feat(cli): dense inline panel + keyboard navigation for parallel agent fan-out#4477
Merged
Conversation
When a tool group is composed entirely of multiple `task_execution`
agent invocations (the shape produced by `/review`'s 9-agent fan-out),
`CompactToolGroupDisplay` collapses the whole group into one line —
`Agent × 9 / <last agent's name>` — even though the agents may run
for minutes each. Status, current activity, elapsed, and tokens are
all hidden, so the most informative span of a multi-agent run is
also the most opaque.
Route those groups to a new `InlineParallelAgentsDisplay` that
surfaces every agent on its own row:
╭─ Parallel agents · 9 · 3/9 done ───────────────────────╮
│ ○ Agent 1: Correctness ReadFile server.ts 3m 38s · 7.0k tok
│ ✔ Agent 2: Security 12s · 8.1k tok
│ ○ Agent 3: Code Quality ReadFile index.ts 3m 38s · 9.0k tok
│ ...
╰────────────────────────────────────────────────────────╯
Each row pulls activity / elapsed live from `BackgroundTaskRegistry`
on a 1s tick (same pattern LiveAgentPanel uses). When the registry
unregisters a finished foreground subagent, the row falls back to
`AgentResultDisplay.executionSummary` for elapsed + tokens so the
column doesn't blank out the moment an agent completes.
A new `InlineAgentClaimContext` lets the inline display claim the
agentIds it's rendering; `LiveAgentPanel` filters those out so the
same agent never appears in both surfaces. Claims are refcounted so
React's commit/cleanup interleave doesn't transiently drop a claim
while the same id is being re-claimed by a remount.
Routing conditions in `ToolGroupMessage`:
- Live phase only (`isPending`). Once the group commits to
`<Static>`, the expanded path with `SubagentScrollbackSummary`
owns the permanent record.
- Group has ≥2 calls and ONLY agent calls — mixed groups keep
the legacy renderer so sibling tools stay visible.
- No pending confirmation — those go through the normal renderer
so the keyboard-focus surface for approval stays intact.
Verified visually in tmux with `/review <pr-url> --comment` against
qwen-code PR 4472: all 9 agents render with live status / activity /
elapsed / tokens, transition glyphs ○→✔ in place, and the footer
LiveAgentPanel correctly suppresses the duplicate rows.
11 new tests + all 57 prior tests on touched files pass.
Contributor
📋 Review SummaryThis PR introduces 🔍 General Feedback
🎯 Specific Feedback🟡 High
🟢 Medium
🔵 Low
✅ Highlights
|
Contributor
Code Coverage Summary
CLI Package - Full Text ReportCore Package - Full Text ReportFor detailed HTML reports, please see the 'coverage-reports-22.x-ubuntu-latest' artifact from the main CI run. |
wenshao
commented
May 24, 2026
wenshao
left a comment
Collaborator
Author
There was a problem hiding this comment.
Summary
Well-structured feature PR that adds a dense inline panel for parallel agent fan-out. The architecture is sound — claim context prevents duplicate rows across surfaces, the 1s tick pattern mirrors LiveAgentPanel, and the routing conditions are clearly documented. tsc clean, ESLint clean, 12 new tests + 57 existing tests all pass.
Two test gaps flagged below as Suggestions; six additional low-confidence items noted in the terminal review output.
— qwen3.7-max via Qwen Code /review
| // suppresses the same rows — no duplication across surfaces. | ||
| const isPureAgentGroup = | ||
| toolCalls.length >= 2 && toolCalls.every(isSubagentToolEntry); | ||
| if (isPending && isPureAgentGroup && !hasSubagentPendingConfirmation) { |
Collaborator
Author
There was a problem hiding this comment.
[Suggestion] The routing branch isPending && isPureAgentGroup && !hasSubagentPendingConfirmation is the primary integration point for the feature but has no test coverage. The existing ToolGroupMessage.test.tsx (32 tests) does not exercise any of the five predicate conditions: (a) isPending=true + pure agent group routes to InlineParallelAgentsDisplay; (b) isPending=false falls through; (c) mixed groups fall through; (d) single agent falls through; (e) pending confirmation falls through.
A regression in any predicate would silently skip the new panel or incorrectly route groups. Consider adding tests for at least the happy path and two fall-through cases.
