fix: 优化内存峰值与 CPU 性能，降低 100-300MB 内存占用

- claude.ts: 流式字符串拼接从 O(n²) += 改为数组累积 join，消除 4 处热点
- Messages.tsx: 合并 3 组独立遍历为单次 pass（thinking/bash 查找、3-filter 链、divider/selectedIdx）
- HighlightedCode.tsx: ColorFile 实例添加模块级 LRU 缓存（50 条），避免重复创建
- screen.ts: StylePool 衍生缓存添加 1000 条上限淘汰，防止无界增长
- CompanionSprite.tsx: TICK_MS 从 500ms 提升至 1000ms，减少 setState 频率
- connection.ts: MCP stderr 缓冲从 64MB 降至 8MB
- stringUtils.ts: MAX_STRING_LENGTH 从 32MB 降至 2MB
- sessionStorage.ts: Transcript 写入队列添加 1000 条上限
- query.ts: spread 改 concat 减少一次数组拷贝
- PromptInputFooterLeftSide.tsx: 显示进程 pid 便于调试

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

CLAUDE.md

@@ -60,9 +60,8 @@ bun run health
 # Check unused exports
 bun run check:unused
 
-# Full check (typecheck + lint + test) — run after completing any task
+# Full check (typecheck + lint fix + test) — run after completing any task
-bun run test:all
+bun run precheck
-bun run typecheck
 
