Repository navigation
feat(chat): Quote / Ask on selected text in ChatPane (Members + split) - #8947
Conversation
Design Review (Fable 5) — ✅ PASSDesign-level review of Design-Verdict: PASS One chat-core seam, store-backed seeding, and host-owned capabilities — the right shape, with every failure path named, caused, and pinned. Suggestions
[DESIGN-REVIEWED] de63206 |
First Principles Review (Fable 5) — 🟡 CONCERNSPremise-level review of All verification is done. The final review follows the contract's output shape. First-Principles-Verdict: CONCERNS The Ask relabel across 13 catalogs and Quote's mirror tooltip rest on design taste — no named reader was failing on "Ask in Side Chat". Not justified as shipped
What this change shipsInventory (10 items) — 7 justifiedIntent: give Members threads and split-view panes the same Quote / Ask-on-selection the main chat has — an ADDITION.
More than 10 exist (app-sdk Watch
Subtractions
[FIRST-PRINCIPLES-REVIEWED] de63206 |
Opus 4.8 Review — ✅ no blocking findingsReviewed Review detailsBased on the candidate file, the discovery pass reported "No candidates." I independently reviewed the core code changes (the Verdict-gated seeding correctly withholds the seed on sync No findings. [OPUS-REVIEWED] de63206 Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
UX Review (Fable 5) — 🟡 CONCERNSUX-level review of Reconciliation is complete. The blind read covered every screenshotted control and got the whole flow right: Quote → "puts the text into my message box as a quotation" (correct, rated a guess), Ask about this → "opens the Side Chat with the text quoted" (correct via the shot 1→2→3 sequence, rated a guess), the off-record marker read as "sure", the draft-gated pill and the Details toggle both read correctly. The drawer details↔side swap has a stated reason, a crossfade, and a committed recording. The UX-Verdict: CONCERNS Cold read comprehended every control correctly — but the new error notice and the tooltips carrying Ask's destination were never shown to anyone. Watch
Evidence gaps
[UX-REVIEWED] de63206 |
GPT 5.6 Review — ✅ no blocking findings (all downgraded on adjudication)GPT 5.6 flagged blocking issues on This comment is updated in place on each push. Review detailsBLOCKING -- docs/feature-map/README.md:87 -- Side Chat endpoint does not exist
Adjudication (Opus 4.8) — is blocking on each finding proportionate?Line 87 ( Harm rung: LOW. The [ADJUDICATION] de63206 total=1 uphold=0 downgrade=1 False positive or not applicable? A repository writer can comment: |
148ebb4 to
75bf525
Compare
|
|
|
|
|
|
|
|
/ai-review override gpt 75bf525: max-two-buttons-per-row on the selection toolbar is a grandfathered row — SelectionToolbar already renders Quote / Ask in Side Chat / Copy for ChatPage on main via useSelectionActions; this diff adds no markup to that row, it only stops ChatPane withholding the two callbacks the shared component takes (same clause as #5895 FollowUpBar and #5892 QueueStack onEdit). |
Human judgment recorded@CrysisDeu marked the gpt AI finding as false positive, not applicable, or explicitly accepted for
This decision applies only to this commit. A new push requires a new judgment. |
75bf525 to
c0ebaf8
Compare
|
|
|
|
|
|
|
|
|
|
|
|
|
… view) Selecting text in a Crew Members thread (or a split-view pane) offered Copy only: AssistantMessage's selection toolbar draws the actions its host hands it, and ChatPage was the only host passing onQuote/onAsk — quote's FlyingQuote-into-composer and ask's open-/side-and-seed were ChatPage-local. Extract both into one chat-core seam, chat-core/composer/selectionActions (useSelectionQuoteAsk + quoteIntoDraft + seedSideChat), and make ChatPage and ChatPane both consume it — no second implementation. Hosts differ only in what they own: the composer draft, and how a Side Chat surface for a slot is brought on screen (openSideChat). Quote is always offered; Ask exactly when the host provides an opener (capability by omission, like onOpenFull). - app-sdk: ChatMessageList/MessageRenderContext gain onQuote/onAsk; the default assistant row passes them through. - ChatPane: wires the seam to its own composer (FlyingQuote lands in the pane) and takes an openSideChat prop. - Split view: SessionGridView threads ChatPage's opener, which re-binds the activity panel to the pane's slot (switchSlot) before opening the side tab. - Members page: the detail drawer gains a Side Chat view bound to the member's slot (reuses existing "Side Chat"/"Details" strings — no new copy). - SideChat: the side-seed event now names its slot; a SideChat bound to another slot ignores it, and the seed poll waits for the named slot's composer (data-side-chat-slot) rather than any Side Chat's. - ChatPage: only change is consuming the hook (behaviour unchanged) plus the one-prop grid wiring. Tests (mutation-oriented pins): selectionActions unit tests, ChatPane selection-actions test, MembersPage Side Chat drawer test, SideChat slot-aware seed, ChatPage Ask seed regression. Capture harness website/scripts/capture-members-selection-quote-ask.mjs with three asserted frames under temp-screenshots/members-selection-quote-ask/.
bolichen97
left a comment
There was a problem hiding this comment.
The two things I most expected to be wrong here are both fine, so let me put them first.
It does not intercept ordinary text selection. The floating popover already appeared in Members threads and split panes with Copy alone; this PR adds two buttons to an existing affordance rather than hooking selection for the first time. It portals at the mouse point +8px (selection rect for keyboard and touch), clamps to the viewport, flips above the anchor on bottom overflow, and dismisses on Escape, on outside mousedown, on a click that collapses the selection, and on picking either action. Copy alone keeps it open with a check.