— qwen3.7-max via Qwen Code /review
- Split InlineAgentClaimContext into read + write contexts so claimers (InlineParallelAgentsDisplay) don't re-render every time some unrelated inline display claims its own ids. The write API now has a stable identity for the lifetime of the provider; only readers (LiveAgentPanel) react to claim-set changes. Add a test that locks in the stability invariant so a future memo on `[version]` is caught before it ships. - Extract `isPureParallelAgentGroup` predicate in ToolGroupMessage alongside the existing `isSubagentToolEntry` / `isTerminalSubagentTool` helpers so the routing intent reads at the call site instead of being spelled out inline. - Document that `recentActivities` is mutated in place by the registry; the rows useMemo's `now`-keyed re-read is what surfaces the latest entry, so `activityLabel` must treat the value as a tick-snapshot rather than closing over the live array. - Document the exhaustiveness of the `hasLiveAgent` status check (`running | background`) against the upstream `AgentResultDisplay.status` union so a future status addition is caught. - Document why `NAME_COL_WIDTH = 26` (fits /review's longest agent label at full length on typical 100-col widths) instead of leaving it as a magic number. - Add an inline comment in gemini.tsx explaining why InlineAgentClaimProvider sits inside BackgroundTaskViewProvider (visible to both surfaces, no provider-boundary crossing).
wenshao
commented
May 24, 2026
wenshao
left a comment
Collaborator
Author
There was a problem hiding this comment.
Second-opinion review (glm-5.1). 9 agents + verification + reverse audit. No Critical issues found. 3 new Suggestions below — code duplication, test gaps for the 1s timer tick, and missing dev-mode logging for registry misses. Prior suggestions (ToolGroupMessage routing tests, LiveAgentPanel isClaimed filter tests) remain open.
| tokenCount?: number; | ||
| } | ||
|
|
||
| // Internal tool name → display name lookup (mirrors LiveAgentPanel so |
Collaborator
Author
There was a problem hiding this comment.
[Suggestion] TOOL_DISPLAY_BY_NAME and activityLabel are structurally identical to their counterparts in LiveAgentPanel.tsx. The comment says "mirrors LiveAgentPanel" but no extraction was done.
If ToolNames/ToolDisplayNames mapping changes or activity formatting logic evolves, both sites must be kept in sync manually — a silent divergence would make the two panels render the same tool differently.
Suggested change
| // Internal tool name → display name lookup (mirrors LiveAgentPanel so | |
| // Extract to a shared utility, e.g. packages/cli/src/ui/utils/toolDisplayUtils.ts: | |
| // export const TOOL_DISPLAY_BY_NAME = Object.fromEntries( | |
| // (Object.keys(ToolNames) as Array<keyof typeof ToolNames>).map((key) => [ | |
| // ToolNames[key], ToolDisplayNames[key], | |
| // ]), | |
| // ); | |
| // export function formatActivityLabel(name: string, description?: string): string { ... } |
— glm-5.1 via Qwen Code /review
| ms = Math.max(0, end - row.startTime); | ||
| } else if (row.fallbackElapsedMs !== undefined) { | ||
| ms = Math.max(0, row.fallbackElapsedMs); | ||
| } |
Collaborator
Author
There was a problem hiding this comment.
[Suggestion] The 1-second setInterval tick that drives live activity/elapsed updates is never exercised in tests. vi.useFakeTimers() is set up in beforeEach but vi.advanceTimersByTime(1000) is never called — the entire live-update loop is dead code in the test suite.
Additionally, statusGlyph branches for failed/cancelled/background (line ~175) and the startTime+endTime elapsed path (line ~200) are untested.
Suggested test additions:
- Render a running agent, assert initial elapsed, then
vi.advanceTimersByTime(1000)and assert elapsed updated - Render agents with
failed/cancelled/backgroundstatus and assert correct glyphs +doneCount - Provide a registry entry with both
startTimeandendTimeand assert fixed-duration elapsed
— glm-5.1 via Qwen Code /review
| const recent = live?.recentActivities?.at(-1); | ||
| return { | ||
| agentId, | ||
| callId: toolCall.callId, |
Collaborator
Author
There was a problem hiding this comment.
[Suggestion] When registry?.get(agentId) returns undefined for a running agent (e.g. deriveAgentId diverges from the core construction, or the registry evicts the entry early), the row silently falls back to empty startTime/recentActivity with no diagnostic output. The panel renders correctly but shows stale/empty data — indistinguishable from "agents are slow."
