GeezerVison

Adds a per-user Collapse assistant steps preference (Agents > Settings, off by default) that folds each turn's consecutive tool-step rows into one disclosure: "Worked for 2m 14s (7 steps)" once the turn completes, or a ticking "Working for 12s" while it is still running.

Stack

Top of a six-PR stack, split from this PR's original single diff for review size. Each layer builds and passes CI on its own.

1. feat: add a collapse_assistant_steps user preference #29416 feat: add a collapse_assistant_steps user preference
2. feat(site): group assistant steps into working blocks #29417 feat(site): group assistant steps into working blocks
3. feat(site): add the working block disclosure #29418 feat(site): add the working block disclosure
4. feat(site): fold assistant steps in the conversation timeline #29419 feat(site): fold assistant steps in the conversation timeline
5. fix(site): keep streaming text alive inside expanded working blocks #29420 fix(site): keep streaming text alive inside expanded working blocks
6. feat(site): add the collapse assistant steps chat preference #29115 feat(site): add the collapse assistant steps chat preference (this PR)

This PR

Adds the Collapse assistant steps switch to Agents > Settings and makes the chat page wait for the preference in its existing loading view (the same skeleton it shows while the chat and messages load), so a saved "collapse" setting never paints rows unfolded and then folds them. If the preference request fails, the transcript renders unfolded, matching how the other display preferences on the page degrade, and the settings switch surfaces the load error.

How the feature works

A step row is assistant output whose last visible block is a tool call or reasoning; narration that precedes a tool call folds with it. Answers, user prompts, system notices, and rows containing ask_user_question, propose_plan, or compaction markers are never folded. Failed steps fold like any other step. Durations are the wall-clock span of part timestamps from the raw messages, so hidden tool-result messages count, parallel tools are not summed, and the idle gap before the next prompt is excluded (#29417).

With the preference on, the live assistant row stays inside the fold for the whole working phase, from the first reasoning delta to the final answer, which streams outside the fold. Between persisted steps the empty live row joins the block it follows instead of rendering a separate Thinking row. The block's clock runs while the chat status is running and stops during interrupting, which still keeps running tools folded; requires_action completes the block and keeps the pending client-executed tool visible (#29419, #29420).

Each block is its own MessageScroller.Item keyed off its newest row, so prepending older pages keeps its element and expansion. While older history is still unloaded the oldest block reports lower bounds. Older rows that join an expanded block at the top scroll the viewport by their own growth so the reading position holds (#29418, #29419).

Two visible effects are accepted rather than fixed: narration that precedes a tool call is briefly visible outside the fold until its tool call streams, and the completed label can be up to about 2s shorter than the last live tick because the completed span uses persisted part timestamps while the live clock also counts the model's pre-answer latency.

Split notes

The stack was rebuilt from the previously reviewed head 7611f70 merged with main. Later review rounds (see the review record) refactored every frontend layer, so the tip no longer equals that head. Rebuilding this branch as the stack's tip required a force-push. After #29416 merged, this branch was restacked onto its new parent as one squashed commit per layer (content unchanged, verified by git patch-id), followed by the review fixes; that required a force-push.

Rebased onto main on 2026-10-04 (clean). A frontend compliance pass (frontend-review and coder-frontend-review) found nothing to change in this layer.

Dogfood UAT

Independent remote UAT (dogfood workspace, real Claude Haiku 4.5 via the AI Gateway, real browser input) ran in seven rounds. Round 1 (605272e) failed on the between-steps timer freeze, round 2 (63d99ee) confirmed that fix and found the in-block reading-position jump, round 3 verified c932002, round 4 (8ddc43a) verified the live-turn fold with a 100ms DOM sampler, round 5 confirmed it on the main merge c822df1, round 6 (a70cc67) found the final-page reading-position jump fixed above, and round 7 verified a83e850 (each proven by git rev-parse HEAD and /api/v2/buildinfo on the remote). The only later commit adds a story and changes no runtime code.

The failed step badge that some rounds below checked was removed afterwards (#29418); rows that mention it describe the earlier build.

Sanitized evidence (screenshots, recordings, sampler JSON, raw API responses, harnesses) is attached to the dogfood chat https://dogfood.cdr.dev/agents/385f856b-7742-4d75-af8c-f94e2a0df558 as r4-uat-archive.tar.gz, r3-uat-archive.tar.gz and r2-final-evidence-archive.tar.gz; rounds 5 to 7 wrote their measurements to the UAT workspace (/home/coder/uat-r5 to uat-r7) because the chat's attachment slots are exhausted.