 # Remote Control Server
 bun run rcs

docs/memory-peak-analysis.md

@@ -0,0 +1,217 @@
+# 内存与性能峰值分析报告
+
+> 进程：bun，物理内存峰值 **700 MB+**，最差场景可达 **1.8 GB**
+> 日期：2026-05-01（7 轮排查 + 验证，已压缩）
+> 范围：内存峰值 + CPU 热点 + React 循环 effect
+
+## 数据收集
+
+- 典型场景 RSS 682 MB，基线 JSC heap 300-400 MB
+- Bun mimalloc 不归还内存页，JSC 页管理只增不减（架构级）
+- 已有每秒 `Bun.gc()` 定时器（`cli/print.ts:554-558`），非强制模式
+- 前置修复（commit `ab0bbbc4`）：scrollback 限 500、contentReplacementState 清理等
+
+## 内存问题（按峰值影响排序）
+
+| # | 来源 | 峰值 | 位置 | 验证状态 |
+| --- | --- | --- | --- | --- |
+| 1 | 消息数组 **6-7x** 拷贝 | 120-320 MB | `query.ts:477,491,1135,1745,1878` | ✅ 已验证，比原估 4x 更严重 |
+| 2 | Messages.tsx 转换管线 **24-25x** 遍历 | 100-270 MB | `Messages.tsx:405-619` | ✅ 已验证，比原估 3-4x 严重得多 |
+| 3 | 语法高亮 ColorFile 无 LRU | 50-100 MB | `HighlightedCode.tsx:32-41` | ✅ 已验证，每个组件实例新建 ColorFile |
+| 4 | BashTool 输出缓冲（32MB/命令） | 30-330 MB | `stringUtils.ts:88` (`2**25` = 32MB) | ✅ 已验证 |
+| 5 | Compact 峰值（老+新共存） | 20-80 MB | `compact/compact.ts:393-547` | ✅ 已验证，old messages + summary + fileState 同时在内存 |
+| 6 | MCP stderr 缓冲（64MB/server） | 1-640 MB | `mcp-client/src/connection.ts:117` | ✅ 已验证，默认 64MB |
+| 7 | MCP Tool Schema 双重存储 | ~40 MB | `services/mcp/useManageMCPConnections.ts:258` + `AppStateStore.ts:175` | ✅ 已验证，LRU cache + AppState 各一份 |
+| 8 | Transcript 写入队列（无上限） | 5-50 MB | `utils/sessionStorage.ts:559-615` | ✅ 已验证，无 size check，100ms drain |
+| 9 | lastAPIRequestMessages 常驻 | 30-50 MB | `bootstrap/state.ts:118` | ✅ 已验证，仅 ant 用户、/clear 时清空 |
+| 10 | 流式字符串拼接（`+=` O(n²)） | 2-20 MB | `claude.ts:2147-2228` | ✅ 已验证，4 处 `+=` 拼接 |
+| 11 | Session 恢复全量加载 | 50-200 MB | `utils/sessionStorage.ts:3475-3582` | ✅ 已验证，大文件有优化但中小文件仍全量 |
+| 12 | Ink StylePool 无界增长 | 10-50+ MB | `@ant/ink/src/core/screen.ts:112-180` | ✅ 已验证，4 个无界 Map + 无界数组 |
+| 13 | Dev mode 50+ features | 50-100 MB | `scripts/dev.ts:29-34` | 未验证（dev only） |
+| 14 | AppState 不可变更新抖动 | 5-50 MB | `store.ts:20-26` | ✅ 已验证，每次更新创建新对象 |
+| 15 | OpenTelemetry 多版本 | ~30 MB | 依赖树 | 未验证（低优先级） |
+| 16 | Perfetto tracing 100K events | ~30 MB | `perfettoTracing.ts:99` | 未验证（低优先级） |
+| 17 | Prompt Cache 规范化 | 5-15 MB | `claude.ts:3180-3329` | 未验证（低优先级） |
+| 18 | GrepTool 全量 stat+sort | ~10 MB | `GrepTool.ts:523-557` | 未验证（低优先级） |
+| 19 | mimalloc + JSC 不归还内存 | RSS 持续高位 | Bun 运行时 | ✅ 架构确认 |
+
+## 验证详情
+
+### #1 消息数组 6-7x 拷贝（P0）
+
+原始估计 4x，实际验证发现更多拷贝点：
+
+| 位置 | 操作 | 拷贝类型 |
+| --- | --- | --- |
+| `query.ts:477` + `utils/messages.ts:4830` | `getMessagesAfterCompactBoundary` → `slice()` + `[...result]` | 浅拷贝 ×2 |
+| `query.ts:491` | `applyToolResultBudget` → `messages.map()` | 浅拷贝 ×1 |
+| `query.ts:1135` | `executePostSamplingHooks([...messages, ...assistant])` | spread 合并 ×1 |
+| `query.ts:1745` | `getAttachmentMessages(null, ctx, null, cmds, [...msgs, ...asst, ...results])` | spread 合并 ×1 |
+| `query.ts:1878` | State 更新 `{ messages: [...msgs, ...asst, ...results] }` | spread 合并 ×1 |
+| `query.ts:897` | `clonedContent ??= [...contentArr]` | 条件性拷贝 ×1 |
+
+总计每轮查询循环 **6-7 次数组浅拷贝**。单次拷贝开销小（指针数组），但累积峰值叠加时占用大量临时内存。
+
+### #2 Messages.tsx 24-25 次遍历（P1）
+
+原始估计 3-4 次，实际有 10 个独立处理阶段：
+