Consider adding a dev-mode warning:
Suggested change
| callId: toolCall.callId, | |
| const rows: RowData[] = useMemo(() => { | |
| const registry = config?.getBackgroundTaskRegistry(); | |
| void now; | |
| if (process.env.NODE_ENV === 'development') { | |
| const liveButMissing = agentEntries | |
| .filter((e) => { | |
| const agentId = deriveAgentId(e.toolCall, e.result); | |
| return (e.result.status === 'running' || e.result.status === 'background') | |
| && registry?.get(agentId) === undefined; | |
| }); | |
| if (liveButMissing.length > 0) { | |
| console.warn( | |
| '[InlineParallelAgentsDisplay] %d running agents not found in registry;' | |
| + ' deriveAgentId may be out of sync', | |
| liveButMissing.length, | |
| ); | |
| } | |
| } |
— glm-5.1 via Qwen Code /review
…anel - Dense panel renders only in committed phase, replacing verbose per-agent ToolMessage expansion with one compact row per agent. - Live phase: pure parallel agent groups return null so LiveAgentPanel below the input is the sole live surface (fixes terminal agents rendering as verbose ToolMessages above input during live phase). - Remove InlineAgentClaimContext (no longer needed without live-phase inline panel).
- Show InlineParallelAgentsDisplay in both live and committed phases with all agents (running + completed). LiveAgentPanel below the composer also shows running agents during live phase (brief overlap that resolves as agents complete and expire from the panel). - Increase LiveAgentPanel DEFAULT_MAX_ROWS from 5 to 12 so all 9 /review agents are visible without truncation. - Add totalAgentCount prop to InlineParallelAgentsDisplay for correct header when rendering a subset of agents.
wenshao
commented
May 24, 2026
wenshao
left a comment
Collaborator
Author
There was a problem hiding this comment.
5 Suggestions (4 test gaps + 1 comment accuracy) and 3 Nice to have items. No Critical issues. Build passes, 70 tests pass, tsc/ESLint clean.
— qwen3.7-max via Qwen Code /review
| return { glyph: '○', color: theme.status.warning }; | ||
| case 'completed': | ||
| return { glyph: '✔', color: theme.status.success }; | ||
| case 'failed': |
Collaborator
Author
There was a problem hiding this comment.
[Suggestion] failed and cancelled arms of statusGlyph are never exercised by any test. The done-tally test only uses status: 'completed' — no test passes status: 'failed' or status: 'cancelled'. Consequently, the doneCount filter that counts failed/cancelled as terminal is also unverified.
Add a test with a mix of terminal statuses:
it('counts failed and cancelled agents in the done tally', () => {
const toolCalls = [
agentToolCall({ callId: 'c1', subagentName: 'g', taskDescription: 'A1', status: 'failed' }),
agentToolCall({ callId: 'c2', subagentName: 'g', taskDescription: 'A2', status: 'cancelled' }),
agentToolCall({ callId: 'c3', subagentName: 'g', taskDescription: 'A3', status: 'running' }),
];
const { lastFrame } = renderInline({ toolCalls });
expect(lastFrame() ?? '').toContain('2/3 done');
});— qwen3.7-max via Qwen Code /review
| // and the trailing suffix stay readable. | ||
| const NAME_COL_WIDTH = 26; | ||
|
|
||
| function truncateMiddle(input: string, max: number): string { |
Collaborator
Author
There was a problem hiding this comment.
[Suggestion] truncateMiddle has three branches (input fits, max <= 1, and actual middle-truncation with Math.ceil/Math.floor head/tail math), but only the first branch is implicitly exercised — all test agent names are ≤26 characters, well under NAME_COL_WIDTH = 26.
Real /review agent labels like "Agent 6c: Maintainer review" exceed 26 chars and WILL trigger truncation. Add a test with a long name:
agentToolCall({ callId: 'c1', subagentName: 'g',
taskDescription: 'Agent 1: Correctness deep review of all modules' }),
// Assert frame contains '…' and the leading prefix portion— qwen3.7-max via Qwen Code /review
| ? { name: recent.name, description: recent.description } | ||
| : undefined, | ||
| tokenCount: | ||
| result.tokenCount ?? |
Collaborator
Author
There was a problem hiding this comment.
[Suggestion] The token-count fallback chain result.tokenCount ?? live?.stats?.totalTokens ?? result.executionSummary?.totalTokens has three tiers, but only tier 3 is tested. Tier 1 (result.tokenCount) and tier 2 (live?.stats?.totalTokens from the registry for a running agent) are never exercised.
