feat(web): collapse tool activity into one line#7152

🟡 Medium chat/MessagesTimeline.tsx:1461

Expanding a live tool group does not show its currently running tool details: aria-expanded changes, but WorkGroupSection filters the inProgress entry through workEntryIndicatesToolNeutralStatus, leaving the expanded group visually empty. Update the expanded-child filtering so the active entry is retained for live groups.

No longer relevant as of 96c8a4df8d37b35ef87d3c9cf78e6bf35d1ece76

chat/MessagesTimeline.tsx:2416

The status text (3 failed / 2 working / Completed) was previously always rendered; gating it on @[32rem]/agent-group means that in a narrow chat column the only remaining failure signal is the aria-hidden colour dot on line 2408, so the state becomes colour-only and invisible to assistive tech. Suggest keeping a text fallback that is only announced when the visible status is hidden.

No longer relevant as of d2095f27f302cd6d4c8bc795bb6f8e38c509d90e

🟡 Medium chat/MessagesTimeline.tsx:2203

tokenizeShellCommand removes the backslashes in quoted Windows paths, so "C:\Program Files\node.exe" script.js becomes C:Program Filesnode.exe; commandProgramName then cannot split the path and displays Running C:Program Filesnode.exe instead of Running node.exe. Preserve Windows path separators while retaining shell escape handling.

✅ Resolved in d4e205311888b825f7796b07c672a484cda751b5

Completed tools jump in timeline

Medium Severity

Merging a tool.started row into its completion copies the completion createdAt (and id) onto the start row. deriveTimelineEntries then re-sorts by createdAt, so a finishing call teleports later in the timeline and can split a live batch around intervening messages.

^{Reviewed by Cursor Bugbot for commit 8408cc3160f1a19b29f0a1afca773f31d71471b6. Configure here.}

src/index.css:143

These two tokens are never consumed: no animate-live-activity-focus / animate-live-activity-focus-counter class exists anywhere, and the @utility live-activity-focus (line 482) and @utility live-activity-focus-counter (line 494) hardcode the same 2.2s steps(24) infinite shorthand. That leaves the animation timing with two sources of truth, one of which is inert — editing the token silently changes nothing. Because this is @theme inline, the variables are not emitted either, so the utilities cannot reference them via var().

Suggest dropping the dead tokens and keeping the shorthand in the utilities as the single owner; the @keyframes stay in the output because Tailwind retains @theme keyframes whose name appears in an emitted animation: declaration.

No longer relevant as of a95aceb534eb0828d41b98fecb21d880c7ad7262

🟡 Medium chat/MessagesTimeline.tsx:1472

A completed successful work-live entry is displayed as if it is still running: liveWorkEntryLabel formats its command as Running <program>. When the trailing active batch contains only completed entries, this leaves the timeline showing a stale running status and hides the separate Thinking indicator until later commentary arrives. Use a completion-aware label/rendering path for completed entries so the row no longer reports an active command.

No longer relevant as of 4816e2c5c929842a3f13ba60496401d64b476023

Grep summarized as web search

Medium Severity

toolGroupAction treats every itemType === "web_search" entry as a web search, so completed groups get labels like Searched the web. In this codebase, codebase grep activities also use that itemType, so the new summary copy mislabels local search as web search.

^{Reviewed by Cursor Bugbot for commit 653175303ca37c3814a5c0719ae02e231aa87347. Configure here.}

The non-highlighted copy of this row is meant to read like the static activity rows it stacks with, but it differs from them in two ways:

• The icon drops opacity-70. WorkGroupToggleTimelineRow's summary icon (line 1527) and PlainWorkEntryRow's icon (line 2602) both render text-icon-muted + opacity-70, so the live row shows the same icon set at a visibly darker tone than the rows directly above and below it, even though its label uses the identical text-secondary-label.
• The content inset is px-1 rather than the px-0.5 those rows use, so the live row's icon/label column sits 2px right of the tool-group summary row and of its own expanded children (which are rendered through PlainWorkEntryRow), giving one disclosure group a ragged left edge.

Suggest restoring the muted icon tone here, and using px-0.5 when an icon is present while keeping px-1 for the icon-less Thinking row so it stays aligned with the "Working for" label above it.