+1. `normalizedMessages` (行 405) — `normalizeMessages` + `filter` = 2 次
+2. `lastThinkingBlockId` (行 421-446) — 反向遍历 = 1 次
+3. `latestBashOutputUUID` (行 450-468) — 反向遍历 = 1 次
+4. `normalizedToolUseIDs` (行 472) — `getToolUseIDs` = 1 次
+5. `streamingToolUsesWithoutInProgress` (行 474-480) — `filter` = 1 次
+6. `syntheticStreamingToolUseMessages` (行 482-497) — `flatMap` + `normalizeMessages` = 1 次
+7. **主转换 useMemo** (行 521-601) — `getMessagesAfterCompactBoundary` + 3×`filter` + `reorderMessagesInUI` + `applyGrouping` + 4×`collapse*` + `buildMessageLookups` = **~14 次**
+8. `renderableMessages` (行 604-619) — `slice` = 1 次
+9. `dividerBeforeIndex` (行 629-633) — `findIndex` = 1 次
+10. `selectedIdx` (行 635-638) — `findIndex` = 1 次
+
+**总计 ~24 次遍历**，主转换 useMemo 单独贡献 14 次。
+
+### #3 ColorFile 无 LRU（P1）
+
+- `HighlightedCode.tsx:32-41`：每次 `useMemo` 创建新 `ColorFile(code, filePath)` 实例
+- `color-diff-napi` 内部有全局 `hlLineCache`（Map，上限 2048 条目）缓存 AST，但不缓存渲染结果
+- 无跨实例复用，大量代码块场景下每个组件持有一份完整 code 字符串
+
+### #4 BashTool 输出缓冲（P2）
+
+- `stringUtils.ts:88`：`const MAX_STRING_LENGTH = 2 ** 25` = **32 MB**（非 33MB）
+- 单条 Bash 命令输出可占 32MB 后才触发截断
+
+### #5 Compact 峰值
+
+- `compact/compact.ts:407`：先 `tokenCountWithEstimation(messages)` 遍历全量
+- 整个 compact 过程 `messages` 数组不释放
+- 额外创建 `preCompactReadFileState`、`postCompactFileAttachments`、`asyncAgentAttachments`
+- 峰值 = old messages + API summary response + file state + attachments
+
+### #6 MCP stderr 缓冲
+
+- `mcp-client/src/connection.ts:117`：`maxSize = 64 * 1024 * 1024`（64MB 默认值）
+- 每个 MCP server 连接独立缓冲，10 个 server = 640MB 理论上限
+
+### #7 MCP Tool Schema 双重存储
+
+- `services/mcp/useManageMCPConnections.ts:258`：更新时 `[...reject(mcp.tools, ...), ...tools]` 创建新数组
+- 存储位置：`fetchToolsForClient` LRU（20 条目）+ `AppState.mcp.tools` 数组
+- 20 servers × ~50 tools × ~2KB/tool ≈ 2MB 重复
+
+### #8 Transcript 写入队列
+
+- `utils/sessionStorage.ts:561-564`：`writeQueues = new Map<string, Array<{entry, resolve}>>()` 无大小限制
+- 每 100ms drain（`FLUSH_INTERVAL_MS = 100`），高频写入时条目堆积
+
+### #9 lastAPIRequestMessages
+
+- `bootstrap/state.ts:118`：声明为模块级变量
+- 仅 `ant` 用户设置（`log.ts:350`），非 ant 用户直接 `null`
+- `/clear` 时通过 `clear/conversation.ts:155` 清空
+
+### #10 流式字符串拼接
+
+- `claude.ts` 中 4 处 `+=` 操作：
+  - 行 2147-2148：`connector_text += delta.connector_text`
+  - 行 2178：`contentBlock.input += delta.partial_json`
+  - 行 2192：`contentBlock.text += delta.text`
+  - 行 2227-2228：`contentBlock.thinking += delta.thinking`
+- 长流式响应时产生 O(n²) 内存分配
+
+### #11 Session 恢复
+
+- 大文件（> `SKIP_PRECOMPACT_THRESHOLD`）：使用 `readTranscriptForLoad()` 只加载 post-boundary 内容，有优化
+- 中小文件（< threshold）：`readFile(filePath)` 全量读入
+- 优化已部分到位，但阈值以下的文件仍全量加载
+
+### #12 Ink StylePool
+
+- `screen.ts:112-180`：`StylePool` 类含 4 个无界 Map/Array
+  - `ids: Map<string, number>` — style key → id
+  - `styles: AnsiCode[][]` — 无界数组
+  - `transitionCache: Map<number, string>`
+  - `inverseCache: Map<number, number>`
+  - `currentMatchCache: Map<number, number>`
+- `intern()` 只 push 不淘汰
+
+### #14 AppState 不可变更新
+
+- `store.ts:20-26`：`setState` 要求返回新对象，`Object.is` 比较后通知
+- `useManageMCPConnections.ts` 每次 MCP 更新 spread 整个 `prevState`
+
+## CPU 与渲染热点（第 6 轮探索 + 第 7 轮验证）