The UAT commits above belong to the pre-split branch. Review fixes since then changed runtime code on every frontend layer (see the review record), so remote UAT has not exercised the current heads.

Review record

• Codex on the pre-split PR: OK for reviewed commit 7611f70 (2026-09-15 20:00Z), CI 31 of 31 green on that head. Three earlier threads were resolved with evidence (fallback tool IDs and refetch-after-PUT declined as hardening outside the diff; the SQL ON CONFLICT cast claim refuted by the passing Postgres-backed toggle test).
• Codex per PR after the split (2026-09-17): "Didn't find any major issues" on the current heads feat: add a collapse_assistant_steps user preference #29416 87292f1, feat(site): group assistant steps into working blocks #29417 97f0bf6, feat(site): add the working block disclosure #29418 5367f18, feat(site): fold assistant steps in the conversation timeline #29419 64d45a6, fix(site): keep streaming text alive inside expanded working blocks #29420 88625f5, feat(site): add the collapse assistant steps chat preference #29115 4e5332d. The FE1 threads on the timeline tests were closed by rewriting them (below); the timestamp-less live-only identity (P2, feat(site): fold assistant steps in the conversation timeline #29419) and the error-fallback mount refetch (P2, this PR) were declined with evidence in their threads and resolved.
• Frontend self-review gate (2026-09-17): the repository frontend-review checklist (FE1 to FE10) and the vercel-react-best-practices rules were run over the composed site/src diff. Fixed: aria-expanded and DOM-node identity assertions in ConversationTimeline.test.tsx replaced by focus retention across the rerender, with the duplicate keyboard-toggle test and the screenshot-only default-collapsed and preference-restore tests removed (feat(site): fold assistant steps in the conversation timeline #29419, fix(site): keep streaming text alive inside expanded working blocks #29420); toBeChecked/toBeDisabled outcomes dropped from CollapseAssistantStepsSettings.test.tsx in favor of the paired stories (this PR); a three-pass map().filter().map() collapsed into one flatMap (feat(site): fold assistant steps in the conversation timeline #29419); the Paginated story now ends in the partial state it exists to capture (feat(site): fold assistant steps in the conversation timeline #29419). Judged compliant and left as is: the settings toggle uses aria-label alone like eight of its nine sibling toggles (FE6), and groupWorkingBlocks in the render body relies on the React Compiler, which this page already uses instead of manual memoization (rerender-memo).
• coder-frontend-review gate (2026-09-20): the eight reviewer roles (simplicity, comments, tests and stories, React patterns, reuse, styling and accessibility, types and naming, data flow) ran over the composed site/src diff and returned 45 findings. 38 were fixed on their owning PRs (feat(site): group assistant steps into working blocks #29417 4d05139, feat(site): add the working block disclosure #29418 b421c87, feat(site): fold assistant steps in the conversation timeline #29419 d0105de, fix(site): keep streaming text alive inside expanded working blocks #29420 f46888b, this PR 73d573c): the story-only now clock prop was removed in favour of the pinned fixture clock, the renderRow and renderLoadingView render helpers were replaced (row contents built once and placed; one loading-view props object), the live item-key reconciliation and expansion bookkeeping moved into pure helpers behind a small hook, stateful story wrappers became one-state static stories, the unread createdAt stream fields, rowKey, the binary search, and the unreachable NaN branch were deleted, isStepRow became getStepRowContent, chatStatus is required, fixtures are shared Mock* constants, comments were trimmed or corrected, and the settings switch gained aria-describedby plus copy in the setting's own vocabulary. The composed site/ diff shrank from +3250/-86 to +2889/-73. Declined with rationale: DF5 (the page-level preference gate was requested by Codex round 4 to avoid an unfolded-then-fold first paint; it lifts on error and the ColdLoadWaitsForCollapsePreference story captures it), RO1 (no existing formatter produces 12s / 2m 14s / 1h 5m, and extending the debug-panel formatter would couple unrelated features), RO3 and SA2 wrapper extraction (the two other dim sites are pre-existing ChatMessageItem markup; the block item now carries the same classes including the transition), RO3 and RO4 nits (scroller decorator shared by two story files; formatWorkingDuration stays beside the span computation it formats and its unit test). The PreferenceChanges story was dropped rather than moved to Vitest: its final frame is the expanded block already captured, and the expansion map trivially survives a