^{Posted via Macroscope — UI Consistency}

This PR moves the activity column to a single scale — the new summary branch here uses a size-6 icon slot with a size-4 glyph and text-sm leading-relaxed, PlainWorkEntryRow moved from size-5/size-3.5/text-[12px] to the same size-6/size-4/text-sm (lines 2557, 2612, 2620), and TurnFoldTimelineRow moved to text-sm leading-relaxed too. The count branch of this same component (lines 1550-1562) keeps the old metrics, so the +N previous log entries toggle now sits 0.25rem left of the rows it reveals and renders at 12px next to their 14px — a ragged left edge inside one group, where the two used to align.

Suggest bringing the count branch to the same slot and type scale:

^{Posted via Macroscope — UI Consistency}

chat/MessagesTimeline.tsx:1468

announceFailure is passed unconditionally by LiveActivityRow (line 1417), and it is not gated on failed here, so the icon slot gets role="img" aria-label="Tool call failed" on every live tool row — including healthy inProgress ones. LiveWorkEntryTimelineRow only sets its own aria-label when the entry failed, so a running row's accessible name is computed from its contents and comes out as "Tool call failed Running psql". Suggest gating the announcement on the failure state so it only fires for the x icon.

No longer relevant as of b06fab9f30e3227956de80097ee960bd400e6f1f

WorkingTimelineRow now renders the same shell as this fold header (border-b border-border/60 pb-2 pt-1 wrapper, px-1 text-sm leading-relaxed … tabular-nums label), so "Working for 12s" and "Worked for 12s" are visual twins that swap in place when a turn settles — but they resolve different tokens: this row stays text-muted-foreground while the working header (line 1306) and every activity line this fold reveals (LiveActivityContent line 1465, the tool-group summary line 1562, PlainWorkEntryRow line 2605) use text-secondary-label. --secondary-label only aliases --muted-foreground in the default palette; imported themes derive it from a different palette entry (themePalette.ts line 1185 textMuted vs line 1183 mutedForeground), so under a custom theme the header visibly changes color the moment the turn completes. Suggest using the same token for both headers.

^{Posted via Macroscope — UI Consistency}

chat/MessagesTimeline.tsx:2508

This card is now a verbatim re-creation of ChangedFilesCard (ChangedFilesTree.tsx lines 54-141): identical shell rounded-xl border border-border/70 bg-secondary p-1.5 dark:border-transparent dark:bg-input/32, identical flex items-center justify-between gap-2 px-1 header, and the same Button size="xs" variant="outline" + Tooltip + @[24rem]/<container>:inline label action. Two call sites now own the same durable, theme-sensitive treatment — including the dark border/background pair — so a change to one silently diverges from the other, which is exactly how the padding/radius already drifted between these cards before this PR realigned them. Suggest extracting the shell plus the responsive icon+label action into a small shared timeline-card primitive, leaving only the container name, contents, and -mx-1/mt-4 positioning at each call site.

No longer relevant as of 4816e2c5c929842a3f13ba60496401d64b476023

Switching ml-auto → ml-1 here (and at line 301) removes the shared right edge these stats were aligned against, but they still render with DiffStatLabel's default layout="aligned", whose only purpose is that alignment: inline-grid grid-cols-[4ch_4ch] gap-2 text-right. Now that each block starts immediately after a variable-length filename, the fixed 4ch columns can never line up across rows — they only add ragged internal whitespace between the name, +N and -N. The new preview row in the same card renders the identical datum in the same hugging position with layout="inline" (line 179), so the two halves of the card now disagree. Suggest using the inline layout in both tree rows (or restoring ml-auto if the alignment column was meant to stay).

^{Posted via Macroscope — UI Consistency}

🟡 Medium chat/MessagesTimeline.logic.ts:358

The live group's identity changes when its first tool call becomes neutral, so an expanded group loses its expandedWorkGroupIds match and remounts as the next call's group. activeWorkAnchor is selected from visibleActiveToolEntries, causing this function to receive a different toolCallId; derive the group key from a stable underlying entry (for example, the first activeToolEntries item) while keeping the latest visible entry for display.

