fix: third-party API user_id validation error (DeepSeek, etc.)

When ANTHROPIC_BASE_URL points to a non-Anthropic endpoint (e.g.
DeepSeek), the JSON-formatted user_id containing {, ", : characters
fails validation against ^[a-zA-Z0-9_-]+$. Send only the hex device_id
for third-party providers.

docs/features/background-agent-selector.md

@@ -0,0 +1,225 @@
+# Background Agent Selector — 底部统一后台 Agent 切换器
+
+> Feature Flag: 无（直接启用）
+> 实现状态：完整可用
+> 依赖：`viewingAgentTaskId` / `enterTeammateView` / `exitTeammateView` 已有机制
+
+## 一、功能概述
+
+Background Agent Selector 是渲染在 PromptInput 下方的常驻状态条，列出当前所有 **backgrounded 的 local_agent 任务**（包括 `/fork` 派生的 fork agent 和 Task/AgentTool 调用 `run_in_background: true` 派生的子 agent）。用户可以用 ↑/↓ 方向键在 `main` 和各 agent 之间切换焦点，按 Enter 把 REPL 主视图替换为所选 agent 的实时 transcript，再按 Enter 选中 `main` 即可回到主对话。
+
+整个机制完全复用官方已有的 teammate transcript 查看基础设施，不引入新的视图层 / 数据流，仅新增一条 footer pill 类型。
+
+### 核心特性
+
+- **统一入口**：`/fork`、Task 派生的 subagent、所有 `run_in_background: true` 的 agent 都在同一栏显示
+- **就地切换**：prompt 为空时按 ↓ 溢出进入底部 selector，↑↓ 选中某行，Enter 即切主视图
+- **实时状态**：每行显示 agent 类型 + 描述 + 运行时长 + 已消耗 token；running 时圆点为绿色
+- **Keep-alive 视图**：agent 完成后在 `evictAfter` grace 窗口内保留一段时间，用户可回看
+- **零界面侵入**：tasks 数为 0 时 selector 完全不渲染，不占屏幕高度
+- **与旧 Dialog 共存**：Shift+↓ 打开的 `BackgroundTasksDialog` 原有行为保留，selector 只作为展示 + 快捷切换
+
+## 二、用户交互
+
+### 触发方式
+
+有任何 background agent 时，selector 自动出现在 `bypass permissions on` 行下方：
+
+```
+  claude-code | Opus 4.7 (1M context) | ctx:4%
+  ▶▶ bypass permissions on (shift+tab to cycle)
+
+  ○ main                                    ↑/↓ to select · Enter to view
+  ● Explore  Research src/hooks              23s · ↓ 10.9k tokens
+  ○ Explore  Research src/components         22s · ↓  9.5k tokens
+  ○ Explore  Research src/utils              21s · ↓ 13.6k tokens
+```
+
+### 键盘路由
+
+| 位置 / 状态 | 按键 | 行为 |
+|---|---|---|
+| PromptInput 非空 | ↑↓ | 光标移动 / 翻历史（不变） |
+| PromptInput 空 + 历史底部 | ↓ | 焦点下放到 selector，高亮到 `● main` |
+| Selector 聚焦（`footerSelection === 'bg_agent'`） | ↓ | 高亮下移，-1 → 0 → ... → N-1 |
+| Selector 聚焦 | ↑ | 高亮上移；在 `main` 再 ↑ → 焦点回 PromptInput |
+| Selector 聚焦 | Enter | `-1` → `exitTeammateView`；`>=0` → `enterTeammateView(agentId)`。焦点保留在 pill |
+| Selector 聚焦 | Esc | `footer:clearSelection`，焦点回 PromptInput |
+
+### 视觉规则
+
+- `● main` / `● <agent>`：当前被**查看**（viewingAgentTaskId 指向）或被**光标聚焦**（pill focused 时以光标为准）的一行
+- running 状态的 agent：圆点渲染为 `success` 色（绿色），与 `BackgroundTasksDialog` 状态语义对齐
+- 右上角 hint 随状态变化：
+  - pill 聚焦：`↑/↓ to select · Enter to view`
+  - 已选中 running agent：`shift+↓ to manage · x to stop`
+  - 已选中 terminal agent：`shift+↓ to manage · x to clear`
+  - 未选中任何 agent：`shift+↓ to manage background agents`
+
+## 三、实现架构
+
+### 3.1 数据层：`useBackgroundAgentTasks`
+
+文件：`src/hooks/useBackgroundAgentTasks.ts`
+
+封装对 `useAppState(s => s.tasks)` 的过滤：
+
+```ts
+export function useBackgroundAgentTasks(): LocalAgentTaskState[] {
+  const tasks = useAppState(s => s.tasks)
+  return useMemo(() => {
+    const now = Date.now()
+    return Object.values(tasks)
+      .filter(isLocalAgentTask)
+      .filter(t => t.agentType !== 'main-session')
+      .filter(t => t.isBackgrounded !== false)
+      .filter(t => t.evictAfter === undefined || t.evictAfter > now)
+      .sort((a, b) => a.startTime - b.startTime)
+  }, [tasks])
+}
+```
+
+`/fork` 和 `AgentTool` 的 `run_in_background: true` 底层都走 `registerAsyncAgent → runAsyncAgentLifecycle`，最终写入同一个 `appState.tasks` Map；此 hook 是唯一数据源，Selector 和 PromptInput 的 `bgAgentList` 都消费它。
+
+### 3.2 状态层：新增两个字段
+
+文件：`src/state/AppStateStore.ts`
+
+```ts
+export type FooterItem =
+  | 'tasks' | 'tmux' | 'bagel' | 'teams' | 'bridge' | 'companion'
+  | 'bg_agent'   // ← 新增
+
+export type AppState = DeepImmutable<{
+  // ...
+  selectedBgAgentIndex: number  // -1 = main, 0..N-1 = 选中的 agent
+}>
+```
+
+- `'bg_agent'` 作为 `FooterItem` 加入 footer pill 体系，享受既有的 `footer:up` / `footer:down` / `footer:openSelected` keybinding 路由
+- `selectedBgAgentIndex` 记录 selector 的光标位置，与 `viewingAgentTaskId`（"正在看什么"）独立；它不可从 `viewingAgentTaskId` 派生——Enter 后光标留在 pill 继续导航，查看目标才变
+
+### 3.3 键盘路由：PromptInput footer pill 分支
+
+文件：`src/components/PromptInput/PromptInput.tsx`
+
+1. **`bg_agent` 进入 footerItems[0]**：保证 prompt ↓ 溢出时（`handleHistoryDown` → `selectFooterItem(footerItems[0])`）直接进入 selector，而不是 `tasks` 等其他 pill
+2. **`footer:up` 分支**：`bgAgentSelected` 时 `selectedBgAgentIndex > -1` 则递减；在 -1 → `selectFooterItem(null)` 退出 pill
+3. **`footer:down` 分支**：`selectedBgAgentIndex < bgAgentList.length - 1` 则递增，到底 clamp
+4. **`footer:openSelected` 分支**：index === -1 → `exitTeammateView`；否则 `enterTeammateView(bgAgentList[i].agentId)`。**不清理 pill 焦点**，光标留在 selector 上继续导航
+5. **`selectFooterItem('bg_agent')`**：入 pill 时重置 `selectedBgAgentIndex = -1`（光标落到 `main`）
+
+### 3.4 渲染层：`BackgroundAgentSelector`
+
+文件：`src/components/tasks/BackgroundAgentSelector.tsx`
+
+纯展示组件，不订阅键盘：
+
+```tsx
+const tasks = useBackgroundAgentTasks()
+const viewingId = useAppState(s => s.viewingAgentTaskId)
+const footerSelection = useAppState(s => s.footerSelection)
+const selectedBgIndex = useAppState(s => s.selectedBgAgentIndex)
+
+if (tasks.length === 0) return null
+
+const pillFocused = footerSelection === 'bg_agent'
+const highlightedId = pillFocused
