalSummaryare now inclient-runtime, and a line's error moved fromreasontodetail.The classifier no longer has the kinds
restart-continuation,pr-watch,delegated-questionandcompaction. Mobile doesn't read them, and those messages behave there as before.Checklist
FORK.md(case text and file-table row) in this PRdocs/j5/product/upstream.mdthread-a2a-rendering.md,crews.mdandupstream.mdrewrittenBuilt by Claude Opus 5.5 in Claude Code.
🤖 Generated with Claude Code