Nothing sends on click. Quote pre-fills this surface's own composer with > -prefixed lines plus a trailing blank line; Ask also only pre-fills, opening the Side Chat (activity Side tab for main chat and split panes, the detail drawer on Members) with the selection quoted and the caret after it. Ask is a write into an in-memory per-slot draft store plus a focus nudge, and it is withheld entirely when the opener refuses (offline, deleted pane session).
Not stale either, which is worth stating since the branch is 20 commits behind: head de63206 is a clean rebase onto c0cf53f, the normalized patch is identical to the already-approved b3a89b0c except for context lines, and the rebase preserved main's newer onDisplayItems/hiddenRow props on the ChatMessageList call it edits. Main's newer drivingSessions change (#9728, MembersPage.tsx:532) sits clear of this PR's insertion at line 544, and main's docs/feature-map/README.md edit at line 169 is clear of this PR's at 74–87.
Four defects, all low:
-
docs/feature-map/README.md:87advertisesPOST /api/chat/slots/{slot}/side/*, which does not exist.routes/chat.pyregisters only/side/open,/side/turn,/side/close— which the Side chat row four lines above already spells out correctly. This is GPT's one BLOCKING finding, downgraded by Opus adjudication as disproportionate, and I agree with the downgrade: it is a doc wildcard. -
chat-core/composer/selectionActions.ts:53— the opener contract is fail-open. It is typedboolean | voidandvoidreads as "opened" (if (verdict === false) return,if (ok !== false)). All three current openers are correct, but a future host that can refuse and forgetsreturn falsewill silently strand a quote in an unopened slot's draft — which is the exact class this seam's own comments say earlier rounds fixed three times. Design raised it as a Suggestion. -
MembersPage.tsx:544—useSideChatDraft(activeSlot)subscribes the whole page to the draft store to decide one pill's visibility, andsideChatDrafts.ts:49mints a new snapshot object on every write. So every keystroke in the member's Side Chat re-renders all 2039 lines of MembersPage — roster, drawer, driving-session list — when the pill only flips at the empty↔non-empty boundary. No lane raised this. -
SideChat.tsx:537-547—consumeSideChatSeed(slot)sits outside theif (el)guard, so a frame where the textarea is not yet resolvable burns the seed and the caret is never placed. The focus is also unconditional with noisTouchDevice()skip, unlikecomposerFocus.revealComposer, which deliberately avoids focus on touch so the soft keyboard does not cover the content. Behaviour carried over from the deletedside-seedhandler, but it now reaches mobile Members threads and split panes for the first time.
Two things for the owner rather than the author, since they are scope questions rather than bugs. The PR also changes two existing surfaces the title does not name: the Ask button is relabelled across all 13 locales, and every Side Chat composer — including the main chat's — gains an accent tint plus a new "Side Chat · not saved to the conversation" label row. First Principles is 🟡 CONCERNS on exactly that, with two explicit Subtractions proposed (restore the old ask value in all 13 catalogs; delete the single-consumer quote_hint), and no disposition targets the current head de63206 — the latest ones target 8a039f3, 34dfc65, 75bf525. UX is also 🟡 CONCERNS: both buttons were comprehended on cold read only as guesses, and the destination copy lives solely in hover title tooltips that touch users never see.
Worth knowing on the mechanics: chenmingwei23 approved this at 08:17Z on b3a89b0c, and the approval was dismissed 7 minutes later by a force-push and again by the 11:14Z one — but the PR's content has been byte-identical since, because both pushes were pure rebases. mergeStateStatus: BLOCKED is a missing re-approval, not a failing gate. CI is green: 62 SUCCESS, 4 skipped, nothing failing or pending.
Summary
Selecting text in a Crew Members thread (and in a split-view pane) offered Copy only.