+  ? (selectedBgIndex === -1 ? null : tasks[selectedBgIndex]?.agentId ?? null)
+  : (viewingId ?? null)
+```
+
+**高亮派生规则**：pill 聚焦 → 跟 `selectedBgAgentIndex`；未聚焦 → 镜像 `viewingAgentTaskId`。这样当用户通过 Shift+↓ Dialog 或 `enterTeammateView` 其它途径切换视图时，selector 也会正确反映。
+
+### 3.5 主视图切换：复用 `viewingAgentTaskId`
+
+REPL.tsx 主体仍复用原有查看逻辑：
+
+```ts
+const viewedTask = viewingAgentTaskId ? tasks[viewingAgentTaskId] : undefined
+const viewedAgentTask = ... (isLocalAgentTask(viewedTask) ? viewedTask : undefined)
+const displayedMessages = viewedAgentTask ? displayedAgentMessages : messages
+```
+
+当 `enterTeammateView(agentId)` 把 `viewingAgentTaskId` 设成某个 local_agent 的 id：
+
+- `viewedAgentTask` 解析成该 agent
+- `displayedMessages` 切换到 agent 的 messages
+- 消息列表、spinner、unseen divider 等一整套组件自动用 agent transcript 重渲染
+- 主对话流被"暂停"（并非销毁，回到 `main` 时仍在原处）
+
+`enterTeammateView` 同步负责：设 `retain: true` 阻止 eviction、清 `evictAfter`、触发 disk bootstrap 从 `agent-<id>.jsonl` 加载完整 transcript 到 `task.messages`。
+
+#### Fork agent prompt 归一化
+
+`/fork` agent 的 transcript 和普通 subagent 不同：它继承 main agent 的上下文，真实初始消息形态是：
+
+```text
+...parent messages
+assistant([...tool_use])
+user([tool_result..., text("<fork-boilerplate>...Your directive: <prompt>")])
+...fork live messages
+```
+
+这里的 prompt 文本混在 `[tool_result..., text]` 多 block user message 里。消息渲染管线会优先把这条 user message 当作 tool-result plumbing 来处理，导致 `<fork-boilerplate>` 里的用户 prompt 不稳定可见。为保证切换到 fork agent 时总能看到用户发起的 fork prompt，REPL.tsx 对 fork 视图做一次展示层归一化：
+
+1. 仅当 `viewedAgentTask.agentType === 'fork'` 时启用，不影响普通 Explore / Task subagent。
+2. 从原始 messages 中识别包含 `<fork-boilerplate>` 的 carrier message。
+3. 剥离 carrier message 里的 boilerplate text block，但保留 `tool_result` blocks，避免破坏父 assistant `tool_use` 的承接关系。
+4. 强制插入一条独立 `createUserMessage({ content: viewedAgentTask.prompt })` 作为可见用户 prompt。
+5. 插入位置优先为 boilerplate carrier 后；如果 sidechain bootstrap 还没读到 carrier，则插到最后一条 inherited `assistant tool_use` 后面，确保 prompt 接在 main 上下文之后，而不是跑到视图顶部。
+
+这个归一化只影响 UI 展示用的 `displayedAgentMessages`，不回写 `task.messages`，也不改变发送给模型的 fork transcript。
+
+### 3.6 生命周期
+
+完全复用官方既有机制：
+
+- **运行中**：`isBackgroundTask()` 谓词为真，selector 列出
+- **完成 / 失败 / 中止**：`completeAgentTask` / `failAgentTask` / `killAsyncAgent` 设 `status` 为 terminal
+- **回访后退出**：`exitTeammateView` 调 `release(task)`——清 `retain`、清 `messages`、terminal 状态下设 `evictAfter = now + PANEL_GRACE_MS (30s)`
+- **evictAfter 过期**：`useBackgroundAgentTasks` 过滤时自然剔除，selector 行消失
+- **手动清除**：`stopOrDismissAgent(taskId)` 设 `evictAfter = 0`，立即消失
+
+## 四、设计决策
+
+1. **数据源单一**：`useBackgroundAgentTasks` 是唯一过滤点，PromptInput 也复用，避免过滤条件散落
+2. **pill 聚焦保留**：Enter 切视图后不松焦，让 ↑↓ 连续导航，贴近官方体验
+3. **`bg_agent` 放 footerItems[0]**：确保 ↓ 溢出直接进入 selector 而非其它 pill
+4. **selector 不订阅键盘**：所有按键路由集中在 PromptInput 的 `footer:*` 分支，避免 selector 组件和 PromptInput 双重 `useInput` 的冲突
+5. **`selectedBgAgentIndex` 存 AppState 而非局部 state**：selector 和 PromptInput 分别在两棵不同子树，需要全局字段协调；该值不能从 `viewingAgentTaskId` 派生
+6. **与 `BackgroundTasksDialog` 共存**：Shift+↓ 行为完全不变，selector 是补充快捷入口；Dialog 仍管 shell / workflow / monitor_mcp 等 selector 不显示的 task 类型
+7. **fork prompt 展示层兜底**：fork prompt 不依赖 boilerplate 自身渲染，统一在 `displayedAgentMessages` 中合成独立用户消息；普通 subagent 不走该分支，避免 prompt 重复
+
+## 五、关键 API 复用
+
+| 官方已有能力 | selector 如何使用 |
+|---|---|
+| `AppState.tasks` | 单一数据源，无需 file watcher / output JSONL 订阅 |
+| `registerAsyncAgent` | `/fork` 和 AgentTool 共用，selector 不区分来源 |
+| `enterTeammateView(id)` | Enter 时调用，负责 retain + disk bootstrap |
+| `exitTeammateView` | Enter 选中 `main` 时调用 |
+| `release(task)` + `PANEL_GRACE_MS` | 30s keep-alive，selector 自动生效 |
+| `useElapsedTime` | 每行时长显示，非 running 自动停 interval |
+| `formatTokens` (`utils/format.ts`) | token 数 1k 缩写 |
+| `footer:up` / `footer:down` / `footer:openSelected` keybinding | 键盘路由复用 Footer context |
+
+## 六、文件索引
+
+| 文件 | 职责 |
+|------|------|
+| `src/hooks/useBackgroundAgentTasks.ts` | 数据过滤 hook（backgrounded local_agent + evictAfter 过滤 + startTime 排序） |
+| `src/components/tasks/BackgroundAgentSelector.tsx` | 底部 selector UI，纯展示 |
+| `src/components/PromptInput/PromptInput.tsx` | 新增 `'bg_agent'` footer pill + 对应的 `footer:up/down/openSelected` 分支 |
+| `src/state/AppStateStore.ts` | `FooterItem` 加 `'bg_agent'`；新增 `selectedBgAgentIndex` 字段 |
+| `src/main.tsx` | `getDefaultAppState` 同步初始化 `selectedBgAgentIndex: -1` |
+| `src/screens/REPL.tsx` | 在 PromptInput + SessionBackgroundHint 之后挂载 `<BackgroundAgentSelector />`；切换 agent 主视图；对 fork transcript 做 prompt 归一化 |
+| `src/components/messages/AssistantToolUseMessage.tsx` | 新增 `defaultCollapsed?: boolean` prop，为后续详情视图默认折叠工具块预留 |
+| `src/components/messages/UserTextMessage.tsx` | 识别 `<fork-boilerplate>`，交给 fork 专用 renderer 处理 |
+| `src/components/messages/UserForkBoilerplateMessage.tsx` | 将 fork boilerplate text 折叠为纯用户 prompt；作为 transcript 中原位渲染的兼容路径 |
+
+## 七、已知限制
+
+- `Date.now()` 在 `useBackgroundAgentTasks` 的 useMemo 里冻结于 `[tasks]` 触发时：若长时间没有新 task 变更事件，某个 terminal agent 的 grace 期过期后不会立即从 selector 消失，要等下一次 tasks 变化才刷新。在典型使用（主对话一直在产生消息）下感知不到，暂不额外加 interval。
+- Selector 当前不处理 Shell Task / Workflow / Monitor MCP 等类型——这些仍走 `BackgroundTasksDialog`（Shift+↓）管理。
+- `AssistantToolUseMessage` 的 `defaultCollapsed` prop 目前无调用方传值，保留作为后续"agent 详情视图内工具块默认折叠"扩展点。