preference flip. A second pass of all eight roles on the fixed diff returned PASS on every checkpoint except two comment placements (fixed: the live item-key reconciliation now owns its doc in liveBlockItemKeys.ts, the expansion recording scheme is documented on recordBlockExpansion) and two nits left as is (RP6 extract a WorkingBlockItem component from a 25-line map branch; DF5 bound the preference gate against a never-settling request, which matches how the pre-existing chat and messages gate already behaves). Codex on the first fixed heads: OK on feat(site): add the working block disclosure #29418 (d489115) and fix(site): keep streaming text alive inside expanded working blocks #29420 (9c3cddb); findings fixed on the rest: P2 hidden user-message boundaries in grouping (feat(site): group assistant steps into working blocks #29417, 2e43056, red-green tests), P1 expanded editing capture (feat(site): fold assistant steps in the conversation timeline #29419, 34a125b), P1 rendered-state waits in the settings test (this PR, 1b6348e). Open human threads answered with fix commits: failed-step badge accessible-name separator (jscottmiller, feat(site): add the working block disclosure #29418), observeContent, LiveLabelProps, stepsLabel, shared MockWorkingBlock, live clock path, memberIds producer assertions (DanielleMaywood, feat(site): add the working block disclosure #29418), and the startedAt doc (feat(site): group assistant steps into working blocks #29417); the feat: add a collapse_assistant_steps user preference #29416 SQL cast P1 was refuted with the Postgres-backed toggle test and resolved. Codex round on the second heads (12:26Z): OK on feat(site): add the working block disclosure #29418 (e1e8afe) and this PR (1b6348e); on feat(site): group assistant steps into working blocks #29417 two follow-on P2s fixed (merged read_file rows now split at a hidden prompt, and a block whose page starts with a hidden prompt is not partial, 4d05139, red-green) and one declined (span cap at raw user IDs is unreachable: hidden timestamped entries are tool results that follow their call's row); on feat(site): fold assistant steps in the conversation timeline #29419 the partial-state P2 is fixed by the same grouping change and the renderer-soft-failure badge P2 declined as designed; on fix(site): keep streaming text alive inside expanded working blocks #29420 the FE9 P1 fixed by wiring the preference query per story again (d0105de, f46888b). ci is green on every head; the separate "AI Documentation Check" fails on four PRs because its dormant doc-check workspace cannot be started by a chat (CODAGT-1025), not because of the diff. Codex re-review requested on the current heads.
• Codex round of 2026-09-20 14:42Z on the current heads: "Didn't find any major issues" on feat(site): add the working block disclosure #29418 b421c87, feat(site): fold assistant steps in the conversation timeline #29419 d0105de, fix(site): keep streaming text alive inside expanded working blocks #29420 f46888b. Findings on feat(site): group assistant steps into working blocks #29417 4d05139 and this PR 73d573c: two P2s, both verified against the code and left open below. No further review rounds will be requested (see the review budget note).
• pr-review-fix pass (2026-09-20, after the 14:42Z round): every unresolved thread across the stack was triaged. Fixed with red-green tests on their owning PRs: feat(site): group assistant steps into working blocks #29417 83d0cb6 caps a block's duration span at the next user message, so a turn that follows a metadata-only prompt and opens with provider-executed parts no longer extends the previous block (Codex P2); feat(site): add the working block disclosure #29418 389d5ba replaces the ResizeObserver height cache with getSnapshotBeforeUpdate, closing the stale-height window DanielleMaywood reproduced (nested expansion and prepend committing before one observer delivery; reproduced and verified in a Chromium probe, over-adjust of exactly the expansion height before, none after), and 11073f5 adds the live-label clock test she asked for; feat(site): fold assistant steps in the conversation timeline #29419 ff84f73 stops an older page's revealed block from aliasing the live block's working:live:head:0 key (Codex P2; duplicate React keys and a remount before); this PR de46163 corrects the setting copy to "Answers and questions stay visible; failed steps are counted on the row." (Codex P2). Thirteen threads were resolved (the four fixes plus DanielleMaywood's and jscottmiller's already-addressed nits and the memberIds coverage); seven stay open, listed below. Current heads: feat(site): group assistant steps into working blocks #29417 83d0cb6, feat(site): add the working block disclosure #29418 11073f5, feat(site): fold assistant steps in the conversation timeline #29419 c8335f4, fix(site): keep streaming text alive inside expanded working blocks #29420 c95ce4e, this PR 1d601ae. No review was re-requested: the stack is past the six-pass budget below, so the earlier Codex OK verdicts do not cover these heads and human review is still required before merge.