AssistantMessagedraws the selection actions its host hands it (useSelectionActions(onQuote, onAsk),AssistantMessage.tsx:212), andChatPagewas the only host passingonQuote/onAsk— Quote'sFlyingQuote-into-composer and Ask's open-/side-and-seed were ChatPage-local.This PR extracts both into one chat-core seam and wires
ChatPaneto it, so the Members page and split view get Quote / Ask about this / Copy with the same behaviour as the main chat. No second implementation is left behind.The seam —
website/src/chat-core/composer/selectionActions.ts—useSelectionQuoteAsk({ slot, setInput, revealComposer, openSideChat })→{ onQuote, onAsk, quoteFlight, endQuoteFlight }. Hosts hand it only what they own: the composer draft, how to bring the composer into view after a quote, and how to bring a Side Chat surface on screen (openSideChatabsent = no Ask; no slot = no Ask; an opener that returnsfalserefused, and the selection is then not seeded; an opener that must re-bind the page first returns aPromiseof that verdict and the seed waits for it). chat-core imports nothing frompages/.quoteDraft.ts—quoteIntoDraft(prev, text), the one blockquote shaping for the main composer, a pane's composer and the Side Chat seed (three copies collapsed).sideChatDrafts.ts— the per-slot Side Chat draft store, in-memory and subscribable (useSideChatDraft(slot)viauseSyncExternalStore). It is the one place a Side Chat's unsent text lives: typing, a failed request handing its text back, and Ask's seed all write it. The seed is a store write (seedSideChatDraft(slot, selection)appends the blockquote and bumps aseedTick), not an event fired at the panel — a store entry waits for the panel to mount and read it, so a Side Chat that comes up a frame late (afterswitchSlot, a tab open and a drawer transition) still finds the selection where an event would have fired unheard.SideChatkeeps only the focus nudge (caret after the quote) keyed onseedTick.Wiring
ChatPage(unchanged behaviour)FlyingQuotesidetab (openActivityToTab('side'))ChatPanein split view (SessionGridView)FlyingQuoteswitchSlot(paneSlot)if it is not the active slot (the panel — and its SideChat — is bound to the active slot), awaited — thesidetab opens and the seed lands only once the switch succeeded and that pane is still the active slot (a switch overtaken by a later one reportsfalse, like the slice ignores its stale fulfilment); a rejected switch (pane session gone) reportsfalse, seeds nothing and surfacespages.chatPage.side_chat_pane_gonethrough the page's ErrorNotice. Split mode is not left.ChatPaneon the Members pageFlyingQuote<SideChat slot={memberSlot}/>), with a "Details" header action back and close resetting the viewChatMessageList/MessageRenderContextgainonQuote/onAsk; the default assistant row passes them through (same shape asonFileOpen).ChatPanetherefore needs no renderer override.sideChatDraftsand subscribes to it; no event listener, no mount poll, no DOM probe.data-side-chat-sloton the composer wrapper remains as a test / capture-harness hook only.handleQuote/handleAskbodies are the hook, byte-for-byte in behaviour) plus threading the one-prop opener intoSessionGridView.components.selectionToolbar.ask, all 13 catalogs +en-XAregenerated) and gains a hint,components.selectionToolbar.ask_hint— the hover tooltip AND the button's accessible description (aria-describedby→ a visually-hidden sibling node), so keyboard and screen-reader users get it too — saying where the question lands and that pressing the button sends nothing. The Members drawer's draft-gated Side Chat header action gains a tooltip,pages.membersPage.side_chat_draft_waiting. Everything else reusespages.chat.sidePanel.menu_side("Side Chat") andpages.membersPage.details. Translator context for all three keys is inen.context.json.SideChat.tsx:border-t border-accent/30 bg-accent/5), including the main chat's right-panel Side Chat, not only the Members drawer. The hazard it answers — an off-record composer next to a live one, same send arrow — exists on the main chat too (side panel open beside the main composer, or beside a split pane), and Side Chat is one component on every surface; scoping the tint to one host would make the same panel read differently depending on where it is docked.AnimatePresence mode="wait",initial={false}so the drawer's own mount animation is not doubled), the header title changes to "Side Chat", a Details header action is the way back, closing the drawer forgets the view, and switching members returns to details. Recording below.chat-core/composer/sideChatDrafts.ts, in-memory, per slot, subscribable): every host unmountsSideChatthrough a control beside its composer — another activity tab, the Members drawer's Details, closing it, switching members — and the draft used to die with each.sendclears it through the same path; a failed submit or failed queued edit restores into the slot the request was FOR (restoreDraftTo(vars.slot)), whichever slot the panel shows by then. In-memory on purpose: a Side Chat is never persisted, so its draft should not outlive the page either.integration/setup.tsresets the store per test (same rule as the Pierre staging reset).Relationship to
feat/members-chat-steer-only(parallel PR)Branched from
origin/main, not stacked. Overlap is limited to the<ChatPane …/>mount line inMembersPage.tsx(both add one prop) and theChatPane.tsximport block /<ChatInput>neighbourhood. None of that PR's steer/composer/busyModelogic is touched. Whichever lands second keeps both props on the mount line.ChatPane vs ChatPage — remaining interaction gaps (list only, not fixed here → P3 / P5 backlog)
Props