src/commands/fork/fork.tsx

@@ -43,8 +43,12 @@ export async function call(
     // Omitting subagent_type triggers implicit fork.
     const input = {
       prompt: directive,
+      fork: true, // 触发 AgentTool 的 fork 路径：继承父会话上下文 + system prompt + 模型
       run_in_background: true, // fork always runs async
-      description: `Fork: ${directive.slice(0, 30)}${directive.length > 30 ? '...' : ''}`,
+      // description 只显示在底部 selector / BackgroundTasksDialog，保持简短标签
+      // 即可；用户输入的 prompt 会作为第一条用户消息呈现在主视图里，这里不要
+      // 重复显示。
+      description: 'forked from main',
     };
 
     // Call AgentTool with proper parameters:

src/components/PromptInput/PromptInput.tsx

@@ -26,6 +26,7 @@ import { useSetPromptOverlayDialog } from '../../context/promptOverlayContext.js
 import { formatImageRef, formatPastedTextRef, getPastedTextRefNumLines, parseReferences } from '../../history.js';
 import type { VerificationStatus } from '../../hooks/useApiKeyVerification.js';
 import { type HistoryMode, useArrowKeyHistory } from '../../hooks/useArrowKeyHistory.js';
+import { useBackgroundAgentTasks } from '../../hooks/useBackgroundAgentTasks.js';
 import { useDoublePress } from '../../hooks/useDoublePress.js';
 import { useHistorySearch } from '../../hooks/useHistorySearch.js';
 import type { IDESelection } from '../../hooks/useIdeSelection.js';
@@ -415,6 +416,16 @@ function PromptInput({
   // First ↓ selects the pill, second ↓ moves to row 0. Prevents double-select
   // of pill + row when both bg tasks (pill) and forked agents (rows) are visible.
   const coordinatorTaskIndex = useAppState(s => s.coordinatorTaskIndex);
+  const selectedBgAgentIndex = useAppState(s => s.selectedBgAgentIndex);
+  const setSelectedBgAgentIndex = useCallback(
+    (v: number | ((prev: number) => number)) =>
+      setAppState(prev => {
+        const next = typeof v === 'function' ? v(prev.selectedBgAgentIndex) : v;
+        if (next === prev.selectedBgAgentIndex) return prev;
+        return { ...prev, selectedBgAgentIndex: next };
+      }),
+    [setAppState],
+  );
   const setCoordinatorTaskIndex = useCallback(
     (v: number | ((prev: number) => number)) =>
       setAppState(prev => {
@@ -501,10 +512,13 @@ function PromptInput({
     (runningTaskCount > 0 || (process.env.USER_TYPE === 'ant' && coordinatorTaskCount > 0)) &&
     !shouldHideTasksFooter(tasks, showSpinnerTree);
   const teamsFooterVisible = cachedTeams.length > 0;
+  const bgAgentList = useBackgroundAgentTasks();
+  const bgAgentFooterVisible = bgAgentList.length > 0;
 
   const footerItems = useMemo(
     () =>
       [
+        bgAgentFooterVisible && 'bg_agent',
         tasksFooterVisible && 'tasks',
         tmuxFooterVisible && 'tmux',
         bagelFooterVisible && 'bagel',
@@ -513,6 +527,7 @@ function PromptInput({
         companionFooterVisible && 'companion',
       ].filter(Boolean) as FooterItem[],
     [
+      bgAgentFooterVisible,
       tasksFooterVisible,
       tmuxFooterVisible,
       bagelFooterVisible,
@@ -540,6 +555,7 @@ function PromptInput({
   const _bagelSelected = footerItemSelected === 'bagel';
   const teamsSelected = footerItemSelected === 'teams';
   const bridgeSelected = footerItemSelected === 'bridge';
+  const bgAgentSelected = footerItemSelected === 'bg_agent';
 
   function selectFooterItem(item: FooterItem | null): void {
     setAppState(prev => (prev.footerSelection === item ? prev : { ...prev, footerSelection: item }));
@@ -547,6 +563,9 @@ function PromptInput({
       setTeammateFooterIndex(0);
       setCoordinatorTaskIndex(minCoordinatorIndex);
     }
+    if (item === 'bg_agent') {
+      setSelectedBgAgentIndex(-1);
+    }
   }
 
   // delta: +1 = down/right, -1 = up/left. Returns true if nav happened
@@ -1808,6 +1827,15 @@ function PromptInput({
   useKeybindings(
     {
       'footer:up': () => {
+        // ↑ in bg_agent pill: move selection up (-1 = main). At -1, leave pill.
+        if (bgAgentSelected) {
+          if (selectedBgAgentIndex > -1) {
+            setSelectedBgAgentIndex(prev => prev - 1);
+          } else {
+            selectFooterItem(null);
+          }
+          return;
+        }
         // ↑ scrolls within the coordinator task list before leaving the pill
         if (
           tasksSelected &&
@@ -1821,6 +1849,13 @@ function PromptInput({
         navigateFooter(-1, true);
       },
       'footer:down': () => {
+        // ↓ in bg_agent pill: move selection down through agents. Clamp at last.
+        if (bgAgentSelected) {
+          if (selectedBgAgentIndex < bgAgentList.length - 1) {
+            setSelectedBgAgentIndex(prev => prev + 1);
+          }
+          return;
+        }
         // ↓ scrolls within the coordinator task list, never leaves the pill
         if (tasksSelected && process.env.USER_TYPE === 'ant' && coordinatorTaskCount > 0) {
           if (coordinatorTaskIndex < coordinatorTaskCount - 1) {
@@ -1906,6 +1941,15 @@ function PromptInput({
             setShowBridgeDialog(true);
             selectFooterItem(null);
             break;
+          case 'bg_agent':
+            if (selectedBgAgentIndex === -1) {
+              exitTeammateView(setAppState);
+            } else {
+              const picked = bgAgentList[selectedBgAgentIndex];
+              if (picked) enterTeammateView(picked.agentId, setAppState);
+            }
+            // Keep the pill focused so ↑/↓ continue to work after Enter.
+            break;
         }
       },
       'footer:clearSelection': () => {