Tier 2 is the primary path for live running agents — the registry accumulates stats.totalTokens while result.tokenCount remains undefined until completion. Extend the "surfaces live activity" test's registry entry to include stats: { totalTokens: 5000 } and assert 5k tok appears. Separately, pass tokenCount: 800 through the seed and assert it takes priority.
— qwen3.7-max via Qwen Code /review
| (Object.keys(ToolNames) as Array<keyof typeof ToolNames>).map((key) => [ | ||
| ToolNames[key], | ||
| ToolDisplayNames[key], | ||
| ]), |
Collaborator
Author
There was a problem hiding this comment.
[Suggestion] The JSDoc claims appendActivity "intentionally mutates that array in place…not by replacing the reference" — but background-tasks.ts:482-494 actually does the opposite:
const next = [...prior, activity]; // new array via spread
entry.recentActivities = next; // reassignment, not mutationThe entry object reference is stable but recentActivities is reassigned on every append. The rows memo still works because it re-reads registry.get(agentId).recentActivities.at(-1) on each tick, but the comment's reasoning about "in-place mutation" is factually wrong and would mislead a future maintainer attempting to optimize the memo.
Correct the comment to describe the actual mechanism.
— qwen3.7-max via Qwen Code /review
This was referenced May 25, 2026
- Add "main" as first entry in LiveAgentPanel, followed by running agents - ↓ from Composer focuses LiveAgentPanel (selects "main") - ↓/↑ navigates between entries with ▸ selection indicator - Enter on agent opens BackgroundTasksDialog directly in detail mode - Esc or ↑-at-top returns focus to Composer - Printable chars auto-unfocus and type through to Composer - Simplify header from "Active agents (N/N)" to "Active agents (N)" - Add livePanelFocused + livePanelSelectedIndex state to BackgroundTaskViewContext, keyboard handling stays in InputPrompt
wenshao
commented
May 25, 2026
wenshao
left a comment
Collaborator
Author
There was a problem hiding this comment.
6 Critical + 6 Suggestion findings in this round (4th opinion). All prior open comments acknowledged.
— qwen3.7-max via Qwen Code /review
| setLivePanelFocused(false); | ||
| return true; | ||
| } | ||
| if (key.sequence && key.sequence.length === 1 && !key.ctrl && !key.meta) { |
Collaborator
Author
There was a problem hiding this comment.
[Suggestion] Two key-handling issues in this catch-all area:
- Ctrl+C swallowed: The
return trueon the next line silently consumes Ctrl+C. In raw terminal mode (Ink), Ctrl+C doesn't generate SIGINT — it's a regular keypress. Users navigating the agent panel can't abort with Ctrl+C without first pressing Esc. - Tab/Backspace match printable-char check:
key.sequence.length === 1matches\t(Tab),\x7f(Backspace), and other non-printable single-byte sequences. These unfocus the panel and fall through to BaseTextInput, potentially inserting a tab or deleting a character.
Suggested change
| if (key.sequence && key.sequence.length === 1 && !key.ctrl && !key.meta) { | |
| if (key.ctrl && key.name === 'c') { | |
| setLivePanelFocused(false); | |
| return false; | |
| } | |
| if (key.sequence && key.sequence.length === 1 && !key.ctrl && !key.meta | |
| && key.name !== 'tab' && key.name !== 'backspace' && key.name !== 'delete') { | |
| setLivePanelFocused(false); | |
| return false; | |
| } |
— qwen3.7-max via Qwen Code /review
| // input — so reverse here back to ASC for rendering. Doing the | ||
| // reverse before the slice means the tail-window math (drop the | ||
| // OLDEST when the list overflows) stays unchanged. | ||
| const visibleAgentsAsc = [...visibleAgents].reverse(); |
Collaborator
Author
There was a problem hiding this comment.
[Suggestion] The diff deletes the 4-line comment explaining WHY [...visibleAgents].reverse() exists. The remaining code is a non-obvious decision — entries arrive newest-first from useBackgroundTaskView but the panel renders top-to-bottom in launch order (oldest at top, newest at bottom above the composer). Without this comment, a future cleanup could plausibly remove the .reverse() as "unnecessary."