ChatPagepasses toAssistantMessagethat the SDK's default assistant row (whatChatPane, and so Members + split view, renders) does not:onFolderOpen,onArtifactOpen,onSessionOpen,sessions,activeSessiononOpenDiff,fileChipStyle,artifactPathsonRegenerate,onSwitchVariant,isRegenerating(variants/variantIdxARE passed, so the arrows render inert)onFork,onPlanFromHere,forkIndex,forkMessageIdonLoadEarlier,loadingOlder,earlierRemaining(the pane uses its own "earlier messages → open session" row instead)onSpeakonApplyPlan,planTaskIdpinned,onTogglePinshareEnabled,prevUserText,linkPreviewsmessageTs,slotKey,slotTitle,modeAlso outside
AssistantMessage: the pane'suserrow is the SDK default (noonEditResend/ edit-and-resend; the parallel PR is adding its ownuseroverride for steer chrome), and the pane has nomcp_oauth/planorchestrator dispatch beyond what the registry carries (#5893). These are the P3 "composer/row capability" seams still to thread; this PR threads exactly one (selection actions) and deliberately nothing else.Tests (written, not run locally per task rules — CI runs them)
src/test/selectionActions.test.tsx—quoteIntoDraftshaping/stacking,sideSeedTargets,seedSideChatwaits for the named slot's composer (old-slot composer present ≠ satisfied), mount-wait exhaustion still dispatches, hook offers Ask only with an opener, Quote appends + one flight, Ask opens for THIS slot + seeds by slot + leaves the draft alone.src/test/ChatPane.selectionActions.test.tsx— the pane handsAssistantMessageonQuotealways /onAskonly withopenSideChat; Quote lands in the pane composer with the flight; stacking; Ask callsopenSideChat(paneSlot)and the seed names the pane slot, draft untouched.src/pages/members/MembersPage.sideChat.test.tsx— the page hands the pane an opener; Ask swaps the drawer to the member-slot Side Chat titled "Side Chat"; Details returns; close forgets the view.src/test/SelectToAsk.test.tsx(+1) — SideChat takes a seed naming its slot, ignores one naming another, marks its composer withdata-side-chat-slot.src/test/ChatPageMoreCoverage.test.tsx— existing Quote/Ask regression pins unchanged; the Ask pin additionally asserts the seed names the active slot.Screenshot evidence
Real components in the capture entries, genuine triple-click selections, every frame state-asserted by the harness scripts (all frames OK):
Members page —
website/scripts/capture-members-selection-quote-ask.mjsdata-testid="member-drawer-side-chat", absent while the Side Chat holds nothing)Split view (selection made in the NON-active pane,
pane-b) —website/scripts/capture-chatpane-selection-quote-ask.mjs, on the sharedprepareSplitChatPagefixturepane-b: the page re-bound topane-b(sidebar highlight moved to Release checklist), Side tab opened,pane-b's Side Chat seeded; split mode kept; both pane composers emptypane-blands inpane-b's own composer;pane-auntouchedRound 1 (head
75bf525a9, rebased ontoorigin/main@ #8951)activeSlot(never the rosterslot_key), andopenMemberresets the drawer view to details on member switch — a slug-collided member can no longer inherit a Side Chat bound to a rejected key. Two new tests pin both halves (MembersPage.sideChat.test.tsx).docs/feature-map/README.mdSide chat + Crew Members rows carry the new reach paths.max-two-buttons-per-rowon the selection toolbar — the diff adds no markup to that row;useSelectionActionsalready renders Quote / Ask / Copy for ChatPage on main (grandfathered). ChatPane merely stops withholding the two callbacks the shared component takes.border-accent/30 bg-accent/5) so the off-record composer reads differently from the live thread composer at a glance.switchSlotis visible as the sidebar highlight moving to the pane's session + the panel header (frame 2 above); split mode is kept. Restoring the prior slot on close was rejected — it would fight a later explicit selection and the grid's ownonCollapsealready uses the same switch.replaceChildren(), not.innerHTML.deadKeysratchet red was main-inherited (fixed on main by fix(i18n): drop the dead pages.chatPage.dismiss_upload_error key #8951; this head is rebased onto it).Round 2 (head
c0ebaf8d7, rebased ontoorigin/main@ #8927)SideSeedDetail.slotis required,sideSeedTargets/sideChatComposerForhave no "any Side Chat" arm, and the hook offers no Ask without a slot. Tests updated (selectionActions.test.tsx,SelectToAsk.test.tsx: a seed naming no slot is ignored).75bf525a9: a PR-body edit after the GPT lane had posted re-triggered the lane and its newest run was cancelled by concurrency, which readiness reads as a failure — re-ran that run (green). This head's body edit landed before the push so the push supersedes it.Round 3 (head
c7f3a6d8e)SideChatand discarded its uncontrolled draft. Fixed at the component: drafts are kept per slot insideChatDrafts.tsandSideChatruns its composer in controlled mode against it — which also closes the same pre-existing loss on the main chat's activity panel when switching tabs. Pinned by a newSelectToAsk.test.tsxcase (unmount → other slot sees nothing → same slot gets the draft back).Round 4 (head
166204b35, rebased ontoorigin/main@ #8958)mergeIntoDraft, i.e. into whatever slot the panel showed at rejection time — after a re-bind (split-view Ask, member switch) A's question landed in B's draft. NowrestoreDraftTo(vars.slot, vars.q): the visible draft when that slot is still the one shown, otherwise straight into that slot's store entry. Pinned by a newSelectToAsk.test.tsxcase (submit on A, re-bind the same instance to B, reject → A's store has the question, B's draft untouched).test_snapshot.py::TestNotificationCopyWhenNoLiveFileExiststiming failure in untouched Python (also seen in round 1); this PR changes no backend code. Re-run only that job if it recurs on this head.Round 5 (head