✅ Resolved in 50b35d866d9d43c5cf2cfa3dbaa4907c8e0716d9

Live row ignores earlier failures

Medium Severity

LiveWorkEntryTimelineRow only runs workEntryIndicatesToolFailure on the latest live entry. Completed tool groups set hasFailure from every grouped entry. If an earlier call failed and a newer parallel call succeeded, the live pill drops the failure affordance even though the expanded batch still contains the failed call.

^{Reviewed by Cursor Bugbot for commit aa78a62bed257493ded9c337ca485de30fdfceee. Configure here.}

chat/MessagesTimeline.tsx:1523

The two icon classifiers still disagree for the non-read cases. toolGroupAction funnels dynamic_tool_call, collab_agent_tool_call and tone-only tool entries into other, so the summary row draws wrench, while workEntryIconName gives those same entries hammer (line 2400) and bot (line 2402). One dynamic tool call therefore renders a hammer in the work-live row, switches to a wrench once it settles into Used 1 tool, and switches back to a hammer on the expanded child row — the icon changes twice for a single unchanged call.

The Read File case was fixed by teaching workEntryIconName about it; suggest closing the remaining gap the same way, by deriving this summary glyph from the group's per-entry workEntryIconName when the entries agree (which needs the entries, or their resolved icon name, carried on the work-toggle row) and keeping the action mapping only as the mixed-group fallback. Mapping other straight to hammer is not enough on its own, since it would then disagree with mcp_tool_call rows.

No longer relevant as of 9ae5f817411576b476a9e57b6680a2dbe9e3ca02

This header toggle loses flex-1 (and, just below, the Show files / Hide files hint), so the disclosure now sizes to chevron + label while the header row still spans the full card width under justify-between. The gap between the label and the Open diff / expand-all actions used to be part of the toggle and is now dead space, and with px-1 py-1 on text-xs leading-4 the remaining target is ~24px tall — below a comfortable coarse-pointer hit target for the card's primary control, whose only affordance left is the chevron. Suggest keeping the button flexible so the whole header row still toggles (the ChangedFilesTree.test.tsx class assertion needs the same update).

^{Posted via Macroscope — UI Consistency}

chat/MessagesTimeline.tsx:1485

The live row strips command before asking workEntryIndicatesToolFailure, but the settled paths introduced alongside it do not: hasFailure for the tool-group summary (MessagesTimeline.logic.ts line 830) and showFailedIndicator on the expanded child row (line 2605) both run the heuristic over detail + command. Because the exception is keyed on which row renders the entry rather than on the entry's state, a completed call that this PR keeps in the live slot (see the "keeps a completed trailing call in the live slot" case) shows no failure while the turn runs, then gains a destructive X plus Tool call failed — and taints the whole group summary — the moment the turn settles, with nothing about the entry having changed. The PR's own test (rg 'command not found' src) is exactly such a command.

Suggest gating the exception on the lifecycle state it was written for, so an entry keeps one failure affordance in both rows:

No longer relevant as of dea09d603789154877b669854f6f10798511d657

Summary failure ignores command strip

Low Severity

Collapsed tool-group hasFailure calls workEntryIndicatesToolFailure on the full entry, including command. The live row strips command before the same check so search invocations like rg 'command not found' do not look failed. Settled summaries can still show failure chrome from benign command text.

^{Reviewed by Cursor Bugbot for commit d4e205311888b825f7796b07c672a484cda751b5. Configure here.}

chat/MessagesTimeline.tsx:2532

This card re-creates the timeline's card surface (border border-border/70 bg-secondary dark:border-transparent dark:bg-input/32) but with different geometry than the only other card in the chat column, ChangedFilesCard (ChangedFilesTree.tsx:59): rounded-xl + p-1.5 here vs rounded-2xl + p-2 there. Both can render under the same assistant turn, so the same surface treatment shows two different corner radii and inset paddings. Suggest aligning this card with the established geometry (or extracting the shared surface so the two call sites cannot drift again).