src/components/messages/AssistantToolUseMessage.tsx

@@ -30,6 +30,7 @@ type Props = {
   inProgressToolCallCount?: number;
   lookups: ReturnType<typeof buildMessageLookups>;
   isTranscriptMode?: boolean;
+  defaultCollapsed?: boolean;
 };
 
 export function AssistantToolUseMessage({
@@ -45,6 +46,7 @@ export function AssistantToolUseMessage({
   inProgressToolCallCount,
   lookups,
   isTranscriptMode,
+  defaultCollapsed,
 }: Props): React.ReactNode {
   const terminalSize = useTerminalSize();
   const [theme] = useTheme();
@@ -167,6 +169,7 @@ export function AssistantToolUseMessage({
         </Box>
         {!isResolved &&
           !isQueued &&
+          !defaultCollapsed &&
           (isClassifierChecking ? (
             <MessageResponse height={1}>
               <Text dimColor>

src/components/messages/UserForkBoilerplateMessage.tsx

@@ -3,28 +3,31 @@
  */
 import type { TextBlockParam } from '@anthropic-ai/sdk/resources/index.mjs';
 import * as React from 'react';
-import { Box, Text } from '@anthropic/ink';
+import { FORK_BOILERPLATE_TAG, FORK_DIRECTIVE_PREFIX } from '../../constants/xml.js';
 import { extractTag } from '../../utils/messages.js';
+import { UserPromptMessage } from './UserPromptMessage.js';
 
 type Props = {
   addMargin: boolean;
   param: TextBlockParam;
+  isTranscriptMode?: boolean;
+  timestamp?: string;
 };
 
-export function UserForkBoilerplateMessage({ param, addMargin }: Props): React.ReactNode {
-  const text = param.text;
-  const extracted = extractTag(text, 'fork-boilerplate');
-  if (!extracted) {
-    return null;
-  }
-
-  const firstLine = extracted.trim().split('\n')[0] ?? '';
-  const preview = firstLine.length > 80 ? firstLine.slice(0, 77) + '...' : firstLine;
+export function UserForkBoilerplateMessage({ param, addMargin, isTranscriptMode, timestamp }: Props): React.ReactNode {
+  if (!extractTag(param.text, FORK_BOILERPLATE_TAG)) return null;
+  const closeTag = `</${FORK_BOILERPLATE_TAG}>`;
+  const afterTag = param.text.slice(param.text.indexOf(closeTag) + closeTag.length).trimStart();
+  const userPrompt = afterTag.startsWith(FORK_DIRECTIVE_PREFIX)
+    ? afterTag.slice(FORK_DIRECTIVE_PREFIX.length)
+    : afterTag;
 
   return (
-    <Box flexDirection="row" marginTop={addMargin ? 1 : 0}>
-      <Text dimColor>[fork] </Text>
-      <Text>{preview}</Text>
-    </Box>
+    <UserPromptMessage
+      addMargin={addMargin}
+      param={{ type: 'text', text: userPrompt }}
+      isTranscriptMode={isTranscriptMode}
+      timestamp={timestamp}
+    />
   );
 }

src/components/messages/UserTextMessage.tsx

@@ -4,6 +4,7 @@ import * as React from 'react';
 import { NO_CONTENT_MESSAGE } from '../../constants/messages.js';
 import {
   COMMAND_MESSAGE_TAG,
+  FORK_BOILERPLATE_TAG,
   LOCAL_COMMAND_CAVEAT_TAG,
   TASK_NOTIFICATION_TAG,
   TEAMMATE_MESSAGE_TAG,
@@ -124,16 +125,21 @@ export function UserTextMessage({
   }
 
   // Fork child's first message: collapse the rules/format boilerplate, show
-  // only the directive. FORK_BOILERPLATE_TAG is inlined so the import doesn't
-  // ship in external builds where feature('FORK_SUBAGENT') is false.
-  if (feature('FORK_SUBAGENT')) {
-    if (param.text.includes('<fork-boilerplate>')) {
-      /* eslint-disable @typescript-eslint/no-require-imports */
-      const { UserForkBoilerplateMessage } =
-        require('./UserForkBoilerplateMessage.js') as typeof import('./UserForkBoilerplateMessage.js');
-      /* eslint-enable @typescript-eslint/no-require-imports */
-      return <UserForkBoilerplateMessage addMargin={addMargin} param={param} />;
-    }
+  // only the user prompt. Independent of FORK_SUBAGENT flag — the fork agent
+  // transcript always needs to render the prompt as a normal user bubble.
+  if (param.text.includes(`<${FORK_BOILERPLATE_TAG}>`)) {
+    /* eslint-disable @typescript-eslint/no-require-imports */
+    const { UserForkBoilerplateMessage } =
+      require('./UserForkBoilerplateMessage.js') as typeof import('./UserForkBoilerplateMessage.js');
+    /* eslint-enable @typescript-eslint/no-require-imports */
+    return (
+      <UserForkBoilerplateMessage
+        addMargin={addMargin}
+        param={param}
+        isTranscriptMode={isTranscriptMode}
+        timestamp={timestamp}
+      />
+    );
   }
 
   // Cross-session UDS message (from another Claude session's SendMessage).