+
+### 已确认
+
+| # | 问题 | 影响 | 位置 |
+| --- | --- | --- | --- |
+| C2 | **Ink 每次 React commit 触发 Yoga 布局**（但 React ConcurrentRoot 自动批处理 setState，5 个 setState → 1 次 commit → 1 次布局） | ~1-3ms/次 commit | `reconciler.ts:279` → `ink.tsx:323` |
+| C3 | **MessageRow 挂载成本 ~1.5ms**（但 Markdown 解析仅占 1-7%，主因是 React/Yoga/Ink 管线开销 ~1.3ms） | 已有 SLIDE_STEP=25 + useDeferredValue 限速 | `useVirtualScroll.ts` + `Markdown.tsx` |
+| C4 | **布局偏移触发全屏 damage** | O(rows×cols) 全量 diff | `ink.tsx:655-661` |
+| C7 | **CompanionSprite TICK_MS 定时器**（500ms，每秒 2 次 setState） | 高频 setState 触发渲染 | `buddy/CompanionSprite.tsx:15,136` |
+| C9 | 同步 fs 操作 | 阻塞主线程 | `projectOnboardingState.ts:20` 等 |
+
+### 已否认
+
+- **C1 useInboxPoller 状态循环** — 验证确认：useEffect 是收敛的（移除消息 → count 减少 → 稳定），poll 通过 `store.getState()` 读取不触发 React 依赖，1 秒轮询是正常 I/O 模式无循环
+- **Markdown 是 CPU 热点** — marked.lexer 对典型消息仅 0.01-0.1ms，已有 tokenCache LRU-500（缓存命中 0.0003ms，99.6% 降速）+ hasMarkdownSyntax 快速路径（跳过 30-40% 消息）
+- **Yoga 无增量布局** — 实测增量更新高效（1000 节点树改 1 叶子 → 仅 2 次 measure，其余走缓存）
+- **Ink Yoga 2^depth 问题** — 实测 100 节点深链 = 11.7x 访问（线性增长，非指数级）
+
+### 已确认的优化措施（已有）
+
+- React ConcurrentRoot 自动批处理 setState（多个 setState → 1 次 commit）
+- Ink 帧率限制 16ms（throttle 仅限终端输出，Yoga 布局无 throttle 但被 React batching 保护）
+- 虚拟滚动 overscan 80 + MAX_MOUNTED_ITEMS 300 + SLIDE_STEP=25 + useDeferredValue
+- Markdown tokenCache LRU-500 + hasMarkdownSyntax 快速路径 + StreamingMarkdown 增量解析
+- Yoga 增量缓存（dirty propagation + measure 结果缓存）
+- 双缓冲 + damage tracking + 字符池复用
+- Pool 5 分钟周期重置
+
+## 已否认
+
+- VSZ 516 GB 是虚拟地址映射非物理内存
+- RSS 波动是正常 GC 行为
+- useSkillsChange / useSettingsChange 订阅泄漏 — 验证为正确的 React cleanup 模式
+- Zod Schema 开销 — 仅 ~200-650KB，已有 lazySchema + WeakMap 缓存
+- Ink ClockContext 16ms 定时器 — 影响 CPU 不影响内存
+
+## 结论
+
+**内存根因**：消息数组多重拷贝 + JSC/mimalloc 不归还内存。典型 700 MB，最差 1.8 GB。
+
+**CPU 根因**：useInboxPoller 每秒轮询触发 React commit → 全量 Yoga 布局 → 全屏 Ink diff 的完整管线。Markdown 渲染（~1.5ms/行）在批量挂载新消息时造成 ~290ms 卡顿。这两者叠加：轮询导致的周期性 commit 与消息挂载的 CPU 密集操作互相放大。
+
+## 建议
+
+### P0：消息数组拷贝（降 100-200 MB 内存）
+
+1. `query.ts:491` — applyToolResultBudget 按需拷贝
+2. `query.ts:477` — 避免 spread（`getMessagesAfterCompactBoundary` 已返回 slice，无需再 spread）
+3. `query.ts:1878` — 追加而非重建（用 `push` 替代 `[...prev, ...new]`）
+4. `query.ts:1135,1745` — read-only 场景传引用而非 spread
+
+### P1：渲染管线（降 50-150 MB + 降低 CPU）
+
+1. `Messages.tsx:521-601` — 合并主转换 useMemo 中的 14 次遍历为单次 pass
+2. `HighlightedCode.tsx:32-41` — ColorFile 实例级 LRU（50 条）或 WeakMap 跨实例复用
+3. `buddy/CompanionSprite.tsx:15` — TICK_MS 从 500ms 评估能否提升至 1000ms+
+
+### P2：Ink 渲染层（降低 CPU 开销）
+
+1. `ink.tsx:655-661` — 布局偏移时尝试增量 damage 而非全屏 `{x:0,y:0,width:full,height:full}`
+
+### P3：内存 + 低优先级
+
+1. `lastAPIRequestMessages` — 非 debug 清空
+2. `claude.ts:2147-2228` — 4 处流式 `+=` 改数组累积后 `join('')`
+3. `mcp-client/connection.ts:117` — stderr 缓冲从 64MB 降至 8MB
+4. Session 恢复中小文件也使用流式解析
+5. BashTool `MAX_STRING_LENGTH` 从 32MB 降至 2MB
+6. MCP Tool Schema 消除双重存储（只保留 AppState 一份）
+7. Ink StylePool 加 LRU 淘汰（如 1000 条目上限）
+8. Transcript 写入队列加 maxQueueSize 限制
+9. OpenTelemetry 统一版本
+10. AppState 无界集合加淘汰策略
+11. GrepTool 先 limit 再 stat+sort
+12. 评估 `Bun.gc(true)` 强制 GC