No longer relevant as of f3e27d751bdb3fe2798e0037fbd65ccd1e8f05e1

Duplicate rows after tool completion

Medium Severity

The new toolLifecycleRowIndex drops a call’s key once the merged row is tool.completed, and when a later same-key event cannot merge it deletes the index and still appends. A delayed tool.updated (or similar) for that toolCallId then becomes a second work-log row instead of folding into the completed call.

^{Reviewed by Cursor Bugbot for commit 266e26758498b080ad8a3966270993f6fc6a2496. Configure here.}

chat/MessagesTimeline.tsx:2544

The status text is now gated on @[32rem]/agent-group, with only an sr-only copy below that width. The container is the card itself, so in a narrow chat column (split view, mobile web) sighted users lose 2 working / 1 failed / Completed entirely, and the sole remaining differentiator between a failed and a completed fleet is the size-1.5 dot — which is aria-hidden and carries no tooltip or label. Before this change the status was always rendered, so this is a color-only state at narrow widths.

Suggest keeping the status visible at all widths and gating only the token total, which is the part that needs the room:

No longer relevant as of b99f41078abb2a18b42980fb85e85800b547e82b

Sudo env assignment mislabeled

Low Severity

commandProgramName skips VAR=value tokens for bare commands and env, but not after sudo. For sudo FOO=bar npm test, the live label becomes Running FOO=bar instead of Running npm, so the activity pill shows the environment assignment rather than the program being run.

^{Reviewed by Cursor Bugbot for commit bb47d9dea0c1715239b6290110709de5c82d0679. Configure here.}

chat/MessagesTimeline.tsx:1455

The new live rows inset at px-1, but every settled row in the same activity column insets at px-0.5: the tool-group summary button (line 1533) and PlainWorkEntryRow (line 2646, whose WorkGroupSection wrapper is -mx-1 px-1, i.e. net zero). So the running tool's icon and label sit 2px right of the identical call once it settles into Used 1 tool, and expanding the live row indents its own children 2px left of the header that discloses them — the column this PR unifies to size-6/size-4/text-sm ends up with two x-origins.

Suggest matching the column's px-0.5 here; the shimmer geometry comes from the live-activity-focus* utilities, so it is unaffected. The gap-1.5 px-1 py-0.5 assertion in MessagesTimeline.test.tsx needs the same update.

No longer relevant as of 4816e2c5c929842a3f13ba60496401d64b476023

chat/MessagesTimeline.tsx:2670

Expanded tool-group children are now the only rows left on the old 12px scale. This PR moves the whole activity column to text-sm leading-relaxed with a min-h-6 row box — the turn fold (line 1110), the working header (line 1300), the live row (line 1412), and both work-toggle branches (lines 1533, 1566) — while the row revealed under those headers keeps text-[12px] leading-5 and, combined with py-0 on line 2647, has no min-h-6, so it renders ~20px tall even though it is still role="button" when the call has output.

The live row makes the mismatch most visible: appendActiveWorkRows renders its grouped entries through this same path, so the entry shown as Running psql at 14px in the header reappears at 12px in the child row directly below it, in the same size-6/size-4 icon column and at the same indent.

Suggest giving the expanded child the column's scale and minimum height (keeping only the tighter vertical padding for density):

No longer relevant as of 2fe727dca96cf12d501bdd9d2e9170c438c170db

hasFailure is only ever rendered by the summary branch of WorkGroupToggleTimelineRow (destructive icon plus the …, tool call failed accessible name, MessagesTimeline.tsx lines 1538-1550). This overflow branch produces the chevron variant (summary: null), which ignores the field entirely, so two collapsed disclosures of the same batch disagree: an all-tool group that hides a failed call announces the failure, while a mixed group that hides the same call stays silent.

The value is also computed over visibleGroupedEntries rather than hiddenEntries, so it would attribute a still-visible failure (which already renders its own X on its own row) to the collapsed set. Suggest scoping it to hiddenEntries and rendering the same failure affordance in the chevron branch, or dropping the field for this variant so the row type only carries state its renderer uses.