Suggested change
| const visibleAgentsAsc = [...visibleAgents].reverse(); | |
| // Entries arrive newest-first from context; reverse so the panel | |
| // reads oldest→top, newest→bottom (above the composer). | |
| const visibleAgentsAsc = [...visibleAgents].reverse(); |
— qwen3.7-max via Qwen Code /review
| <AgentRow key={entry.agentId} entry={entry} now={now} /> | ||
| {visible.map((entry, idx) => ( | ||
| <AgentRow | ||
| key={entry.agentId} |
Collaborator
Author
There was a problem hiding this comment.
[Suggestion] The hint text "^ N more above (↓ to view all)" was unchanged, but its meaning shifted with this PR. Previously ↓ from the composer opened the dialog. Now ↓ enters the panel's keyboard navigation, which does NOT scroll — overflow entries (the oldest agents, dropped by .slice(-maxRows)) are unreachable via keyboard. The hint should be updated (e.g., "Enter on main to view all in dialog") or the panel should support scroll-through-window during keyboard navigation.
— qwen3.7-max via Qwen Code /review
| * (e.g. only terminal agents during the live phase). When omitted, | ||
| * defaults to the number of agent entries in `toolCalls`. | ||
| */ | ||
| totalAgentCount?: number; |
Collaborator
Author
There was a problem hiding this comment.
[Suggestion] totalAgentCount is declared as an optional prop with JSDoc, but the sole caller in ToolGroupMessage.tsx never passes it. The prop describes a scenario ("toolCalls is a subset") that no code path exercises. Per the project's simplicity-first principle, remove the prop and simplify to const total = rows.length. If a future caller needs it, add it then.
Suggested change
| totalAgentCount?: number; | |
| contentWidth: number; | |
| } |
— qwen3.7-max via Qwen Code /review
| agentId, | ||
| callId: toolCall.callId, | ||
| name: result.taskDescription || result.subagentName, | ||
| status: result.status, |
Collaborator
Author
There was a problem hiding this comment.
[Suggestion] status: result.status reads from the static toolCalls prop, while activity/elapsed/tokens reconcile from the live registry on each tick. When an agent finishes and unregisterForeground fires, the registry entry disappears but this glyph stays ○ (running) until the parent re-renders with an updated toolCalls prop. During the same window, activity and elapsed are correctly refreshed from the registry — creating a visual contradiction: a "running" glyph next to stale activity.
Consider reading live?.status from the registry in the rows useMemo and preferring it over result.status when available.
— qwen3.7-max via Qwen Code /review
| {headerLabel} | ||
| </Text> | ||
| </Box> | ||
| {rows.map((row) => ( |
Collaborator
Author
There was a problem hiding this comment.
[Suggestion] rows.map renders every agent without any maxRows cap or windowing. LiveAgentPanel has maxRows = 12 and windows from the tail. For a /review 9-agent fan-out during the live phase, the user sees a bordered 12-line panel here PLUS up to 12 more rows in LiveAgentPanel below the composer — 20+ terminal lines of duplicate content. On smaller terminal heights, conversation context is pushed far off-screen.
Consider adding a maxRows prop with the same tail-windowing pattern, or gate this component on !isPending so the live phase renders only LiveAgentPanel and the committed phase renders only this panel.
— qwen3.7-max via Qwen Code /review
Add livePanelFocused, livePanelSelectedIndex, enterBgDetailFromPanel, setBgSelectedIndex, bgEntries.length to InputPrompt handleInput deps. Add isDetailMode to BackgroundTasksDialog effect deps. Remove unused setBgPillFocused import.
- Use bgAgentCount (agent-only) instead of bgEntries.length for keyboard navigation bounds (was counting shells/monitors/dreams). - Auto-clear livePanelFocused when no agent entries remain (prevents stuck focus state when LiveAgentPanel returns null).
pomelo-nwu
previously approved these changes
May 26, 2026
pomelo-nwu
left a comment
Collaborator
There was a problem hiding this comment.
Tested locally with tmux — looks great
Built the branch and ran a few parallel fan-outs in tmux. The dense inline panel + LiveAgentPanel coexistence works as advertised. Approving with a couple of small things to flag.