• coder-frontend-review loop (2026-09-26, LOC-reduction focus): round 1 ran all eight roles over the composed diff, round 2 re-ran the six that failed, and round 3 the three that still failed; round 3 returned no findings. Fixes landed on their owning PRs (feat(site): group assistant steps into working blocks #29417 ffc8d67 and 5f523b1, feat(site): add the working block disclosure #29418 ea0bec9, feat(site): fold assistant steps in the conversation timeline #29419 6841434 and 685a716, fix(site): keep streaming text alive inside expanded working blocks #29420 054ea94, this PR 8934352) and remove about 416 added site/ lines across the stack. origin/main was merged bottom-up (feat: add a collapse_assistant_steps user preference #29416 c8ba3e9), resolving three conflicts in feat(site): fold assistant steps in the conversation timeline #29419 and moving the stories to the vendored message scroller. DanielleMaywood's getSnapshotBeforeUpdate gating and missing-story threads on feat(site): add the working block disclosure #29418 were fixed and resolved. Declined: extracting a WorkingBlockItem component (adds lines), SmoothText resume after dispose (useSyncExternalStore resubscribes on the StrictMode remount), and building row contents only for visible rows (minor).
• Owner decisions and pr-review-fix pass (2026-09-27): the clock now stops during interrupting (feat(site): fold assistant steps in the conversation timeline #29419 f9b04a4); the live-only identity arm was added with a red-green test (feat(site): fold assistant steps in the conversation timeline #29419 8210761); the feat(site): add the working block disclosure #29418 scroll coverage is deferred to the feat(site): fold assistant steps in the conversation timeline #29419 Pixel story; the feat(site): group assistant steps into working blocks #29417 startedAt comment was deleted (7fb9334); jakehwll's void nit (fix(site): keep streaming text alive inside expanded working blocks #29420 de1636f) and props-bag pattern (this PR 7d1d659) were fixed. A delta coder-frontend-review (eight roles) found two comment fixes and a duplicated test fixture, all fixed. All ten open threads were replied to and resolved. Codex returned "Didn't find any major issues" on feat(site): fold assistant steps in the conversation timeline #29419 8210761. On this PR its rounds on 7d1d659 and 7a71975 left four threads, all fixed: chat load errors no longer wait on the preference (7a71975, red-green via a probe on DetailQueryErrorWhilePreferenceLoads), a CollapseAssistantStepsEnabled story, and two tidy-ups (e238b30). Codex then returned "Didn't find any major issues" on e238b30, CI is green on all five heads, and no threads are open.
• Review budget (completed review passes per PR, every reviewer, history never resets, cap 6): feat: add a collapse_assistant_steps user preference #29416 3 (Codex 2, human 1); feat(site): group assistant steps into working blocks #29417 22 (Codex 5, human 1, agent role passes 16); feat(site): add the working block disclosure #29418 29 (Codex 9, human 4, agent 16); feat(site): fold assistant steps in the conversation timeline #29419 29 (Codex 13, agent 16); fix(site): keep streaming text alive inside expanded working blocks #29420 29 (Codex 13, agent 16); this PR 44 (Codex 28, agent 16). The 16 agent passes are the eight coder-frontend-review roles times two passes over the composed frontend diff on 2026-09-20; agent-side reviews from earlier sessions are not included, so the totals are lower bounds. feat(site): add the working block disclosure #29418, feat(site): fold assistant steps in the conversation timeline #29419, fix(site): keep streaming text alive inside expanded working blocks #29420 and this PR were already past the cap on Codex rounds alone before the 14:42Z request, and feat(site): group assistant steps into working blocks #29417 was past it once the agent passes are counted; the 14:42Z round is the last Codex round. The 2026-09-26 frontend loop added 17 agent passes (8, 6 and 3 roles) to every PR at Mike's explicit request. Codex verdicts for this PR and feat(site): group assistant steps into working blocks #29417 are therefore findings-open on the exact current heads, and the three OK verdicts hold only for the heads listed above. Mike lifted the cap for the 2026-09-27 pass, which added 8 agent passes to every PR and re-requested Codex on feat(site): fold assistant steps in the conversation timeline #29419 and this PR.

This PR was prepared by Xum, an AI coding agent, on behalf of @ibetitsmike.