^{Posted via Macroscope — UI Consistency}

chat/MessagesTimeline.logic.ts:717

The settled grouping path deliberately keeps tone === "error" entries out of tool-group summarization (line 816-821, covered by the new "keeps error entries visible instead of summarizing them as tools" test), but this active-turn collector only excludes agentSpawn. workLogEntryIsToolLike returns true for tone error (session-logic.ts:168), and runtime.error / tool.denied activities are emitted with the turn's turnId (ProviderRuntimeIngestion.ts:433-439), so while the turn is running such an entry is added to activeWorkEntryIds and skipped as its own row (line 791-793).

Result: a provider error or denied tool call is visible only if it happens to be the tail of the live batch — any later tool call hides it behind the collapsed live row until the turn settles, at which point it pops back out as a standalone destructive row. Even as the tail it renders through LiveActivityContent, which has no equivalent of PlainWorkEntryRow's destructive label treatment. Suggest applying the same exclusion here so error rows stay outside the live batch in both states.

No longer relevant as of 4dcdd9652028633c90a77a8adc340a8943583457

chat/MessagesTimeline.tsx:1487

This button and the settled tool-group summary button (line 1535) are the same disclosure — both call ctx.onToggleWorkGroup with the same work-group:tool:<turn>:<callId> id and reveal the same children — but only the settled one gets pointer feedback (transition-colors hover:bg-accent/20, also used by the overflow toggle at line 1569 and by expandable PlainWorkEntryRow at line 2664). The live row declares cursor-pointer and aria-expanded with no hover or press state and no chevron, so while a turn is running the control looks inert, then grows a hover tint the moment the batch settles into Used 1 tool. The new test at MessagesTimeline.test.tsx (expect(markup).not.toContain("hover:bg-accent/20")) pins that divergence, so it would need updating with the fix.

group/live-work also has no group-*/live-work consumer anywhere, so it can go.

No longer relevant as of 4816e2c5c929842a3f13ba60496401d64b476023

The aggregate failure icon here is not gated on collapse, so an expanded summary group swaps its category icon (eye / terminal / square-pen) for a destructive x and announces …, tool call failed while every failing child row below already renders its own x plus Tool call failed (PlainWorkEntryRow, lines 2620-2686).

The overflow toggle immediately below now avoids exactly that with showHiddenFailure = row.hasFailure && !row.expanded (line 1564), so the two disclosures in the same activity column disagree about aggregate failure state — one drops back to its normal icon when the failing rows become visible, the other doesn't. LiveWorkEntryTimelineRow (line 1482) has the same gap for its expanded batch.

Suggest deriving const showGroupFailure = row.hasFailure && !row.expanded; and using it for the icon name, wrapper color, role/aria-label, and the button's aria-label here (and the equivalent in the live row), so the aggregate marker only stands in for hidden rows.

^{Posted via Macroscope — UI Consistency}

Intentional: the group summary keeps its aggregate failure identity while expanded. Expanded children explain which calls failed; they do not replace the status of the group. The overflow toggle differs because it describes only hidden rows.

chat/MessagesTimeline.tsx:919

isExpandedToolGroupHeader keys on row.onlyToolEntries && row.expanded, but the renderer only draws a group header (children below it) when row.onlyToolEntries && row.summary (WorkGroupToggleTimelineRow, line 1531). The overflow toggle — summary: null, pushed after its revealed rows in MessagesTimeline.logic.ts (lines 881-906) — can also carry onlyToolEntries: true, since it is now derived from hiddenEntries and workLogEntryIsToolLike is true for tone: "error" rows. With MAX_VISIBLE_WORK_LOG_ENTRIES = 1, a two-entry mixed group (e.g. one runtime.error row plus one tool row) hits exactly that: expanding it gives the trailing +1 previous tool call row pb-0 instead of pb-2, so it collapses against the next unrelated row while every revealed row above it keeps normal spacing.

Suggest gating on the same condition the renderer uses so the two decisions cannot drift:

No longer relevant as of 96c8a4df8d37b35ef87d3c9cf78e6bf35d1ece76

🟡 Medium chat/MessagesTimeline.tsx:2232

commandProgramName returns null for env -v npm test, so the UI reports Running command instead of Running npm. COMMAND_WRAPPER_FLAGS.env includes --debug but not GNU env's equivalent -v; add the alias so this valid flag is skipped.