Screenshot 1 — Live phase (5 parallel agents, 4 done + 1 running)
╭──────────────────────────────────────────────────────────────────────────╮
│ Parallel agents · 5 · 4/5 done │
│ ✔ Search useState count 42s · 1.0k tok │
│ ○ Search useEffect count Glob '**/*.tsx' in path '/… 2m 8s · 921 tok │
│ ✔ Search useMemo count 51s · 2.3k tok │
│ ✔ Search useCallback count 41s · 1.2k tok │
│ ✔ Search useRef count 46s · 1.3k tok │
╰──────────────────────────────────────────────────────────────────────────╯
... (composer)
YOLO mode (Shift + Tab to toggle) · 1 local agent, 1 monitor
main
○ Search useEffect count (Glob '**/*.tsx' in path … ▶ 2m 8s · 182k tokens
Dense panel keeps all 5 rows with status / activity / elapsed / tokens; LiveAgentPanel below the composer correctly filters down to only the still-running one. The "brief visual overlap" you described in the PR is barely noticeable in practice — clean.
Screenshot 2 — Keyboard navigation (↓↓ from input)
main
▸ ○ Search useEffect count (Glob '**/*.tsx' in path… ▶ 2m 46s · 182k tokens
↑↓ navigate · Enter detail · Esc back
▸ indicator, the navigate hint line, and the focus transitions all render correctly.
Findings worth a look
1. CI Lint is red — must-fix before merge. This is what's failing CI, not a real test failure:
packages/cli/src/ui/components/InputPrompt.tsx
1301:5 warning React Hook useCallback has missing dependencies:
'bgEntries.length', 'enterBgDetail', 'livePanelFocused',
'livePanelSelectedIndex', 'setBgSelectedIndex', 'setLivePanelFocused',
'setLivePanelSelectedIndex'.
✖ 1 problem (0 errors, 1 warning)
ESLint found too many warnings (maximum: 0).
Easy fix; just add the missing deps (or intentionally suppress with a comment if any need to be stable).
2. Enter-to-detail mapping may be off when monitor tasks are mixed in. Repro: LiveAgentPanel ▸ selecting an agent row (grep useMemo in tsx), then Enter — but the detail that opened was a different [monitor] Wait for sleep 25 ... task, not the selected agent. The selectedIndex from the panel doesn't seem to map cleanly to the dialog's target when the background-task list contains both agents and monitors. Worth a second look — this is core to the new keyboard-navigation UX.
3. ← from detail doesn't return to panel selection (at least for monitor detail). PR description says: "← from detail: returns to LiveAgentPanel selection (not the dialog list)". In my repro ← took me back to the BackgroundTasksDialog list view instead. Could be that the detail-from-panel mode only kicks in for agents and monitors fall through to the old path — if so, worth a comment in the code.
4. Stale comment in InlineParallelAgentsDisplay.tsx:15. Header doc says "Rendered in the committed phase only; during the live phase LiveAgentPanel below the composer owns the per-agent roster." — but the implementation now renders in both phases (per commit 205fa024 "dense panel both phases"). Quick comment tidy.
Verdict
Core feature (dense panel) is a huge UX improvement over Agent × 9, design is right. Approving. The two keyboard-nav issues (#2, #3) aren't blockers for me — feature still useful without them — but please address before they become muscle memory.