src/components/tasks/BackgroundAgentSelector.tsx

@@ -0,0 +1,63 @@
+import { Box, Text } from '@anthropic/ink';
+import { useBackgroundAgentTasks } from '../../hooks/useBackgroundAgentTasks.js';
+import { useElapsedTime } from '../../hooks/useElapsedTime.js';
+import { useAppState } from '../../state/AppState.js';
+import type { LocalAgentTaskState } from '../../tasks/LocalAgentTask/LocalAgentTask.js';
+import { formatTokens } from '../../utils/format.js';
+
+function AgentRow({ task, selected }: { task: LocalAgentTaskState; selected: boolean }) {
+  const elapsed = useElapsedTime(task.startTime, task.status === 'running');
+  const tokens = task.progress?.tokenCount ?? 0;
+  const isRunning = task.status === 'running';
+  return (
+    <Box flexDirection="row" width="100%" justifyContent="space-between">
+      <Box flexDirection="row" flexShrink={1}>
+        <Text color={isRunning ? 'success' : undefined}>{selected ? '● ' : '○ '}</Text>
+        <Text bold={selected} wrap="truncate-end">
+          {task.agentType} <Text dimColor>{task.description}</Text>
+        </Text>
+      </Box>
+      <Box flexShrink={0}>
+        <Text dimColor>
+          {elapsed} · ↓ {formatTokens(tokens)} tokens
+        </Text>
+      </Box>
+    </Box>
+  );
+}
+
+function getHint(pillFocused: boolean, viewedTask: LocalAgentTaskState | null): string {
+  if (pillFocused) return '↑/↓ to select · Enter to view';
+  if (!viewedTask) return 'shift+↓ to manage background agents';
+  return viewedTask.status === 'running' ? 'shift+↓ to manage · x to stop' : 'shift+↓ to manage · x to clear';
+}
+
+export function BackgroundAgentSelector(): React.ReactNode {
+  const tasks = useBackgroundAgentTasks();
+  const viewingId = useAppState(s => s.viewingAgentTaskId);
+  const footerSelection = useAppState(s => s.footerSelection);
+  const selectedBgIndex = useAppState(s => s.selectedBgAgentIndex);
+
+  if (tasks.length === 0) return null;
+
+  const pillFocused = footerSelection === 'bg_agent';
+  const highlightedId = pillFocused
+    ? selectedBgIndex === -1
+      ? null
+      : (tasks[selectedBgIndex]?.agentId ?? null)
+    : (viewingId ?? null);
+  const mainHighlighted = pillFocused ? selectedBgIndex === -1 : viewingId === undefined;
+  const viewedTask = viewingId ? (tasks.find(t => t.agentId === viewingId) ?? null) : null;
+
+  return (
+    <Box flexDirection="column" width="100%">
+      <Box flexDirection="row" width="100%" justifyContent="space-between">
+        <Text bold={mainHighlighted}>{mainHighlighted ? '● ' : '○ '}main</Text>
+        <Text dimColor>{getHint(pillFocused, viewedTask)}</Text>
+      </Box>
+      {tasks.map(task => (
+        <AgentRow key={task.agentId} task={task} selected={task.agentId === highlightedId} />
+      ))}
+    </Box>
+  );
+}

src/hooks/useBackgroundAgentTasks.ts

@@ -0,0 +1,19 @@
+import { useMemo } from 'react'
+import { useAppState } from '../state/AppState.js'
+import {
+  isLocalAgentTask,
+  type LocalAgentTaskState,
+} from '../tasks/LocalAgentTask/LocalAgentTask.js'
+
+export function useBackgroundAgentTasks(): LocalAgentTaskState[] {
+  const tasks = useAppState(s => s.tasks)
+  return useMemo(() => {
+    const now = Date.now()
+    return Object.values(tasks)
+      .filter(isLocalAgentTask)
+      .filter(t => t.agentType !== 'main-session')
+      .filter(t => t.isBackgrounded !== false)
+      .filter(t => t.evictAfter === undefined || t.evictAfter > now)
+      .sort((a, b) => a.startTime - b.startTime)
+  }, [tasks])
+}

src/main.tsx

@@ -3485,6 +3485,7 @@ async function run(): Promise<CommanderCommand> {
             : 'none',
         showTeammateMessagePreview: isAgentSwarmsEnabled() ? false : undefined,
         selectedIPAgentIndex: -1,
+        selectedBgAgentIndex: -1,
         coordinatorTaskIndex: -1,
         viewSelectionMode: 'none',
         footerSelection: null,

src/screens/REPL.tsx

@@ -244,7 +244,14 @@ import {
   formatCommandInputTags,
 } from '../utils/messages.js';
 import { generateSessionTitle } from '../utils/sessionTitle.js';
-import { BASH_INPUT_TAG, COMMAND_MESSAGE_TAG, COMMAND_NAME_TAG, LOCAL_COMMAND_STDOUT_TAG } from '../constants/xml.js';
+import {
+  BASH_INPUT_TAG,
+  COMMAND_MESSAGE_TAG,
+  COMMAND_NAME_TAG,
+  FORK_BOILERPLATE_TAG,
+  LOCAL_COMMAND_STDOUT_TAG,
+} from '../constants/xml.js';
+import { FORK_SUBAGENT_TYPE } from '@claude-code-best/builtin-tools/tools/AgentTool/forkSubagent.js';
 import { escapeXml } from '../utils/xml.js';
 import type { ThinkingConfig } from '../utils/thinking.js';
 import { gracefulShutdownSync } from '../utils/gracefulShutdown.js';
@@ -336,6 +343,7 @@ import {
 import { isBgSession, updateSessionName, updateSessionActivity } from '../utils/concurrentSessions.js';
 import { isInProcessTeammateTask, type InProcessTeammateTaskState } from '../tasks/InProcessTeammateTask/types.js';
 import { restoreRemoteAgentTasks } from '../tasks/RemoteAgentTask/RemoteAgentTask.js';
+import { BackgroundAgentSelector } from '../components/tasks/BackgroundAgentSelector.js';
 import { useInboxPoller } from '../hooks/useInboxPoller.js';
 // Dead code elimination: conditional import for loop mode
 /* eslint-disable @typescript-eslint/no-require-imports */
@@ -800,6 +808,21 @@ export type Props = {
 
 export type Screen = 'prompt' | 'transcript';
 
+// Boilerplate carrier lives in a mixed user message ([tool_result..., text])
+// that AgentTool/forkSubagent.buildForkedMessages emits as the fork child's
+// first user turn. The text block wraps <FORK_BOILERPLATE_TAG>...</..> + the
+// user prompt; tool_result siblings keep the parent's tool calls closed.
+const FORK_BOILERPLATE_OPEN_TAG = `<${FORK_BOILERPLATE_TAG}>`;
+
+function isForkBoilerplateTextBlock(block: { type: string; text?: string }): boolean {
+  return block.type === 'text' && typeof block.text === 'string' && block.text.includes(FORK_BOILERPLATE_OPEN_TAG);
+}
+
+function isForkBoilerplateMessage(message: MessageType): boolean {
+  if (message.type !== 'user' || !Array.isArray(message.message?.content)) return false;
+  return message.message.content.some(isForkBoilerplateTextBlock);
+}
+
 export function REPL({
   commands: initialCommands,
   debug,
@@ -5548,8 +5571,72 @@ export function REPL({
   const usesSyncMessages = showStreamingText || !isLoading;
   // When viewing an agent, never fall through to leader — empty until
   // bootstrap/stream fills. Closes the see-leader-type-agent footgun.
+  const rawAgentMessages = viewedAgentTask?.messages;
+  // Fork sidechain encodes the user prompt inside a mixed user message alongside
+  // tool_result blocks; surface the prompt as a standalone bubble and strip the
+  // boilerplate text from its original carrier while preserving tool_results.
+  const displayedAgentMessages = useMemo(() => {
+    if (!viewedAgentTask) return undefined;
+    const agentMessages = rawAgentMessages ?? [];
+    if (
+      !isLocalAgentTask(viewedAgentTask) ||
+      viewedAgentTask.agentType !== FORK_SUBAGENT_TYPE ||
+      !viewedAgentTask.prompt
+    ) {
+      return agentMessages;
+    }
+    // Single pass: locate boilerplate carrier, check whether the prompt text is
+    // already present elsewhere, and find the fallback insertion point (after
+    // the last parent assistant tool_use).
+    const trimmedPrompt = viewedAgentTask.prompt.trim();
+    let boilerplateIndex = -1;
+    let lastAssistantToolUseIndex = -1;
+    let promptAlreadyRendered = false;
+    for (let i = 0; i < agentMessages.length; i++) {
+      const m = agentMessages[i]!;
+      if (m.type === 'user' && Array.isArray(m.message?.content)) {
+        const hasBoilerplate = m.message.content.some(isForkBoilerplateTextBlock);
+        if (hasBoilerplate) {
+          boilerplateIndex = i;
+        } else if (!promptAlreadyRendered) {
+          const firstText = m.message.content.find(b => b.type === 'text' && typeof b.text === 'string') as
+            | { type: 'text'; text: string }
+            | undefined;
+          if (firstText && firstText.text.trim() === trimmedPrompt) promptAlreadyRendered = true;
+        }
+        continue;
+      }
+      if (m.type === 'assistant' && Array.isArray(m.message?.content)) {
+        if (m.message.content.some(b => b.type === 'tool_use')) lastAssistantToolUseIndex = i;
+      }
+    }
+
+    const stripped =
+      boilerplateIndex === -1
+        ? agentMessages
+        : agentMessages.map((m, i) => {
+            if (i !== boilerplateIndex) return m;
+            if (!Array.isArray(m.message?.content)) return m;
+            return {
+              ...m,
+              message: {
+                ...m.message,
+                content: m.message.content.filter(b => !isForkBoilerplateTextBlock(b)),
+              },
+            };
+          });
+
+    if (promptAlreadyRendered) return stripped;
+
+    const insertAt = boilerplateIndex !== -1 ? boilerplateIndex + 1 : lastAssistantToolUseIndex + 1;
+    const synthetic = createUserMessage({
+      content: viewedAgentTask.prompt,
+      timestamp: new Date(viewedAgentTask.startTime).toISOString(),
+    });
+    return [...stripped.slice(0, insertAt), synthetic, ...stripped.slice(insertAt)];
+  }, [viewedAgentTask, rawAgentMessages]);
   const displayedMessages = viewedAgentTask
-    ? (viewedAgentTask.messages ?? [])
+    ? (displayedAgentMessages ?? [])
     : usesSyncMessages
       ? messages
       : deferredMessages;
@@ -6286,6 +6373,7 @@ export function REPL({
                       voiceInterimRange={voice.interimRange}
                     />
                     <SessionBackgroundHint onBackgroundSession={handleBackgroundSession} isLoading={isLoading} />
+                    <BackgroundAgentSelector />
                   </>
                 )}
                 {cursor && (