48436bb9e){A, ''}, re-bind to B, A fails and restores into the store only, re-bind to A → the stale cached empty string was shown and the next keystroke overwrote the restored question. Invariant now: the store is the single source of truth —SideChatreadsreadSideChatDraft(slot)on every render and keeps state only as a render tick, so nothing can shadow a restore. The failed-submit test now also drives the return to A and asserts the restored question is shown and appended to, not overwritten.vi.fnper property access, somockImplementationand the call assertion targeted different fns;sideTurn/sideOpenare now stable mocks (theSideChat.oversizeQuestionpattern).Round 6 (head
8a039f385)editQueued.onError(a failed queued-question edit) still merged the text into the draft of whatever slot the panel showed. NowrestoreDraftTo(vars.slot, vars.content). Rather than another point fix,restoreDraftTo's doc now carries the table of every draft writer and why each lands on the right slot (typing / chips / send-clear → the shown slot by definition; seed → gated bysideSeedTargets; queue release → read fromslotSide[slot], waits in the store for a hidden slot; failed submit and failed edit →restoreDraftTo(vars.slot)), with the rule that amergeIntoDraftinside a mutation callback is the wrong-slot bug again. (The Opus adjudicator flagged the finding as naming non-existent code; it was reading the ChatPane queue hook — themergeIntoDraftis inSideChat.tsx's owneditQueued.onError, so the finding was real and is fixed.) New test inSideChat.steerQueue.test.tsx: edit on A, re-bind to B and type, A's edit fails → A's store draft holds the wording, B's draft untouched.test_snapshot.py::TestNotificationCopyWhenNoLiveFileExists— the known flake tracked in flaky: TestNotificationCopyWhenNoLiveFileExists ordering asserts race on loaded CI shards #8893 (open); no backend code in this PR.Round 7 (head
8226e607e)8a039f385.side-seedevent + 20-frame rAF poll +data-side-chat-slotDOM probe — all three are deleted.sideChatDraftsmoved into chat-core and became subscribable (useSyncExternalStore); Ask writes the quote into the slot's store entry (seedSideChatDraft) and the panel renders it whenever it mounts, so a late-mounting panel cannot miss the selection. Draft state now lives in exactly one place.revealComposeris a host callback (selectionActions.tsno longer imports frompages/). Tests: store contract (selectionActions.test.tsx, incl. the seed-before-mount case),SelectToAsk.test.tsxseed cases rewritten against the store, host tests assert the store instead of listening for an event.main, this PR does not change them.ChatInputbehaviour on every split pane onmain; this PR does not touchChatInput.Round 8 (head
c2b4fd126, rebased ontoorigin/main@ #8998) — readiness passed on8226e607e; two advisories fixedseedTickwas never reset, so every later remount of a once-seeded slot (reopening the Side tab, a member switch and back) re-ran the focus nudge into an empty composer. The nudge now consumes the seed (consumeSideChatSeed), so a non-zero tick means "a seed is waiting for the caret", never "was seeded once". Store contract test added.useSideChatDraft(activeSlot)holds text; newMembersPage.sideChat.test.tsxcase (absent without a draft, present with one, gone when it clears).needs-a-decisionin round 7); labels pre-exist onmain, default is to ship as-is.Round 9 (head
f936f404a)switchSlotunguarded while every otherswitchSlotinChatPage(tab strip, sidebar row,?siddeep link, tab close) is gated onconnected— offline, the rejected switch clears the active pane's messages and the transcript the reader just selected from disappears until reconnect. Two-part fix: the grid is handedopenSideChatonly while connected, so panes' toolbars offer Copy / Quote only offline (capability by omission, no Ask into the void and no orphan seed); andopenSideChatForPaneitself checksconnectedRefbefore the switch, covering the frame between the drop and the re-render — it does nothing rather than open a Side Chat bound to some other slot. Two newChatPageMoreCoverage.test.tsxcases through a stubbedSessionGridView: connected Ask re-binds to the pane's slot and opens thesidetab; aftersseDisconnectedthe capability is gone and a slipped-through call moves nothing.MochiChatPanel.coverage.test.tsx › replaces the streamed text with the committed message— afindByText('partial')timeout (1034 ms) in the Mochi app's streaming footer; this PR touches nothing underapps/mochiand the file has a prior stabilisation for the same streaming assertion (test: stabilize flaky widget-tag streaming assertion (#3314) #3378). Not this PR's; Coverage Merge is its cascade.test_snapshotflake again — also red onmain's own runs34035151321/34033073642at the same time.Round 10 (head
af095d91e) — round 9's new connected-path test, fixedre-binds the activity panel to a split pane's slot…assertedactiveSlot === 'chat-2'but the harness knew onlychat-1, and ChatPage undoes a switch to a slot its list does not hold (the mode-guard clearsactiveSlot, the auto-select falls back to the first known slot) — the test was switching to a stranger.renderChatPagenow takes aslotsoption; the split tests render withchat-1+chat-2. Production code unchanged from round 9.f936f404a(GPT ✅ no blocking, Opus ✅, all advisory lanes settled); Coverage Merge / Coverage Gate were the cascade.Round 11 (head
38a1bf806, rebased ontoorigin/main@ #9037) — CONFLICTING clearedaf095d91ewas green on every frontend check and every AI lane; the two backend reds were inherited (shard 4 = flaky: TestNotificationCopyWhenNoLiveFileExists ordering asserts race on loaded CI shards #8893test_snapshot; shard 2 = threeTimeout >120sintest_dashboard_state_ws/test_history_locking_remediation— no backend code here). Main moved under the PR (28 commits), so it was rebased instead of rerun.docs/feature-map/README.mdCrew Members row: feat(members): show each member's auto-patrol (monitor loop) status #8936's auto-patrol wording and this PR's Side Chat wording merged as a union.MembersPage.tsxauto-merged with feat(members): show each member's auto-patrol (monitor loop) status #8936;tsc -b --force+ eslint clean. No code change.Round 12 (head