✅ Resolved in f19ee2b971e375e51d5637424f1687537acd5fa9

chat/MessagesTimeline.logic.ts:855

summarizeToolGroup now counts only terminal lifecycle entries (toolGroupSummaryEntries, line 354), but hiddenCount, hasFailure, and the rows revealed on expand still use the full visibleGroupedEntries, so the header and its own contents disagree.

For the id-less lifecycle shape this PR keeps in the work log (does not leave id-less tool lifecycles running: a statusless tool.updated plus its tool.completed, session-logic.test.ts), the update marker is not neutral — workEntryIndicatesToolSuccess is true for an undefined status — so it survives the line 826 filter. The collapsed row reads Used 1 tool, and expanding it lists two rows for that one call, the intermediate marker rendered with the same settled treatment as the completion.

Suggest deriving the entry set once and using it for the label, the count, the failure flag, and the expanded children (toolGroupSummaryEntries already falls back to the raw entries when every entry is non-terminal, so an in-flight-only group still renders).

Same split inside the helper: toolGroupSummaryKind filters for the action set (line 367) but the icon fallback still maps over the unfiltered entries (line 374), so a non-terminal marker with a different itemType can push the group icon to mixed/hammer while the label ignores that entry.

No longer relevant as of f19ee2b971e375e51d5637424f1687537acd5fa9

chat/MessagesTimeline.logic.ts:736

visibleActiveToolEntries never applies the terminal-entry rule that toolGroupSummaryEntries (line 355) applies to the settled group, so an id-less lifecycle pair is counted twice while the turn is running.

deriveWorkLogEntries deliberately keeps a statusless tool.updated marker alongside its tool.completed row for providers that emit no toolCallId (session-logic.test.ts, "does not leave id-less tool lifecycles running" — ids ["legacy-tool-update", "legacy-tool-complete"], statuses [undefined, "completed"]). Both are tool-like and neither is neutral (workEntryIndicatesToolSuccess is true for an undefined status), so both land in groupedEntries: the live row's label advances through the marker, and expanding it reveals two rows for one call. The moment the turn settles, the same batch renders as Used 1 tool with a single child (summarizeToolGroup test "counts id-less lifecycle markers as one completed tool"), so the same tool call visibly changes count on settle.

Suggest dropping superseded lifecycle markers here — entries whose sourceActivityKind is tool.started/tool.updated with an undefined toolLifecycleStatus once a later entry in the batch is terminal — rather than reusing toolGroupSummaryEntries as-is, which would also drop genuinely in-progress parallel calls whenever any sibling has completed.

No longer relevant as of 8f9910e4fb3c0cf05c6ba766ee6da472d901f96d

🟡 Medium src/session-logic.ts:1075

An unfinished same-type/same-label tool row from an older turn makes a current turn's adjacent tool.started/tool.completed pair remain as two rows, so the lifecycle is not collapsed. hasCompetingIdlessCompletionTarget scans all earlier collapsed rows without matching candidate.turnId to previous.turnId; restrict that search to the same turn, as shouldCollapseToolLifecycleEntries does.

✅ Resolved in 374c3d5638887c480344a72db072c15d8126b28b

🟡 Medium chat/MessagesTimeline.logic.ts:386

With concurrent same-label calls ordered started(A), started(B), completed(A), the live timeline drops B's still-running marker and shows only the completed call. The backward scan records one shared identity for A and B, so A's terminal marker supersedes both; correlate lifecycle events with a unique call identity before suppressing markers.

No longer relevant as of 6d25760e673c6eb1110d8ecf7692b1a75dab44e3

Thinking work vanishes from live timeline

High Severity

The live batch treats every non-error workLogEntryIsToolLike row as a tool, including thinking rows without agentSpawn such as background task progress. Those ids are stripped from the chronological list, but isVisibleActiveToolEntry still hides neutral thinking rows, so they never land in work-live either and disappear while the turn is active. After settle, the same predicate folds them into a tool summary.

^{Reviewed by Cursor Bugbot for commit 039583025bbe4ecfc209f957aa2309983aef4ace. Configure here.}