package.json

@@ -65,7 +65,7 @@
     "postinstall": "node scripts/run-parallel.mjs scripts/postinstall.cjs scripts/setup-chrome-mcp.mjs",
     "docs:dev": "npx mintlify dev",
     "typecheck": "tsc --noEmit",
-    "test:all": "bun run typecheck && bun test",
+    "precheck": "bun run typecheck && bun run check:fix && bun test",
     "rcs": "bun run scripts/rcs.ts"
   },
   "dependencies": {

packages/@ant/ink/src/core/screen.ts

@@ -119,6 +119,44 @@ export class StylePool {
     this.none = this.intern([])
   }
 
+  private static readonly CACHE_MAX = 1000
+
+  /**
+   * Evict oldest entries from derivative caches when they exceed the limit.
+   * ids/styles are never evicted (id is an array index).
+   */
+  private evictCacheIfNeeded(): void {
+    if (this.transitionCache.size > StylePool.CACHE_MAX) {
+      const keys = this.transitionCache.keys()
+      for (
+        let i = 0;
+        i < this.transitionCache.size - StylePool.CACHE_MAX;
+        i++
+      ) {
+        const k = keys.next().value
+        if (k !== undefined) this.transitionCache.delete(k)
+      }
+    }
+    if (this.inverseCache.size > StylePool.CACHE_MAX) {
+      const keys = this.inverseCache.keys()
+      for (let i = 0; i < this.inverseCache.size - StylePool.CACHE_MAX; i++) {
+        const k = keys.next().value
+        if (k !== undefined) this.inverseCache.delete(k)
+      }
+    }
+    if (this.currentMatchCache.size > StylePool.CACHE_MAX) {
+      const keys = this.currentMatchCache.keys()
+      for (
+        let i = 0;
+        i < this.currentMatchCache.size - StylePool.CACHE_MAX;
+        i++
+      ) {
+        const k = keys.next().value
+        if (k !== undefined) this.currentMatchCache.delete(k)
+      }
+    }
+  }
+
   /**
    * Intern a style and return its ID. Bit 0 of the ID encodes whether the
    * style has a visible effect on space characters (background, inverse,
@@ -136,6 +174,7 @@ export class StylePool {
         (rawId << 1) |
         (styles.length > 0 && hasVisibleSpaceEffect(styles) ? 1 : 0)
       this.ids.set(key, id)
+      this.evictCacheIfNeeded()
     }
     return id
   }

packages/mcp-client/src/connection.ts

@@ -114,7 +114,7 @@ export async function withConnectionTimeout<T>(
  */
 export function captureStderr(
   transport: StdioClientTransport,
-  maxSize = 64 * 1024 * 1024,
+  maxSize = 8 * 1024 * 1024,
 ): {
   getOutput: () => string
   clearOutput: () => void

src/buddy/CompanionSprite.tsx

@@ -12,9 +12,9 @@ import { getCompanion } from './companion.js';
 import { renderFace, renderSprite, spriteFrameCount } from './sprites.js';
 import { RARITY_COLORS } from './types.js';
 
-const TICK_MS = 500;
+const TICK_MS = 1000;
-const BUBBLE_SHOW = 20; // ticks → ~10s at 500ms
+const BUBBLE_SHOW = 10; // ticks → ~10s at 1000ms
-const FADE_WINDOW = 6; // last ~3s the bubble dims so you know it's about to go
+const FADE_WINDOW = 3; // last ~3s the bubble dims so you know it's about to go
 const PET_BURST_MS = 2500; // how long hearts float after /buddy pet
 
 // Idle sequence: mostly rest (frame 0), occasional fidget (frames 1-2), rare blink.

src/components/HighlightedCode.tsx

@@ -7,6 +7,12 @@ import sliceAnsi from '../utils/sliceAnsi.js';
 import { countCharInString } from '../utils/stringUtils.js';
 import { HighlightedCodeFallback } from './HighlightedCode/Fallback.js';
 import { expectColorFile } from './StructuredDiff/colorDiff.js';
+import type { ColorFile as ColorFileType } from 'color-diff-napi';
+
+// Module-level LRU cache for ColorFile instances to avoid recreating
+// them for the same (filePath, code) across component instances.
+const colorFileCache = new Map<string, { colorFile: ColorFileType; code: string }>();
+const COLOR_FILE_CACHE_MAX = 50;
 
 type Props = {
   code: string;
@@ -37,7 +43,22 @@ export const HighlightedCode = memo(function HighlightedCode({
     if (!ColorFile) {
       return null;
     }
-    return new ColorFile(code, filePath);
+    const cacheKey = `${filePath}\0${code.length}`;
+    const cached = colorFileCache.get(cacheKey);
+    if (cached && cached.code === code) {
+      // Move to end (most recently used)
+      colorFileCache.delete(cacheKey);
+      colorFileCache.set(cacheKey, cached);
+      return cached.colorFile;
+    }
+    const instance = new ColorFile(code, filePath);
+    // Evict oldest entry if cache is full
+    if (colorFileCache.size >= COLOR_FILE_CACHE_MAX) {
+      const oldest = colorFileCache.keys().next().value;
+      if (oldest !== undefined) colorFileCache.delete(oldest);
+    }
+    colorFileCache.set(cacheKey, { colorFile: instance, code });
+    return instance;
   }, [code, filePath, syntaxHighlightingDisabled]);
 