f3aeec2f0, rebased ontoorigin/main@ #8694) — picks up the #8893 fix38a1bf806everything was green except Backend shard 4 (test_snapshot … survives_the_READER, flaky: TestNotificationCopyWhenNoLiveFileExists ordering asserts race on loaded CI shards #8893) and its Coverage Gate cascade; shard 1's rerun passed. The shard-4 red was a stale merge ref, not a flake to rerun: fix(test): read the copy ordering before the worker is released #9047 fixed that test onmainat 16:03 UTC, but the run was built at 15:47 andgh run rerunreuses the same merge SHA, so it kept hitting the pre-fix assertion. Rebased so the merge ref contains fix(test): read the copy ordering before the worker is released #9047. No code change;tsc -b --forceclean.Round 13 (head
2e22aed33, rebased ontoorigin/main@ #9407) — CONFLICTING cleared; UX Watch items closed in-PRdocs/feature-map/README.mdCrew Members row (main's star/filter wording ∪ this PR's Side Chat wording) andMembersPage.tsxactivate()(main'ssafeSetItem(LAST_MEMBER_KEY)+ this PR'ssetDrawerView('details'), both kept). One semantic break the rebase could not see: main movedconnectedRefout ofChatPageintouseChatPageSessionController, soopenSideChatForPane's offline guard lost its ref — re-declared locally next to the callback (same read-at-click-time shape).tsc -b --forceclean.f3aeec2f0: "Ask in Side Chat: a guess (I don't know what a 'Side Chat' is yet)", carried as a first-use-guess Watch in rounds 7, 8 and 12 (round-7 disposition quotes it and asked the maintainer, who directed the relabel), so the button no longer names the panel: "Ask about this" says what happens to the selection, and a tooltip (ask_hint) says where it lands and that the main chat is untouched — the panel's own header then reads "Side Chat" when it opens. Shorter than the old label, so the width thecapture/selection-toolbar-ask.tsxentry reviews only shrinks.SelectionActiongains an optionalhint(renders astitle; the accessible name stays the label). All 13 catalogs updated (zh-CNkeeps its existing "随手问", which already followed this pattern);en.context.jsonentries rewritten/added;AssistantMessage.test.tsx+SelectToAsk.test.tsxfollow the label, and the latter pins the tooltip.pages.membersPage.side_chat_draft_waiting("An unsent question is waiting in this member's Side Chat").MembersPage.sideChat.test.tsxpins the tooltip.Round 14 (head
aff53f5b3, rebased ontoorigin/main@ #9507) — CONFLICTING cleared, no code change2e22aed33was fully green (61/61, MERGEABLE) and sat ~7 h awaiting merge whilemainmoved 34 commits. Rebased; three conflicting files, all resolved onmain's structure:MembersPage.tsx—main's react-query rewrite replaced theslots/errorsmaps withthreadOutcome(slot_key/collision/failed). Kept:useSideChatDraft(activeSlot)(now readsthreadOutcome.slot_key), thesetDrawerView('details')reset inactivate(), andopenSideChat={openMemberSideChat}on the<ChatPane>mount besidemain's newhideEmptyHint. Dropped: this PR'ssetErrors(...)cleanup, whose state no longer exists.ChatPane.tsx—hideEmptyHint(main) andopenSideChat(this PR) are both additive props; union.en.context.json—pages.membersPage.edit_member(main) ∪pages.membersPage.side_chat_draft_waiting(this PR).MembersPage.sideChat.test.tsxfollowsmain's harness: the api mock providesdefaultAgent/autonudgeList(the page's new reads) instead of the retiredkirocrewAgents, and the collision case waits onmember-thread-error(main'sErrorNoticetest id) rather than a barerole="alert".tsc -b --force, eslint,i18n:check(19, baseorigin/main) clean; both capture harnesses re-run green on the rebased page (stills re-shot, pinned above).Round 15 (head
a98d3d7ad)mainshippedcapture-queue-edit-multiline.mjswith the same 34-line split-harness stanza (splitLayouts/json/FIXTURES/check) this PR'scapture-chatpane-selection-quote-ask.mjshad, so the 0%-duplication gate tripped on the merge ref. Hoisted the stanza intoscripts/lib/prepare-split-chat-page.mjs(TWO_PANE_SPLIT_LAYOUTS,jsonResponder,splitPaneFixtures,makeChecker) — the place that file's own header says the shared wiring belongs — and this PR's script imports it.main's copy is left as-is (not this PR's file); jscpd is 0 clones with one copy gone. Harness re-run green.member-thread-error;main's rewrite renders a collision under its ownmember-thread-collisionid (the-errorid is the failed-POST notice). Test retargeted.Appchunk over by 1.1 KB. Measured on a pristinemain@6ae74179dworktree: 3,440,273 B against the 3,440,640 B ceiling — 367 B (0.01%) headroom, the drift classcheck-bundle-size.mjsitself documents. chore(bundle): re-measure the App chunk ceiling (3360 -> 3530 KB) #9523 (opened independently) re-measures it to 3530 KB; this PR picks it up when it merges. Not folded in here, per the file's own convention.test_work_ledger.py::test_two_conductors_binding_one_worker_at_once…— identical failure onmain's own run at this base (34266985826).subscribeSideChatDraftsis module-private;useSideChatDraftis the one subscription surface. The store test that called the raw subscriber now pins the same notify contract through the hook.f3aeec2f0("Ask in Side Chat: a guess — I don't know what a 'Side Chat' is yet"), raised in rounds 7, 8 and 12 as a first-use-guess Watch, and the maintainer directed that it be resolved in-PR by relabelling rather than overridden.ask_hintnow carries the no-commit reassurance the blind reader lacked ("Opens Side Chat with the selection quoted — nothing is sent until you send, and the main chat is untouched"), all 13 catalogs +en-XA; and because a hovertitleis invisible to touch/keyboard, the hint is also the button'saria-description.Round 16 (head