中文说明（点开）
本地拉分支用 tmux 跑了几次并行 fan-out 验证，dense panel + LiveAgentPanel 同屏共存确实如描述所说。Approve，但有几点提一下：
截图 1：5 个并行 agent，4 个 ✔ 完成 + 1 个 ○ 进行中，dense panel 完整保留 5 行（含状态/activity/耗时/token），LiveAgentPanel 自动过滤只剩还在跑的那个。"brief visual overlap" 在实际使用中几乎不察觉，干净。
截图 2：键盘导航 ↓↓ 后 ▸ 指示器和 ↑↓ navigate · Enter detail · Esc back 提示都正常出现。
需要看一下的几点：
-
CI Lint 红灯，合并前必修 —— 不是测试失败，是
InputPrompt.tsx:1301的 useCallback 缺 7 个依赖，max-warnings 0把 CI 顶掉了。 -
Enter 进 detail 时索引疑似错位：LiveAgentPanel 上
▸明明选中的是 agent 行，Enter 后打开的却是 monitor 任务的 detail。当后台列表里 agent + monitor 混合时映射有问题，这是新键盘导航的核心 UX，值得排查。 -
← 返回不符合 PR 描述：PR 说"← from detail: returns to LiveAgentPanel selection (not the dialog list)"，实测 ← 回到的是
BackgroundTasksDialoglist 而不是 panel。可能detail-from-panel模式只对 agent 生效，monitor 没走新路径。 -
文件头注释过时：
InlineParallelAgentsDisplay.tsx:15注释还写"Rendered in the committed phase only"，但 commit205fa024已改成 both phases。
核心功能体验比旧 Agent × 9 提升明显，方向对。Approve。键盘导航的 2/3 不算 blocker（即使有问题主功能也已经有用），但建议尽快修复，避免变成肌肉记忆。
Collaborator
Author
🧪 本地构建验证报告 (维护者)在隔离 worktree 中对 PR 合并后状态做了完整验证。整体方向 OK, 视觉效果与 PR mockup 一致, 但有一处关联测试遗漏需要补。 验证环境
静态检查
单元测试
|
| 维度 | 状态 |
|---|---|
| 构建 / 类型 / lint | ✅ 全清 |
| PR 自带 3 个测试文件 (63 用例) | ✅ 全过 |
| 视觉渲染 vs PR mockup | ✅ 三种场景帧对帧一致 |
| 真实 CLI 烟测 | ✅ 启动 / 无 agent 状态下 ↓ 不崩 |
关联测试 (AgentTabBar.test.tsx) |
建议: 补上 AgentTabBar.test.tsx 的 3 行 mock + 断言更新, 再 merge. 改动行为本身 (Ctrl+N → setLivePanelFocused) 与 PR 整体的"输入框 → LiveAgentPanel"导航链是一致的, 只是 PR 没把这个旁路 test 跟着更新, 而 vi.mock(...) + as never 联合压住了 TS 在 mock 不完整时本该报的错, 才让这个回归在作者本地的 typecheck/vitest run 改动文件 时没暴露出来.
Diff 本身约束良好, 实现按照 PR 描述的"≥2 agent / 无 non-agent tool / 无 approval pending"门控, LiveAgentPanel 与 InlineParallelAgentsDisplay 的双阶段渲染策略也符合 mockup, 整体可以 merge.
pomelo-nwu
approved these changes
May 26, 2026
Collaborator
Author
✅ 本地真实测试验证报告 (PR #4477)环境：Linux 6.12 / Node 22 / 本地 tmux + 真实模型（glm-4.7）/ 仓库已 checkout 到该 PR 分支 (HEAD 一、自动化测试
二、tmux 真机交互测试在 tmux (200×60) 中启动 1. 密集 inline 面板（
|
| 步骤 | 操作 | 实际渲染（捕获自 tmux） | 判断 |
|---|---|---|---|
| 1 | 输入框空，按 ↓ | ▸ main + 2 行 ○ 执行 50 次 grep 搜索 + 提示行 ↑↓ navigate · Enter detail · Esc back |
✅ 焦点正确进入面板，初始选中 main，提示行只在 focused 时出现 |
| 2 | 再按 ↓ | main 上无 ▸，第一个 agent 行变 ▸ ○ ... |
✅ 选择下移到 agent 1 |
| 3 | 再按 ↓ | ▸ 移到 agent 2 |
✅ 选择下移到 agent 2 |
| 4 | agent 2 行再按 ↓ | ▸ 仍停在 agent 2 |
✅ 边界 clamp 正确（if (livePanelSelectedIndex < maxIdx)） |
| 5 | 按 ↑ | ▸ 上移到 agent 1，且 activity 列变为实时内容 (Grep 'paint\s+\w+' in path ...) |
✅ 选择上移；同时确认 1s tick 持续刷新 activity |
| 6 | agent 行按 Enter | 弹出 BackgroundTasksDialog，标题 general-purpose › 执行 50 次 grep 搜索，含 Progress trail / Prompt 摘要 / token 计数，底部提示 ← back · Esc close · x stop |
✅ 进入 detail-from-panel 模式，enterBgDetailFromPanel() + 预选目标 agent 都生效 |
| 7 | 详情中按 ← | 关闭详情，回到面板，▸ 仍停在刚才那个 agent |
✅ exitDetail 在 'detail-from-panel' 模式下正确返回 panel（而非 list），保留选择 |
| 8 | Esc 路径 | 走代码 if (key.name === 'escape') { setLivePanelFocused(false); return true; } |
由于 grep agent 在按键之间完成、面板已自动 unfocus（hasAgentEntries 变 false 触发 setLivePanelFocusedRaw(false)），Esc 透传到了模型 turn 的取消路径。属于设计内行为：面板不在了就没什么好释放。代码路径本身正确 |
备注：截屏文件已保存到
/tmp/qwen-pr-test/nav-*.txt共 9 份，每一步的 tmux pane 都有原始捕获。