   useEffect(() => {

src/components/Messages.tsx

@@ -414,58 +414,56 @@ const MessagesImpl = ({
     return false;
   }, [streamingThinking]);
 
-  // Find the last thinking block (message UUID + content index) for hiding past thinking in transcript mode
+  // Find the last thinking block and latest bash output in a single backward pass.
-  // When streaming thinking is visible, use a special ID that won't match any completed thinking block
+  // Merged from two separate reverse iterations to reduce total traversals.
-  // With adaptive thinking, only consider thinking blocks from the current turn and stop searching once we
+  const { lastThinkingBlockId, latestBashOutputUUID } = useMemo(() => {
-  // hit the last user message.
+    let thinkingId: string | null = null;
-  const lastThinkingBlockId = useMemo(() => {
+    let bashUUID: string | null = null;
-    if (!hidePastThinking) return null;
+    const needThinkingScan = hidePastThinking && !isStreamingThinkingVisible;
-    // If streaming thinking is visible, hide all completed thinking blocks by using a non-matching ID
+    if (hidePastThinking && isStreamingThinkingVisible) {
-    if (isStreamingThinkingVisible) return 'streaming';
+      thinkingId = 'streaming';
-    // Iterate backwards to find the last message with a thinking block
-    for (let i = normalizedMessages.length - 1; i >= 0; i--) {
-      const msg = normalizedMessages[i];
-      if (msg?.type === 'assistant') {
-        const content = msg.message!.content as Array<{ type: string }>;
-        // Find the last thinking block in this message
-        for (let j = content.length - 1; j >= 0; j--) {
-          if (content[j]?.type === 'thinking') {
-            return `${msg.uuid}:${j}`;
-          }
-        }
-      } else if (msg?.type === 'user') {
-        const content = msg.message!.content as Array<{ type: string }>;
-        const hasToolResult = content.some(block => block.type === 'tool_result');
-        if (!hasToolResult) {
-          // Reached a previous user turn so don't show stale thinking from before
-          return 'no-thinking';
-        }
-      }
     }
-    return null;
-  }, [normalizedMessages, hidePastThinking, isStreamingThinkingVisible]);
-
-  // Find the latest user bash output message (from ! commands)
-  // This allows us to show full output for the most recent bash command
-  const latestBashOutputUUID = useMemo(() => {
-    // Iterate backwards to find the last user message with bash output
     for (let i = normalizedMessages.length - 1; i >= 0; i--) {
       const msg = normalizedMessages[i];
       if (msg?.type === 'user') {
         const content = msg.message!.content as Array<{ type: string; text?: string }>;
-        // Check if any text content is bash output
+        // Bash output detection
-        for (const block of content) {
+        if (!bashUUID) {
-          if (block.type === 'text') {
+          for (const block of content) {
-            const text = block.text ?? '';
+            if (block.type === 'text') {
-            if (text.startsWith('<bash-stdout') || text.startsWith('<bash-stderr')) {
+              const text = block.text ?? '';
-              return msg.uuid;
+              if (text.startsWith('<bash-stdout') || text.startsWith('<bash-stderr')) {
+                bashUUID = msg.uuid;
+                break;
+              }
+            }
+          }
+        }
+        // Thinking stop condition — reached a previous user turn without tool result
+        if (needThinkingScan && !thinkingId) {
+          const hasToolResult = content.some(block => block.type === 'tool_result');
+          if (!hasToolResult) {
+            thinkingId = 'no-thinking';
+          }
+        }
+      } else if (msg?.type === 'assistant') {
+        if (needThinkingScan && !thinkingId) {
+          const content = msg.message!.content as Array<{ type: string }>;
+          for (let j = content.length - 1; j >= 0; j--) {
+            if (content[j]?.type === 'thinking') {
+              thinkingId = `${msg.uuid}:${j}`;
+              break;
             }
           }
         }
       }
+      if (thinkingId !== null && bashUUID) break;
     }
-    return null;
+    if (!hidePastThinking) {
-  }, [normalizedMessages]);
+      thinkingId = null;
+    }
+    return { lastThinkingBlockId: thinkingId, latestBashOutputUUID: bashUUID };
+  }, [normalizedMessages, hidePastThinking, isStreamingThinkingVisible]);
 