41cb6fc06) — two own reds from round 15's additionsaria-descriptiontripsjsx-a11y/role-supports-aria-props(the rule's ARIA table predates the 1.3 attribute) and CI lints at--max-warnings 0. Replaced witharia-describedby→ asr-onlysibling node per action (useIdprefix, since split view mounts several toolbars). The node is a sibling, not a child, so the button's own text stays the label — the capture harnesses andgetByRolename lookups read it as such.SelectToAsk.test.tsxpinstoHaveAccessibleDescription(/nothing is sent/)and the clean text. Full-treeeslint src/ --max-warnings 0clean locally.ask_hintused formal आप, tripping thehiStyleratchet (style/hi.md§4: address the user as तुम). Reworded. Audited every string this PR adds against the sibling style gates and fixed the same register slip in de (Sie→du), fr (vous→tu) and ru (вы→ты) before their ratchets could catch it; es/it/pt/zh/ja/ko were already in register.Round 17 (head
84afb49ae, rebased ontoorigin/main@ #9523) — bundle red cleared; two advisories taken41cb6fc06— inherited, see round 15 — clears on this merge ref. 11 commits, no conflicts;en-XAregenerated against the mergeden.json(byte-identical tree).quoteIntoDraftre-export fromselectionActions.tsis gone; the only importer (its test) readsquoteDraft.tsdirectly.aria-describedby→sr-onlysibling (useId), so its "an unsent question is waiting" reason reaches keyboard and screen-reader users, not only hover.MembersPage.sideChat.test.tsxpinstoHaveAccessibleDescription.titletooltips, drawn by the OS outside the page, which a Chromiumpage.screenshot()cannot paint; the text itself is pinned by tests (SelectToAsk.test.tsx,MembersPage.sideChat.test.tsx) and, since round 16, is also the element's accessible description.41cb6fc06green: GPT ✅, Opus ✅, Design PASS.Round 18 (head
6a8247bc1) — UX Watch on the composer band, takenpages.chat.sideChat.off_record_marker("Side Chat · not saved to the conversation"), 13 catalogs +en-XA,en.context.jsonentry — so the seeded composer is recognisably where "Ask about this" led and states that text here does not reach the main conversation. Frames 02 and 2 re-shot (marker visible above the seeded quote). Not an in-page tooltip: the reader's failure was the box, not the button, so the words go on the box.titlewith an in-page tooltip) — declined for this PR: the hint copy is now reachable three ways (hovertitle,aria-describedbyfor AT, and — for the destination question specifically — the marker above the composer the button opens), which is the reader's actual gap; a tooltip component change on the shared selection toolbar is a separate surface decision.84afb49aefinished fully green (61/61, MERGEABLE): GPT ✅, Opus ✅, Design PASS; this head changes only the Side Chat composer band and catalogs.Round 19 (head
dae03e58a) — harness helpers get their other consumerscapture-queue-edit-multiline.mjs(main's script that shared the stanza) importsTWO_PANE_SPLIT_LAYOUTS/jsonResponder/splitPaneFixtures/makeCheckerinstead of carrying the paste, and this PR'scapture-members-selection-quote-ask.mjsimportsmakeCheckerinstead of its inlinecheck/failed. All three harnesses re-run green through the shared code; jscpd 0 clones. Scripts only — no product code changes on this head.6a8247bc1finished fully green (60/60, MERGEABLE): GPT ✅, Opus ✅, Design PASS.Round 20 (head
f09a32ae4) — the pair's hints now disambiguate each otherquote_hint) — taken: round 13 gave Ask a hint but left Quote'stitlerestating its label, so the pair's tooltips did not tell the two destinations apart.components.selectionToolbar.quote_hint("Quotes the selection into the composer of this conversation — your reply goes to the main chat"), 13 catalogs +en-XA,en.context.jsonentry; wired through the samehint→title+aria-describedbypath.SelectToAsk.test.tsxpins that Quote's description names the main chat and Ask's names the Side Chat. No screenshot change (tooltips).dae03e58afinished fully green (61/61, MERGEABLE): GPT ✅ (all findings downgraded on adjudication), Opus ✅, Design PASS.Round 21 (head
34dfc65e6) — GPT F1 (fenced, security-class): a refused opener no longer seedsselectionActions.ts:79, fixed rather than overridden):onAskcalledopenSideChat(slot)and then seeded the slot's draft store unconditionally, so when the opener refused —ChatPage.openSideChatForPanereturning early offline (round 9's guard) — the quote was still parked in that slot's in-memory draft and would surface at the next unrelated open. The adjudicator rated the residual LOW (additive, visible, clearable, never persisted or auto-sent) and pre-drafted an override; the fix is one line cheaper than arguing it. The opener contract is now(slot) => boolean | void: returnfalseto refuse, andonAskskips the seed.openSideChatForPanereturnsfalseon the offline path andtrueotherwise; the Members opener always opens (void= opened).ChatPane/SessionGridViewprop types follow. Tests:selectionActions.test.tsx— a refusing opener leaves the store empty;ChatPageMoreCoverage.test.tsx— the slipped-through offline call returnsfalse.f09a32ae4: every other check green (59), Opus ✅, Design PASS; UX/FP advisory dispositions on record.Round 22 (head