src/services/api/claude.ts

@@ -509,10 +509,30 @@ export function getAPIMetadata() {
     }
   }
 
+  const deviceId = getOrCreateUserID()
+
+  // Third-party API providers (DeepSeek, etc.) validate user_id against
+  // ^[a-zA-Z0-9_-]+$ which rejects JSON strings containing {, ", :, etc.
+  // When using a non-Anthropic base URL, send only the device_id (hex string).
+  const baseUrl = process.env.ANTHROPIC_BASE_URL
+  const isThirdParty =
+    baseUrl &&
+    (() => {
+      try {
+        return new URL(baseUrl).host !== 'api.anthropic.com'
+      } catch {
+        return false
+      }
+    })()
+
+  if (isThirdParty) {
+    return { user_id: deviceId }
+  }
+
   return {
     user_id: jsonStringify({
       ...extra,
-      device_id: getOrCreateUserID(),
+      device_id: deviceId,
       // Only include OAuth account UUID when actively using OAuth authentication
       account_uuid: getOauthAccountInfo()?.accountUuid ?? '',
       session_id: getSessionId(),

src/services/mcp/client.ts

@@ -74,7 +74,10 @@ import {
 } from '../../utils/errors.js'
 import { getMCPUserAgent } from '../../utils/http.js'
 import { maybeNotifyIDEConnected } from '../../utils/ide.js'
-import { maybeResizeAndDownsampleImageBuffer } from '../../utils/imageResizer.js'
+import {
+  type ImageLimits,
+  maybeResizeAndDownsampleImageBuffer,
+} from '../../utils/imageResizer.js'
 import { logMCPDebug, logMCPError } from '../../utils/log.js'
 import {
   getBinaryBlobSavedMessage,
@@ -102,6 +105,7 @@ import {
   isPersistError,
   persistToolResult,
 } from '../../utils/toolResultStorage.js'
+import { checkStatsigFeatureGate_CACHED_MAY_BE_STALE } from '../analytics/growthbook.js'
 import {
   type AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
   logEvent,
@@ -2486,15 +2490,23 @@ export function prefetchAllMcpResources(
 export async function transformResultContent(
   resultContent: PromptMessage['content'],
   serverName: string,
+  limits?: ImageLimits,
+  includeMeta = false,
 ): Promise<Array<ContentBlockParam>> {
   switch (resultContent.type) {
-    case 'text':
-      return [
-        {
-          type: 'text',
-          text: resultContent.text,
-        },
-      ]
+    case 'text': {
+      const block: ContentBlockParam = {
+        type: 'text',
+        text: resultContent.text,
+      }
+      if (includeMeta) {
+        const meta = resultContent._meta
+        if (meta) {
+          ;(block as { _meta?: unknown })._meta = meta
+        }
+      }
+      return [block]
+    }
     case 'audio': {
       const audioData = resultContent as {
         type: 'audio'
@@ -2516,6 +2528,7 @@ export async function transformResultContent(
         imageBuffer,
         imageBuffer.length,
         ext,
+        limits,
       )
       return [
         {
@@ -2551,6 +2564,7 @@ export async function transformResultContent(
             imageBuffer,
             imageBuffer.length,
             ext,
+            limits,
           )
           const content: MessageParam['content'] = []
           if (prefix) {
@@ -2671,6 +2685,7 @@ export async function transformMCPResult(
   result: unknown,
   tool: string, // Tool name for validation (e.g., "search")
   name: string, // Server name for transformation (e.g., "slack")
+  limits?: ImageLimits, // Image processing limits, plumbed to transformResultContent
 ): Promise<TransformedMCPResult> {
   if (result && typeof result === 'object') {
     if ('toolResult' in result) {
@@ -2694,7 +2709,9 @@ export async function transformMCPResult(
     if ('content' in result && Array.isArray(result.content)) {
       const transformedContent = (
         await Promise.all(
-          result.content.map(item => transformResultContent(item, name)),
+          result.content.map(item =>
+            transformResultContent(item, name, limits, true),
+          ),
         )
       ).flat()
       return {
@@ -2729,8 +2746,15 @@ export async function processMCPResult(
   result: unknown,
   tool: string, // Tool name for validation (e.g., "search")
   name: string, // Server name for IDE check and transformation (e.g., "slack")
+  limits?: ImageLimits, // Image processing limits, plumbed to transformMCPResult
+  skipLargeOutput = false, // If true, skip large-output handling for non-image content
 ): Promise<MCPToolResult> {
-  const { content, type, schema } = await transformMCPResult(result, tool, name)
+  const { content, type, schema } = await transformMCPResult(
+    result,
+    tool,
+    name,
+    limits,
+  )
 
   // IDE tools are not going to the model directly, so we don't need to
   // handle large output.
@@ -2738,6 +2762,12 @@ export async function processMCPResult(
     return content
   }
 