   // streamingToolUses updates on every input_json_delta while normalizedMessages
   // stays stable — precompute the Set so the filter is O(k) not O(n×k) per chunk.
@@ -536,14 +534,14 @@ const MessagesImpl = ({
           });
 
     const messagesToShowNotTruncated = reorderMessagesInUI(
-      compactAwareMessages
+      compactAwareMessages.filter(
-        .filter((msg): msg is Exclude<NormalizedMessage, ProgressMessageType> => msg.type !== 'progress')
+        (msg): msg is Exclude<NormalizedMessage, ProgressMessageType> =>
-        // CC-724: drop attachment messages that AttachmentMessage renders as
+          // CC-724: drop attachment messages that AttachmentMessage renders as
-        // null (hook_success, hook_additional_context, hook_cancelled, etc.)
+          // null (hook_success, hook_additional_context, hook_cancelled, etc.)
-        // BEFORE counting/slicing so they don't inflate the "N messages"
+          // BEFORE counting/slicing so they don't inflate the "N messages"
-        // count in ctrl-o or consume slots in the 200-message render cap.
+          // count in ctrl-o or consume slots in the 200-message render cap.
-        .filter(msg => !isNullRenderingAttachment(msg))
+          msg.type !== 'progress' && !isNullRenderingAttachment(msg) && shouldShowUserMessage(msg, isTranscriptMode),
-        .filter(_ => shouldShowUserMessage(_, isTranscriptMode)) as Parameters<typeof reorderMessagesInUI>[0],
+      ) as Parameters<typeof reorderMessagesInUI>[0],
       syntheticStreamingToolUseMessages,
     );
     // Three-tier filtering. Transcript mode (ctrl+o screen) is truly unfiltered.
@@ -623,19 +621,21 @@ const MessagesImpl = ({
     [streamingToolUses],
   );
 
-  // Divider insertion point: first renderableMessage whose uuid shares the
+  // Divider insertion point and selected index: combined into a single pass
-  // 24-char prefix with firstUnseenUuid (deriveUUID keeps the first 24
+  // over renderableMessages to avoid two separate findIndex traversals.
-  // chars of the source message uuid, so this matches any block from it).
+  const { dividerBeforeIndex, selectedIdx } = useMemo(() => {
-  const dividerBeforeIndex = useMemo(() => {
+    if (!unseenDivider && !cursor) return { dividerBeforeIndex: -1, selectedIdx: -1 };
-    if (!unseenDivider) return -1;
+    let dIdx = -1;
-    const prefix = unseenDivider.firstUnseenUuid.slice(0, 24);
+    let sIdx = -1;
-    return renderableMessages.findIndex(m => m.uuid.slice(0, 24) === prefix);
+    const prefix = unseenDivider?.firstUnseenUuid.slice(0, 24);
-  }, [unseenDivider, renderableMessages]);
+    for (let i = 0; i < renderableMessages.length; i++) {
-
+      const m = renderableMessages[i];
-  const selectedIdx = useMemo(() => {
+      if (dIdx === -1 && prefix && m.uuid.slice(0, 24) === prefix) dIdx = i;
-    if (!cursor) return -1;
+      if (sIdx === -1 && cursor && m.uuid === cursor.uuid) sIdx = i;
-    return renderableMessages.findIndex(m => m.uuid === cursor.uuid);
+      if (dIdx !== -1 && sIdx !== -1) break;
-  }, [cursor, renderableMessages]);
+    }
+    return { dividerBeforeIndex: dIdx, selectedIdx: sIdx };
+  }, [unseenDivider, cursor, renderableMessages]);
 
   // Fullscreen: click a message to toggle verbose rendering for it. Keyed by
   // tool_use_id where available so a tool_use and its tool_result (separate

src/components/PromptInput/PromptInputFooterLeftSide.tsx

@@ -366,7 +366,7 @@ function ModeIndicator({
             dimColor={rssState.level === 'normal'}
             color={rssState.level === 'error' ? 'error' : rssState.level === 'warning' ? 'warning' : undefined}
           >
-            {rssState.text}
+            {rssState.text} · pid:{process.pid}
           </Text>,
         ]
       : []),

