Feature request: fold pre-result thinking/tool rows into a "work" group in the chat view

[mattamior]

Motivation

Long tool-heavy turns leave the chat view crowded with intermediate process rows — consecutive Think/reasoning blocks and tool calls — even after the turn has settled. The reader has to scroll past all of that to reach the final answer.

Proposal

Fold each turn's pre-result activity into one collapsible group behind a work · {duration} header, gated by a Settings toggle:

• Consecutive reasoning/tool rows collapse into a single group header once the turn settles; the group stays expanded while the agent is running.
• The turn closing answer always stays visible outside the group (never folded).
• One Settings toggle owns the preference end-to-end.

[mattamior]

Implementation ready

A working implementation of this feature is already prepared on my fork, ready for review:

• Branch: fold-work-rows
• Single commit d72e4a1 — 21 files, +1053/−32

What it does:

• New GroupFold primitive (ui-primitives) wraps a run of activity rows in a shared DisclosureRow header with aggregate run state.
• flow-grouping (ui-conversation) classifies activity nodes, folds maximal consecutive runs, and keeps boundary rows (closing answer, user/steering messages, turn footer, terminal failure notices) as standalone rows so the result is never swallowed.
• One fold-rows Settings toggle owns the preference end-to-end; the group stays expanded while running and folds once the turn settles.

Verification: focused vitest run on the four owning specs (85 passed); tsc -b tsconfig.client.json and oxlint clean.

I could not open a pull request because the repository currently has pull requests disabled. Happy to open one as soon as that is enabled, or point review at the fork branch directly.

[mattamior]

Update: fold is now the default

Follow-up commit pushed to fold-work-rows (885e97d): the work-group fold now ships enabled instead of opt-in. DEFAULT_FOLD_ROWS = true feeds both the durable settings schema default and the process-local policy store, so a fresh session folds pre-result reasoning/tool rows without visiting Settings. The fold-rows toggle still owns the preference end-to-end for opting out.

Verification: focused vitest run on the owning specs (updated for the new default), plus tsc -b tsconfig.client.json clean via the pre-push gate. Branch is now two commits: d72e4a1 (feature) + 885e97d (default on).

[little3tar]

补充一个 #4863 实现需要覆盖的用例，供 fold-work-rows 分支参考：

问题不在"该不该折过程行"，而在折叠上界只认"最后一段答复"。 本机 0.1.5-rc.2 的 @deepseek-ai/dsh-client-ui-chat/lib/client.js 里：

// L6751-6755：折叠锚点只看本轮最后一个 step
function latestAnswer(turn) {
  const data = turn.steps.at(-1)?.data.get("assistant-step");
  if (!isFinalAssistant(data) || !hasAssistantReplyContent(data.blocks)) return null;
  return data.blocks.some((block) => block.kind === "tool-call") ? null : data;
}
// L6781：折叠上界 = 那一段的起点
answerAnchorSeq: answer.finalNode.seq,
// L1556：于是它之前的一切都成了 processMember
const processMember = … && routedNode.anchorSeq < processSpec.answerAnchorSeq;

结果：一轮中间的助手正文段落也被折进「过程」，不只有 thinking/tool 行。实测一轮 正文A（完整诊断 + 选项）→ ask_user_question → 用户回答 → 正文B（"明白…"收尾），轮次结束后界面上只剩正文B，正文A 要展开折叠行或切「轨迹」才能找回；「标准」模式下正文A 正常可见。

对 fold-work-rows 的两点建议：

1. 本 PR 的 "closing answer always stays visible (never folded)" 建议改成 "every user-facing answer stays visible"：按节点判定，凡 hasAssistantReplyContent 的 text block（无论它在第几个 step）都留在折叠组之外，只把 reasoning / tool / 其它过程行折进组里。
2. 若为控制篇幅确实要折中间正文，折叠组的标题上应体现（如「work · 含 2 段说明」），避免"关键结论静默消失"。

完整复现与定位已单独开贴：#6254