3. 视觉密度对比
旧版（被取代）：Agent × 9 / <last name> 折叠成一行，N 分钟跑过去只看到一个名字。
新版（本 PR）：N 行 × (状态 · 名字 · activity · 耗时 · token)，9 行就完整覆盖 /review 全部并发面，信息密度提升一个数量级，且 elapsed/token 一直右对齐到边界，肉眼可扫。
三、Edge-case 复盘
- ✔ Pending approval 抑制 dense 面板：测试中 4 个 agent 同时被起，其中一个等审批，dense 面板自动让位给审批 UI（
isPureParallelAgentGroup && !hasSubagentPendingConfirmation守卫） - ✔ Queued approval 渲染：第二个 agent 在排队等待时显示
⏳ Queued approval: general-purpose，符合 first-come-first-served 锁 - ✔ 终态后 8s 内可见：单 agent 完成后短暂保留在 LiveAgentPanel 顶部（
TERMINAL_VISIBLE_MS = 8000），随后自动剔除 - ✔ 元数据兜底链：
tokenCount ?? live?.stats?.totalTokens ?? executionSummary?.totalTokens在 unregisterForeground 之后仍能正确展示 - ✔ 超长 agent 名字截断：
NAME_COL_WIDTH = 26+ middle-truncate 工作正常（本次测试未触发，但代码逻辑正确）
四、结论
| 维度 | 结论 |
|---|---|
| 自动化测试 | 全绿（64 单测 + tsc + eslint + build） |
| Dense 面板渲染 | live / committed 两阶段都正确，pending approval 时正确退让 |
| 键盘导航 | ↓ 入栈 / ↑↓ 移动 / 边界 clamp / Enter → 详情 / ← 回 panel / 选择保持 —— 全部按设计工作 |
| 与现有 LiveAgentPanel/BackgroundTasksDialog 集成 | 兼容，无重复 UI，detail-from-panel 新模式打通了 ← back 闭环 |
| 实测体验 | 显著优于旧版「Agent × N」一行；在 9-agent /review 场景下尤其有价值 |
建议：可合并。 没有发现回归或交互坑。
本验证在本地 tmux + 真实模型对话中完成；所有捕获原始数据可在 /tmp/qwen-pr-test/ 找到。
yiliang114
approved these changes
May 26, 2026
10 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why
Commands like
/reviewfan out into 9 parallel agents, each running for minutes. The old display had two problems:Agent × 9 / <last name>line — almost zero information for a multi-minute run.ToolMessagewith very low information density — many lines per agent with mostly redundant chrome.Additionally, there was no keyboard path from the input box to the running agents — users had to know about the hidden footer pill to access the background tasks dialog.
What changes
1. Dense inline panel (
InlineParallelAgentsDisplay)For tool groups composed entirely of ≥2 parallel agent invocations, render a dense panel with one row per agent showing status / name / activity / elapsed / tokens:
Display strategy:
○with live activity from the registry on a 1s tick; completed agents show✔with fallback data fromexecutionSummary.<Static>as the permanent scrollback record.2. LiveAgentPanel keyboard navigation
▸selection indicatorBackgroundTasksDialogdirectly in detail mode (full agent view)When this kicks in
Other changes
LiveAgentPanel.DEFAULT_MAX_ROWSincreased from 5 → 12.Active agents (N/N)toActive agents (N).totalAgentCountprop toInlineParallelAgentsDisplayfor correct header when rendering a subset.livePanelFocused,livePanelSelectedIndex,enterDetailFromPaneltoBackgroundTaskViewContext.detail-from-paneldialog mode so ← returns to panel instead of list.Test plan
npx tsc --noEmitclean.中文说明
为什么
/review等命令会并行启动 9 个 agent，每个运行数分钟。旧的显示有两个问题：Agent × 9 / <last name>，几乎没有信息。ToolMessage，显示密度很低。此外，从输入框到运行中的 agent 没有键盘导航路径。
改了什么
1. 密集内联面板（
InlineParallelAgentsDisplay）纯并行 agent 组（≥2 个），每个 agent 一行，显示状态/名称/活动/耗时/token。Live + Committed 两阶段都渲染。LiveAgentPanel 在下方同时显示运行中的 agent（短暂重复，完成后自动淘汰）。
2. LiveAgentPanel 键盘导航
mainmain和 agent 行间导航，选中行显示▸触发条件
其他改动
DEFAULT_MAX_ROWS5 → 12Active agents (N)detail-from-paneldialog 模式，← 回到 panel 而非 list效果截屏