1a21d0fbe) — GPT F1 (fenced): concurrent failed restores composed through the storeSideChat.tsx:198, fixed):restoreDraftTotook a shortcut when the failed request's slot was the one on screen —mergeIntoDraft(text)— and in controlled mode that resolves its functional updater against the composer's render-time draft (useComposerDraft.ts:201). Two edits rejected in the same tick therefore both built on the same stale base and the second restore erased the first. Now every restore, shown slot or not, goeswriteSideChatDraft(target, append(readSideChatDraft(target), text)): the store read is synchronous and current, so successive restores compose, and the shown composer re-renders from the store as it does for every other write. TheslotRefthe shortcut needed is gone. NewSideChat.steerQueue.test.tsxcase: two queued edits fail in one tick → both wordings are in the slot's draft.34dfc65e6: every other check green (59), Opus ✅, Design PASS; FP/UX dispositions on record.Round 23 (head
2b044a8af) — GPT F1 (upheld): a rejected pane re-bind no longer strands the seed; own test red fixedChatPage.tsx:4360, fixed):openSideChatForPanedispatchedswitchSlot(slot)and returnedtruesynchronously, so when the server rejected the switch (the pane's session deleted under it → 404 →switchSlot.rejectedfalls back to the prior slot) the seed had already been written for a slot the panel never bound to, with no visible error. The opener now returnsdispatch(switchSlot(slot)).unwrap().then(…): on success it opens thesidetab and resolvestrue; on rejection it reports through the page'sErrorNotice(showActionError, new keypages.chatPage.side_chat_pane_gone, 13 catalogs) and resolvesfalse. The seam (selectionActions.ts) acceptsboolean | void | Promise<boolean | void>: a sync verdict seeds synchronously (same-slot Ask, Members), a Promise defers the seed to its verdict.ChatPane/SessionGridViewprop types follow. Tests:selectionActions.test.tsx(Promise-true seeds after settle; Promise-false seeds nothing),ChatPageMoreCoverage.test.tsx(connected re-bind resolvestrue; a 404'd re-bind resolvesfalse, falls back tochat-1, opens no Side tab, and the notice is on screen).getByRole('button', { expanded: false }), which matched otheraria-expandedcontrols on the page; it now finds the stack as the[role=button][aria-expanded]ancestor of a card's text.1a21d0fbe: Opus ✅, Design PASS, every other check green (52 at the time of writing, shards still running).Round 24 (head
94f490c1d) — GPT F1 (fenced): a re-bind overtaken by a later switch no longer opens the wrong pane; rebasedChatPage.tsx:4373, fixed): round 23 resolvedtrueonswitchSlot.fulfilledalone. If the user moved on to pane A while pane B's switch was still in flight,chatSliceignores B's stale fulfilment (activeSlot !== key → return) but the opener did not: it opened the Side tab — now bound to A — and the seam seeded B's draft, so A's Side Chat came up while B's quote sat hidden. The opener now checksstore.getState().chat.activeSlot === slotonce the switch resolves and reportsfalseotherwise; the seam seeds nothing and no tab opens. Test:ChatPageMoreCoverage.test.tsx"a re-bind overtaken by a later switch reports false" (B's detail fetch held open,switchSlot('chat-1')in between, then released).[vs-base]red on2b044a8afwas main drift: the 26 addedapp-detailfindings all sit inAppDetailPage.tsx, untouched by this PR; the base tree CI exported (d4e59297) was 18 commits past the branch's merge-base. Rebased ontoorigin/main(02921dfc1), clean.2b044a8af: Frontend Tests 1–4 ✅ (round 22/23 tests pass), Opus ✅, Design PASS, everything else green.Round 25 (head
b3a89b0cf) — opener reads the Provider's store, not the module singletonstore.getState()from../store— the module singleton. In production that is the store the page renders under, butChatPageMoreCoveragerenders ChatPage under its owncreateTestStore, so the happy-path Ask resolvedfalse. The opener now takes the bound store viauseAppStore()(the patternInstancesViewport/Appalready use) and reads that after the await — the same store its dispatches went to. No behaviour change in the app.94f490c1d: all five review lanes green (GPT ✅ for the first time since round 20, Opus ✅, UX ✅, Design ✅, First Principles ✅); Frontend Tests 1/2/4 ✅; E2E[vs-base]was still running.Round 26 (head
ebf99636b) — rebase only: main drifted to CONFLICTINGb3a89b0cfwas 61/61 green and MERGEABLE; main then moved (67716d722, members roster "+" + the ChatPane pin/onDisplayItemsprops) and the PR went CONFLICTING. Rebased; two textual conflicts, both unions:ChatPane.tsxpasses main'sonDisplayItems/hiddenRowand this PR'sonQuote/onAsktoChatMessageList; the Crew Members row indocs/feature-map/README.mdtakes main's badge wording and keeps this PR's Side Chat additions. No code change otherwise; tsc / eslint / i18n clean, both harnesses re-verified.Round 27 (head
de6320652) — rebase only: picks up the #9682 fix for the inherited shard-4 redebf99636bwas red only on Backend Tests (3.12, 4) / (Windows) (4) — inherited from main: fix(workflows): bootstrap the platform context #9066'stest_main_boots_platform_before_binding_serverpatchedserver.ThreadingHTTPServerwhilemain()constructs_Server, so the test bound a real socket and hung (Timeout >120son Linux, xdist worker crash on Windows); main's own run on67716d722failed the same jobs. Fixed on main by fix: patch _Server in workflows entrypoint tests (#9675) #9682 (_Serverretarget + aserver_bindtripwire); my duplicate fix(test): patch the server class workflows main() actually constructs #9683 closed as superseded. Rebased ontoc0cf53f3e, clean, no code change; tsc / eslint / i18n clean, both harnesses re-verified.Local gates
tsc -b✓ ·eslinton every touched file ✓ ·i18n:checkwithI18N_BASE_REF=origin/main✓ (19 checks, incl. the diff-scoped catalog QA) ·check-theme-colors✓ ·check-phantom-classes✓ ·jscpd .✓ (0 clones). vitest/pytest deliberately not run locally.no linked issue: dispatched from a crew work item (session task), not a GitHub issue; #8964 is a follow-up filed by this PR, #5893 / #5895 / #5892 are referenced as precedent only.