fix(tasks): scope a subagent's task list to the agent

`getTaskListId()` fell back to the session id, and a subagent runs
in-process, so every task one created landed in its parent session's list
and rendered in the UI as if the assistant had planned it. Observed with a
workflow whose scan agent filed three review tasks into the main session.

The reach was the bug, not the capability: an agent tracking its own work
is what lets it hold a goal across a long run. So the list is scoped by
agent rather than the tools being withheld. `TodoWriteTool` has always
keyed on `context.agentId ?? getSessionId()`; this brings the four task
tools in line with it.

Teammates are resolved before the agent branch and keep sharing the
leader's list, which is the point of a team. Also corrects the docstring's
claim that `CLAUDE_CODE_TEAM_NAME` is consulted — `getTeamName()` never
reads it.

src/tools/TaskCreateTool/TaskCreateTool.ts

@@ -78,7 +78,7 @@ export const TaskCreateTool = buildTool({
     return null
   },
   async call({ subject, description, activeForm, metadata }, context) {
-    const taskId = await createTask(getTaskListId(), {
+    const taskId = await createTask(getTaskListId(context?.agentId), {
       subject,
       description,
       activeForm,
@@ -108,7 +108,7 @@ export const TaskCreateTool = buildTool({
     }
 
     if (blockingErrors.length > 0) {
-      await deleteTask(getTaskListId(), taskId)
+      await deleteTask(getTaskListId(context?.agentId), taskId)
       throw new Error(blockingErrors.join('\n'))
     }
 

src/tools/TaskGetTool/TaskGetTool.ts

@@ -70,8 +70,8 @@ export const TaskGetTool = buildTool({
   renderToolUseMessage() {
     return null
   },
-  async call({ taskId }) {
-    const taskListId = getTaskListId()
+  async call({ taskId }, context) {
+    const taskListId = getTaskListId(context?.agentId)
 
     const task = await getTask(taskListId, taskId)
 

src/tools/TaskListTool/TaskListTool.ts

@@ -62,8 +62,8 @@ export const TaskListTool = buildTool({
   renderToolUseMessage() {
     return null
   },
-  async call() {
-    const taskListId = getTaskListId()
+  async call(_input, context) {
+    const taskListId = getTaskListId(context?.agentId)
 
     const allTasks = (await listTasks(taskListId)).filter(
       t => !t.metadata?._internal,

src/tools/TaskUpdateTool/TaskUpdateTool.ts

@@ -134,7 +134,7 @@ export const TaskUpdateTool = buildTool({
     },
     context,
   ) {
-    const taskListId = getTaskListId()
+    const taskListId = getTaskListId(context?.agentId)
 
     // Auto-expand task list when updating tasks
     context.setAppState(prev => {

src/utils/taskListId.test.ts

@@ -0,0 +1,56 @@
+import { afterEach, beforeEach, describe, expect, test } from 'bun:test'
+import { runWithTeammateContext, type TeammateContext } from './teammateContext.js'
+import { getTaskListId } from './tasks.js'
+
+// The real context API rather than a module stub: stubbing the module drops
+// its other exports and breaks every importer of it.
+const TEAMMATE: TeammateContext = {
+  agentId: 'researcher@review-crew',
+  agentName: 'researcher',
+  teamName: 'review-crew',
+  planModeRequired: false,
+  parentSessionId: 'leader-session',
+  isInProcessTeammate: true,
+}
+
+describe('getTaskListId', () => {
+  const originalTaskList = process.env.CLAUDE_CODE_TASK_LIST_ID
+
+  beforeEach(() => {
+    delete process.env.CLAUDE_CODE_TASK_LIST_ID
+  })
+
+  afterEach(() => {
+    if (originalTaskList) process.env.CLAUDE_CODE_TASK_LIST_ID = originalTaskList
+  })
+
+  test('a subagent keeps its own list instead of writing the session list', () => {
+    // The bug this guards: a subagent runs in-process, so without agent
+    // scoping every task it created landed in the parent session's list and
+    // rendered in the UI as if the assistant had planned it. The agent keeps
+    // its tracking — it just no longer shares the session's.
+    expect(getTaskListId('a7e7846c7975c3665')).toBe('a7e7846c7975c3665')
+    expect(getTaskListId('a7e7846c7975c3665')).not.toBe(getTaskListId())
+  })
+
+  test('two subagents never share a list', () => {
+    // Twenty parallel workflow agents on one list is also a write race.
+    expect(getTaskListId('agent-one')).not.toBe(getTaskListId('agent-two'))
+  })
+
+  test('the session keeps a stable list of its own', () => {
+    expect(getTaskListId()).toBe(getTaskListId())
+    expect(getTaskListId()).toBeTruthy()
+  })
+
+  test('an explicit task list id still wins over the agent', () => {
+    process.env.CLAUDE_CODE_TASK_LIST_ID = 'explicit-list'
+    expect(getTaskListId('some-agent')).toBe('explicit-list')
+  })
+
+  test('a teammate keeps sharing the leader list — that is the point of a team', () => {
+    runWithTeammateContext(TEAMMATE, () => {
+      expect(getTaskListId('teammate-agent')).toBe('review-crew')
+    })
+  })
+})

src/utils/tasks.ts

@@ -192,11 +192,20 @@ export async function resetTaskList(taskListId: string): Promise<void> {
  * Priority:
  * 1. CLAUDE_CODE_TASK_LIST_ID - explicit task list ID
  * 2. In-process teammate: leader's team name (so teammates share the leader's task list)
- * 3. CLAUDE_CODE_TEAM_NAME - set when running as a process-based teammate
+ * 3. Team name from the active team context
  * 4. Leader team name - set when the leader creates a team via TeamCreate
- * 5. Session ID - fallback for standalone sessions
+ * 5. Agent ID - a subagent keeps its own list instead of writing the session's
+ * 6. Session ID - fallback for standalone sessions
+ *
+ * A subagent runs in-process, so without step 5 every task it created landed
+ * in its parent's list and surfaced in the UI as if the assistant had planned
+ * it. Scoping by agent keeps the agent's own tracking — which is what lets it
+ * hold a goal across a long run — while keeping it out of the session's list.
+ * `TodoWriteTool` has always keyed on `context.agentId ?? getSessionId()`;
+ * this brings the task tools in line with it. Teammates are checked first and
+ * so keep sharing the leader's list, which is the point of a team.
  */
-export function getTaskListId(): string {
+export function getTaskListId(agentId?: string): string {
   if (process.env.CLAUDE_CODE_TASK_LIST_ID) {
     return process.env.CLAUDE_CODE_TASK_LIST_ID
   }
@@ -206,7 +215,7 @@ export function getTaskListId(): string {
   if (teammateCtx) {
     return teammateCtx.teamName
   }
-  return getTeamName() || leaderTeamName || getSessionId()
+  return getTeamName() || leaderTeamName || agentId || getSessionId()
 }
 
 /**