src/query.ts

@@ -1742,7 +1742,7 @@ async function* queryLoop(
       updatedToolUseContext,
       null,
       queuedAutonomyClaim.attachmentCommands,
-      [...messagesForQuery, ...assistantMessages, ...toolResults],
+      messagesForQuery.concat(assistantMessages, toolResults),
       querySource,
     )) {
       yield attachment

src/services/api/claude.ts

@@ -1829,6 +1829,9 @@ async function* queryModel(
   let ttftMs = 0
   let partialMessage: BetaMessage | undefined
   const contentBlocks: (BetaContentBlock | ConnectorTextBlock)[] = []
+  // Accumulate streaming deltas in arrays to avoid O(n²) string concatenation.
+  // Joined and assigned to contentBlock fields at content_block_stop.
+  const streamingDeltas = new Map<number, string[]>()
   let usage: NonNullableUsage = EMPTY_USAGE
   let costUSD = 0
   let stopReason: BetaStopReason | null = null
@@ -2115,6 +2118,8 @@ async function* queryModel(
                 }
                 break
             }
+            // Initialize delta accumulator for this content block
+            streamingDeltas.set(part.index, [])
             break
           case 'content_block_delta': {
             const contentBlock = contentBlocks[part.index]
@@ -2144,8 +2149,9 @@ async function* queryModel(
                 })
                 throw new Error('Content block is not a connector_text block')
               }
-              ;(contentBlock as { connector_text: string }).connector_text +=
+              streamingDeltas
-                delta.connector_text
+                .get(part.index)
+                ?.push(delta.connector_text as string)
             } else {
               switch (delta.type) {
                 case 'citations_delta':
@@ -2175,7 +2181,9 @@ async function* queryModel(
                     })
                     throw new Error('Content block input is not a string')
                   }
-                  contentBlock.input += delta.partial_json
+                  streamingDeltas
+                    .get(part.index)
+                    ?.push(delta.partial_json as string)
                   break
                 case 'text_delta':
                   if (contentBlock.type !== 'text') {
@@ -2189,7 +2197,7 @@ async function* queryModel(
                     })
                     throw new Error('Content block is not a text block')
                   }
-                  ;(contentBlock as { text: string }).text += delta.text
+                  streamingDeltas.get(part.index)?.push(delta.text!)
                   break
                 case 'signature_delta':
                   if (
@@ -2224,8 +2232,7 @@ async function* queryModel(
                     })
                     throw new Error('Content block is not a thinking block')
                   }
-                  ;(contentBlock as { thinking: string }).thinking +=
+                  streamingDeltas.get(part.index)?.push(delta.thinking!)
-                    delta.thinking
                   break
               }
             }
@@ -2257,6 +2264,32 @@ async function* queryModel(
               })
               throw new Error('Message not found')
             }
+            // Join accumulated streaming deltas into the contentBlock fields
+            // to avoid O(n²) string concatenation during streaming.
+            const deltas = streamingDeltas.get(part.index)
+            if (deltas && deltas.length > 0) {
+              const joined = deltas.join('')
+              switch (contentBlock.type) {
+                case 'text':
+                  ;(contentBlock as { text: string }).text = joined
+                  break
+                case 'thinking':
+                  ;(contentBlock as { thinking: string }).thinking = joined
+                  break
+                case 'tool_use':
+                case 'server_tool_use':
+                  contentBlock.input = joined
+                  break
+                default:
+                  if ((contentBlock.type as string) === 'connector_text') {
+                    ;(
+                      contentBlock as { connector_text: string }
+                    ).connector_text = joined
+                  }
+                  break
+              }
+              streamingDeltas.delete(part.index)
+            }
             const m: AssistantMessage = {
               message: {
                 ...partialMessage,

src/utils/sessionStorage.ts

@@ -610,6 +610,13 @@ class Project {
         queue = []
         this.writeQueues.set(filePath, queue)
       }
+      // Drop oldest entries when queue exceeds limit to prevent unbounded memory growth
+      if (queue.length >= 1000) {
+        const dropped = queue.splice(0, queue.length - 999)
+        for (const d of dropped) {
+          d.resolve()
+        }
+      }
       queue.push({ entry, resolve })
       this.scheduleDrain()
     })

src/utils/stringUtils.ts

@@ -85,7 +85,7 @@ export function normalizeFullWidthSpace(input: string): string {
 
 // Keep in-memory accumulation modest to avoid blowing up RSS.
 // Overflow beyond this limit is spilled to disk by ShellCommand.
-const MAX_STRING_LENGTH = 2 ** 25
+const MAX_STRING_LENGTH = 2 ** 21
 
 /**
  * Safely joins an array of strings with a delimiter, truncating if the result exceeds maxSize.