+  // Caller opted out of large-output handling (e.g., result already truncated
+  // upstream); only continue if the content has images that may need handling.
+  if (skipLargeOutput && !contentContainsImages(content)) {
+    return content
+  }
+
   // Check if content needs truncation (i.e., is too large)
   if (!(await mcpContentNeedsTruncation(content))) {
     return content
@@ -2775,9 +2805,15 @@ export async function processMCPResult(
   // Generate a unique ID for the persisted file (server__tool-timestamp)
   const timestamp = Date.now()
   const persistId = `mcp-${normalizeNameForMCP(name)}-${normalizeNameForMCP(tool)}-${timestamp}`
-  // Convert to string for persistence (persistToolResult expects string or specific block types)
+  // When the large-string format gate is on, unwrap a single bare text block
+  // (no annotations, no _meta) into raw text so the model gets plain text in
+  // the persisted file instead of a JSON-wrapped block. The `_meta` check is
+  // why transformResultContent preserves _meta on text blocks.
+  const unwrappedText = unwrapSingleTextBlock(content)
   const contentStr =
-    typeof content === 'string' ? content : jsonStringify(content, null, 2)
+    typeof content === 'string'
+      ? content
+      : (unwrappedText ?? jsonStringify(content, null, 2))
   const persistResult = await persistToolResult(contentStr, persistId)
 
   if (isPersistError(persistResult)) {
@@ -2798,7 +2834,10 @@ export async function processMCPResult(
     persistedSizeChars: persistResult.originalSize,
   } as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS)
 
-  const formatDescription = getFormatDescription(type, schema)
+  const formatDescription =
+    unwrappedText !== undefined
+      ? getFormatDescription('toolResult')
+      : getFormatDescription(type, schema)
   return getLargeOutputInstructions(
     persistResult.filepath,
     persistResult.originalSize,
@@ -2806,6 +2845,39 @@ export async function processMCPResult(
   )
 }
 
+/**
+ * Returns true when the large-string output format is enabled (matches
+ * binary's `sf8()`). When enabled, processMCPResult unwraps a single bare
+ * text block to raw text for persistence instead of JSON-wrapping it.
+ *
+ * Gating sources, in order:
+ *   1. MCP_TRUNCATION_PROMPT_OVERRIDE env var (anything except "legacy" enables)
+ *   2. Statsig gate `tengu_mcp_subagent_prompt`
+ */
+function isLargeStringFormatEnabled(): boolean {
+  const override = process.env.MCP_TRUNCATION_PROMPT_OVERRIDE
+  if (override) return override !== 'legacy'
+  return checkStatsigFeatureGate_CACHED_MAY_BE_STALE(
+    'tengu_mcp_subagent_prompt',
+  )
+}
+
+/**
+ * Unwraps a single bare text content block to its raw text when the
+ * large-string format gate is on. Returns undefined when the gate is off,
+ * the content is not an array, the array doesn't contain exactly one text
+ * block, or the block carries annotations or _meta. Matches binary mg5's
+ * `M=...` computation.
+ */
+function unwrapSingleTextBlock(content: MCPToolResult): string | undefined {
+  if (!isLargeStringFormatEnabled()) return undefined
+  if (!Array.isArray(content) || content.length !== 1) return undefined
+  const block = content[0]
+  if (!block || block.type !== 'text') return undefined
+  if ('annotations' in block || '_meta' in block) return undefined
+  return block.text
+}
+
 /**
  * Call an MCP tool, handling UrlElicitationRequiredError (-32042) by
  * displaying the URL elicitation to the user, waiting for the completion
@@ -2827,6 +2899,8 @@ export async function callMCPToolWithUrlElicitationRetry({
   signal,
   setAppState,
   onProgress,
+  imageLimits,
+  hasResultSizeAnnotation = false,
   callToolFn = callMCPTool,
   handleElicitation,
 }: {
@@ -2838,6 +2912,8 @@ export async function callMCPToolWithUrlElicitationRetry({
   signal: AbortSignal
   setAppState: (f: (prev: AppState) => AppState) => void
   onProgress?: (data: MCPProgress) => void
+  imageLimits?: ImageLimits
+  hasResultSizeAnnotation?: boolean
   /** Injectable for testing. Defaults to callMCPTool. */
   callToolFn?: (opts: {
     client: ConnectedMCPServer
@@ -2846,6 +2922,8 @@ export async function callMCPToolWithUrlElicitationRetry({
     meta?: Record<string, unknown>
     signal: AbortSignal
     onProgress?: (data: MCPProgress) => void
+    imageLimits?: ImageLimits
+    hasResultSizeAnnotation?: boolean
   }) => Promise<MCPToolCallResult>
   /** Handler for URL elicitations when no hook handles them.
    * In print/SDK mode, delegates to structuredIO. In REPL, falls back to queue. */
@@ -2865,6 +2943,8 @@ export async function callMCPToolWithUrlElicitationRetry({
         meta,
         signal,
         onProgress,
+        imageLimits,
+        hasResultSizeAnnotation,
       })
     } catch (error) {
       // The MCP SDK's Protocol creates plain McpError (not UrlElicitationRequiredError)
@@ -3041,6 +3121,8 @@ async function callMCPTool({
   meta,
   signal,
   onProgress,
+  imageLimits,
+  hasResultSizeAnnotation = false,
 }: {
   client: ConnectedMCPServer
   tool: string
@@ -3048,6 +3130,8 @@ async function callMCPTool({
   meta?: Record<string, unknown>
   signal: AbortSignal
   onProgress?: (data: MCPProgress) => void
+  imageLimits?: ImageLimits
+  hasResultSizeAnnotation?: boolean
 }): Promise<{
   content: MCPToolResult
   _meta?: Record<string, unknown>
@@ -3176,7 +3260,13 @@ async function callMCPTool({
       })
     }
 
-    const content = await processMCPResult(result, tool, name)
+    const content = await processMCPResult(
+      result,
+      tool,
+      name,
+      imageLimits,
+      hasResultSizeAnnotation,
+    )
     return {
       content,
       _meta: result._meta as Record<string, unknown> | undefined,

src/state/AppStateStore.ts

@@ -85,6 +85,7 @@ export type FooterItem =
   | 'teams'
   | 'bridge'
   | 'companion'
+  | 'bg_agent'
 
 export type AppState = DeepImmutable<{
   settings: SettingsJson
@@ -97,6 +98,9 @@ export type AppState = DeepImmutable<{
   // Optional - only present when ENABLE_AGENT_SWARMS is true (for dead code elimination)
   showTeammateMessagePreview?: boolean
   selectedIPAgentIndex: number
+  // Selection index for the bottom BackgroundAgentSelector.
+  // -1 = main, 0..N-1 = index into useBackgroundAgentTasks().
+  selectedBgAgentIndex: number
   // CoordinatorTaskPanel selection: -1 = pill, 0 = main, 1..N = agent rows.
   // AppState (not local) so the panel can read it directly without prop-drilling
   // through PromptInput → PromptInputFooter.
@@ -477,6 +481,7 @@ export function getDefaultAppState(): AppState {
     isBriefOnly: false,
     showTeammateMessagePreview: false,
     selectedIPAgentIndex: -1,
+    selectedBgAgentIndex: -1,
     coordinatorTaskIndex: -1,
     viewSelectionMode: 'none',
     footerSelection: null,

src/utils/imageResizer.ts

@@ -162,6 +162,16 @@ interface CompressedImageResult {
   originalSize: number
 }
 
