[xiaozhou26]

Linked issue

Closes #1240

Follows up on the question and permission data from Stack-Cairn/K-brain#162 and Stack-Cairn/K-brain#135 (both in K-brain v0.107.22).

Summary

1. Question headers. Each question tab shows its header. When there is no header, the tab falls back to Q1, Q2, …
2. Multi-select. A question with multiple: true uses checkboxes and a "Select all that apply" hint. The custom answer can be combined with the listed options. K-brain has no field for several labels yet, so when more than one option is picked, the answer is sent as the labels joined with , plus custom: true. A single picked option is sent as a normal answer. When the card is shown again from history, the joined text is split back into the listed options.
3. History replay of the question tool. In history, K-brain's own question calls now render as the question card with the recorded answers, including custom text. A dismissed question shows as cancelled, and one with no answer shows as timed out. Both the single and the questions: [...] forms are handled. The original arguments are kept in a hidden __kbrainQuestionArguments field, so a transcript sent back to K-brain still uses the question name and its original arguments.
4. Gateway direct link. crates/agent-gateway already passed multi-question question.requested events through. It now also handles the rest:
   • It renames and normalizes question calls, both live and in history.
   • It turns question results into card answers.
   • It accepts selectedLabels from the WebUI and sends them as joined text, with custom: true when there is more than one label.
   • It forwards permission.requested.preview to the approval card.
5. Edit approval preview. When permission.requested.preview has a diff, the approval card renders K-brain's diff rows (<line> <mark> <text>), including the … +N more lines overflow note. new_file adds a "This creates a new file." line. truncated with an empty diff says the change is too large to preview. Without a preview, the card falls back to the old preview built from the tool arguments.
6. Command run info. exit_code, timed_out and duration_ms from shell tools (Bash, bash, ManagedProcess, ProcessWait, ProcessStop) are read from tool.result events and from history tool_calls[]. They show in the Shell header, for example exit 1 · 2.3s or timed out · 30.0s. K-brain's native bash has no Shell display, so its result shows the same thing as one line.

K-brain protocol need

A selected_labels: []string answer field, validated against the options when multiple is set. Today a multi-select answer with more than one option has to be sent as joined labels with custom: true. That has two problems:

• K-brain's AskUserQuestion result text presents the answer as something the user typed.
• A label that contains , can't be split back reliably. LiveAgent splits on the longest listed label, which is a best effort.

Observation on K-brain's preview rows (no change needed here)

editDiffRows prints the trailing context row as the last line of the old text, o[len(o)-1]. It does not print the first line after the change. For a whole-file preview, that row is the file's last line. In the edit screenshot below the trailing row is 7 }, while the line right after the change is line 5. LiveAgent renders the rows exactly as K-brain sends them.

Change scope

• Modules: agent-ui (question card, approval card, diff view, tool result display, shared helpers, i18n), agent-gui (K-brain turn, history and question adapters), agent-gateway (Go K-brain client and WebUI tests)
• Key paths:
  • crates/agent-ui/src/components/chat/AskUserQuestionCard.tsx: header tabs and multi-select
  • crates/agent-ui/src/lib/chat/askUserQuestion.ts: question argument normalization, result parsing, multi-select answers
  • crates/agent-ui/src/lib/chat/toolApprovalArgs.ts: approvalPreviewFromKBrain()
  • crates/agent-ui/src/lib/chat/diffRows.ts (new), crates/agent-ui/src/components/chat/EditDiffView.tsx, crates/agent-ui/src/components/chat/ToolApprovalBar.tsx: preview rows, new file and too-large messages
  • crates/agent-ui/src/lib/chat/toolRunInfo.ts (new), crates/agent-ui/src/components/chat/assistant-bubble/ToolResultDisplay.tsx: run info
  • crates/agent-gui/src/lib/kbrain/{history,questions,turn,types}.ts, crates/agent-gui/src/pages/chat/turns/runKBrainConversationTurn.ts
  • crates/agent-gateway/internal/kbrain/question_tool.go (new), events.go, history.go, interactive.go, client.go

Screenshots / preview

All screenshots come from the real desktop debug app (macOS arm64). It was a cargo build of this branch with the Vite dev server from this worktree, and it ran the real K-brain v0.107.22 release binary through LIVEAGENT_KBRAIN_BINARY. HOME and LIVEAGENT_DEV_HOME pointed at temporary directories. The model was a local scripted OpenAI-compatible fixture. It picks the tool from the request's tool list and returns:

• a question call with three questions (Format, Checks with multiple: true, Rollout)
• an edit on an existing file
• a write of a new file
• a bash command that exits with status 3

K-brain produced everything else itself: the question.requested events, the permission requests with preview (in Always ask mode), the history, and the run info. The screenshots are only cropped (sidebar removed) and not otherwise edited.

Multi-question card with headers (live question.requested from the question tool):

Multi-select question (two options checked):

Edit approval rendered from permission.requested.preview:

Write approval for a new file (new_file: true):

History replay (after a reload): the question tool call from K-brain history shows as the settled card, and the joined multi-select answer is split back into its options:

Run info on K-brain's native bash, read from history tool_calls[] (exit_code: 3, duration_ms: 16):

The Gateway direct link was not run live in this pass. It is covered by the Go tests and the WebUI tests listed below.

Verification

• pnpm typecheck:ui, pnpm typecheck:gui, pnpm typecheck:webui: pass
• pnpm lint (biome): no errors
• pnpm test:gui: 4122 pass, 12 skipped, 0 fail. The backend tests ran with LIVEAGENT_KBRAIN_BINARY set to the v0.107.22 release binary.
• pnpm test:webui: 887 pass
• cd crates/agent-gateway && go test ./...: pass (after pnpm --filter @liveagent/gateway-webui build)
• New tests:
  • test/providers/kbrain-question-tool.test.mjs: normalization, result parsing, multi-select answers, history replay and the round trip back to K-brain, run info, and a live turn with the question tool and a Bash exit code
  • internal/kbrain/question_tool_test.go: the Gateway side of the same cases, plus the preview reaching the approval card
  • web/test/kbrain-question-preview.test.mjs
• Extended tests: the question card (header tabs, multi-select, settled joined answers), the approval card (preview rows, new file, too large), the diff rows parser, a full turn where the permission preview reaches both the desktop and the WebUI, and run info on generic results.
• The e2e run above also caught a bug that the unit tests missed: K-brain's native bash result showed the run info as raw JSON. It is fixed in the second commit and covered by a test.