+/**
+ * Per-call image processing limits, overriding the module-level defaults.
+ */
+export interface ImageLimits {
+  targetRawSize: number
+  maxWidth: number
+  maxHeight: number
+  maxBase64Size: number
+}
+
 /**
  * Extracted from FileReadTool's readImage function
  * Resizes image buffer to meet size and dimension constraints
@@ -170,7 +180,13 @@ export async function maybeResizeAndDownsampleImageBuffer(
   imageBuffer: Buffer,
   originalSize: number,
   ext: string,
+  limits?: ImageLimits,
 ): Promise<ResizeResult> {
+  const targetRawSize = limits?.targetRawSize ?? IMAGE_TARGET_RAW_SIZE
+  const maxWidth = limits?.maxWidth ?? IMAGE_MAX_WIDTH
+  const maxHeight = limits?.maxHeight ?? IMAGE_MAX_HEIGHT
+  const maxBase64Size = limits?.maxBase64Size ?? API_IMAGE_MAX_BASE64_SIZE
+
   if (imageBuffer.length === 0) {
     // Empty buffer would fall through the catch block below (sharp throws
     // "Unable to determine image format"), and the fallback's size check
@@ -189,7 +205,7 @@ export async function maybeResizeAndDownsampleImageBuffer(
 
     // If dimensions aren't available from metadata
     if (!metadata.width || !metadata.height) {
-      if (originalSize > IMAGE_TARGET_RAW_SIZE) {
+      if (originalSize > targetRawSize) {
         // Create fresh sharp instance for compression
         const compressedBuffer = await sharp(imageBuffer)
           .jpeg({ quality: 80 })
@@ -210,9 +226,9 @@ export async function maybeResizeAndDownsampleImageBuffer(
 
     // Check if the original file just works
     if (
-      originalSize <= IMAGE_TARGET_RAW_SIZE &&
-      width <= IMAGE_MAX_WIDTH &&
-      height <= IMAGE_MAX_HEIGHT
+      originalSize <= targetRawSize &&
+      width <= maxWidth &&
+      height <= maxHeight
     ) {
       return {
         buffer: imageBuffer,
@@ -226,20 +242,19 @@ export async function maybeResizeAndDownsampleImageBuffer(
       }
     }
 
-    const needsDimensionResize =
-      width > IMAGE_MAX_WIDTH || height > IMAGE_MAX_HEIGHT
+    const needsDimensionResize = width > maxWidth || height > maxHeight
     const isPng = normalizedMediaType === 'png'
 
     // If dimensions are within limits but file is too large, try compression first
     // This preserves full resolution when possible
-    if (!needsDimensionResize && originalSize > IMAGE_TARGET_RAW_SIZE) {
+    if (!needsDimensionResize && originalSize > targetRawSize) {
       // For PNGs, try PNG compression first to preserve transparency
       if (isPng) {
         // Create fresh sharp instance for each compression attempt
         const pngCompressed = await sharp(imageBuffer)
           .png({ compressionLevel: 9, palette: true })
           .toBuffer()
-        if (pngCompressed.length <= IMAGE_TARGET_RAW_SIZE) {
+        if (pngCompressed.length <= targetRawSize) {
           return {
             buffer: pngCompressed,
             mediaType: 'png',
@@ -258,7 +273,7 @@ export async function maybeResizeAndDownsampleImageBuffer(
         const compressedBuffer = await sharp(imageBuffer)
           .jpeg({ quality })
           .toBuffer()
-        if (compressedBuffer.length <= IMAGE_TARGET_RAW_SIZE) {
+        if (compressedBuffer.length <= targetRawSize) {
           return {
             buffer: compressedBuffer,
             mediaType: 'jpeg',
@@ -275,14 +290,14 @@ export async function maybeResizeAndDownsampleImageBuffer(
     }
 
     // Constrain dimensions if needed
-    if (width > IMAGE_MAX_WIDTH) {
-      height = Math.round((height * IMAGE_MAX_WIDTH) / width)
-      width = IMAGE_MAX_WIDTH
+    if (width > maxWidth) {
+      height = Math.round((height * maxWidth) / width)
+      width = maxWidth
     }
 
-    if (height > IMAGE_MAX_HEIGHT) {
-      width = Math.round((width * IMAGE_MAX_HEIGHT) / height)
-      height = IMAGE_MAX_HEIGHT
+    if (height > maxHeight) {
+      width = Math.round((width * maxHeight) / height)
+      height = maxHeight
     }
 
     // IMPORTANT: Always create fresh sharp(imageBuffer) instances for each operation.
@@ -298,7 +313,7 @@ export async function maybeResizeAndDownsampleImageBuffer(
       .toBuffer()
 
     // If still too large after resize, try compression
-    if (resizedImageBuffer.length > IMAGE_TARGET_RAW_SIZE) {
+    if (resizedImageBuffer.length > targetRawSize) {
       // For PNGs, try PNG compression first to preserve transparency
       if (isPng) {
         const pngCompressed = await sharp(imageBuffer)
@@ -308,7 +323,7 @@ export async function maybeResizeAndDownsampleImageBuffer(
           })
           .png({ compressionLevel: 9, palette: true })
           .toBuffer()
-        if (pngCompressed.length <= IMAGE_TARGET_RAW_SIZE) {
+        if (pngCompressed.length <= targetRawSize) {
           return {
             buffer: pngCompressed,
             mediaType: 'png',
@@ -331,7 +346,7 @@ export async function maybeResizeAndDownsampleImageBuffer(
           })
           .jpeg({ quality })
           .toBuffer()
-        if (compressedBuffer.length <= IMAGE_TARGET_RAW_SIZE) {
+        if (compressedBuffer.length <= targetRawSize) {
           return {
             buffer: compressedBuffer,
             mediaType: 'jpeg',
@@ -407,11 +422,11 @@ export async function maybeResizeAndDownsampleImageBuffer(
       imageBuffer[1] === 0x50 &&
       imageBuffer[2] === 0x4e &&
       imageBuffer[3] === 0x47 &&
-      (imageBuffer.readUInt32BE(16) > IMAGE_MAX_WIDTH ||
-        imageBuffer.readUInt32BE(20) > IMAGE_MAX_HEIGHT)
+      (imageBuffer.readUInt32BE(16) > maxWidth ||
+        imageBuffer.readUInt32BE(20) > maxHeight)
 
     // If original image's base64 encoding is within API limit, allow it through uncompressed
-    if (base64Size <= API_IMAGE_MAX_BASE64_SIZE && !overDim) {
+    if (base64Size <= maxBase64Size && !overDim) {
       logEvent('tengu_image_resize_fallback', {
         original_size_bytes: originalSize,
         base64_size_bytes: base64Size,
@@ -423,7 +438,7 @@ export async function maybeResizeAndDownsampleImageBuffer(
     // Image is too large and we failed to compress it - fail with user-friendly error
     throw new ImageResizeError(
       overDim
-        ? `Unable to resize image — dimensions exceed the ${IMAGE_MAX_WIDTH}x${IMAGE_MAX_HEIGHT}px limit and image processing failed. ` +
+        ? `Unable to resize image — dimensions exceed the ${maxWidth}x${maxHeight}px limit and image processing failed. ` +
             `Please resize the image to reduce its pixel dimensions.`
         : `Unable to resize image (${formatFileSize(originalSize)} raw, ${formatFileSize(base64Size)} base64). ` +
             `The image exceeds the 5MB API limit and compression failed. ` +