diff --git a/apps/ai-game-creator-shell/src/App.tsx b/apps/ai-game-creator-shell/src/App.tsx
index aca14fac6..57ccaf972 100644
--- a/apps/ai-game-creator-shell/src/App.tsx
+++ b/apps/ai-game-creator-shell/src/App.tsx
@@ -5628,6 +5628,7 @@ export function App({
             PROJECT_SUPERVISOR_AGENT_ID,
             sessionId,
             submissionRunProfile,
+            'project-supervisor-game-chat',
           )
         : null;
       let autoPreviewAfterRevision = 0;
diff --git a/apps/ai-game-creator-shell/src/features/agent-runtime/model.ts b/apps/ai-game-creator-shell/src/features/agent-runtime/model.ts
index 65a2a05d4..32e2714c7 100644
--- a/apps/ai-game-creator-shell/src/features/agent-runtime/model.ts
+++ b/apps/ai-game-creator-shell/src/features/agent-runtime/model.ts
@@ -725,6 +725,7 @@ export function matchingAgentRuntimeForSteer(
   requestedRunProfile: NonNullable<
     AgentRuntimeState['runProfile']
   > = 'standard',
+  requestedSource?: string,
 ) {
   if (!sessionId) {
     return null;
@@ -735,6 +736,7 @@ export function matchingAgentRuntimeForSteer(
         runtime?.agentId === agentId &&
         runtime.sessionId === sessionId &&
         (runtime.runProfile ?? 'standard') === requestedRunProfile &&
+        (!requestedSource || runtime.source === requestedSource) &&
         isAgentRuntimeSteerableState(runtime),
     ) ?? null
   );
@@ -834,6 +836,7 @@ export async function submitProjectSupervisorRuntimeTask({
     PROJECT_SUPERVISOR_AGENT_ID,
     sessionId,
     runProfile,
+    source,
   );
   if (steerRuntime) {
     const steer = await invoke<AgentRuntimeSteerResult>(
@@ -1330,7 +1333,9 @@ export function projectSupervisorCollaboratingAgentRuntimes(
   const runtimesByAgentId = new Map<string, AgentRuntimeState>();
   for (const runtime of Object.values(runtimeByAgentId)) {
     const isVisibleChildSource =
-      ['agent-delegate', 'agent-delegate-retry'].includes(runtime?.source ?? '') ||
+      ['agent-delegate', 'agent-delegate-retry'].includes(
+        runtime?.source ?? '',
+      ) ||
       (supervisorRuntime.source === 'project-supervisor-game-chat' &&
         runtime?.source === 'agent-ready-task-scheduler');
     if (
diff --git a/apps/ai-game-creator-shell/tests/agentRuntimeModel.test.ts b/apps/ai-game-creator-shell/tests/agentRuntimeModel.test.ts
index 9a9721f8d..367cca150 100644
--- a/apps/ai-game-creator-shell/tests/agentRuntimeModel.test.ts
+++ b/apps/ai-game-creator-shell/tests/agentRuntimeModel.test.ts
@@ -82,7 +82,8 @@ describe('Game Chat stream identity and source', () => {
       eventPath: 'events.jsonl',
     } satisfies AgentRuntimeResult;
     const invoke = vi.fn(
-      async (_command: string, _args?: Record<string, unknown>) => runtimeResult,
+      async (_command: string, _args?: Record<string, unknown>) =>
+        runtimeResult,
     );
 
     await submitProjectSupervisorRuntimeTask({
@@ -116,6 +117,7 @@ describe('Game Chat stream identity and source', () => {
       ...runtimeResult.state,
       runId: 'steer-run',
       runProfile: 'autonomous-game-build',
+      source: 'project-supervisor-game-chat',
       status: 'running',
       phase: 'planning',
     } satisfies AgentRuntimeState;
@@ -157,6 +159,24 @@ describe('Game Chat stream identity and source', () => {
       runProfile: 'autonomous-game-build',
     });
     expect(invoke.mock.calls[0]?.[1]).not.toHaveProperty('source');
+
+    invoke.mockClear();
+    await submitProjectSupervisorRuntimeTask({
+      invoke,
+      projectPath: '/tmp/game-chat',
+      sessionId: 'session-provider-retry',
+      prompt: 'continue from the app',
+      runtime: {
+        ...steerRuntime,
+        source: 'project-supervisor-cli',
+      },
+      runProfile: 'autonomous-game-build',
+      source: 'project-supervisor-game-chat',
+    });
+    expect(invoke).toHaveBeenCalledWith(
+      'start_game_creator_supervisor_runtime_task',
+      expect.objectContaining({ source: 'project-supervisor-game-chat' }),
+    );
   });
 
   test('game-chat hydration keeps identical text from a different runtime response identity', () => {
@@ -178,10 +198,9 @@ describe('Game Chat stream identity and source', () => {
       },
     ];
 
-    expect(mergeGameChatRuntimeResponseMessagesIntoHistory(history, pending)).toEqual([
-      pending[0],
-      history[0],
-    ]);
+    expect(
+      mergeGameChatRuntimeResponseMessagesIntoHistory(history, pending),
+    ).toEqual([pending[0], history[0]]);
     expect(
       mergeGameChatRuntimeResponseMessagesIntoHistory(history, [
         { ...pending[0], messageId: 'persisted-message-1' },
diff --git a/apps/ai-game-creator-shell/tests/appSurface/harness.ts b/apps/ai-game-creator-shell/tests/appSurface/harness.ts
index 7555d2a6d..71d427a97 100644
--- a/apps/ai-game-creator-shell/tests/appSurface/harness.ts
+++ b/apps/ai-game-creator-shell/tests/appSurface/harness.ts
@@ -471,9 +471,7 @@ function createProjectSupervisorRuntimeHarness({
         currentProjectMessages.push({
           schemaVersion: 'game-creator-conversation.v1',
           ...message,
-          ...(args?.messageId
-            ? { messageId: String(args.messageId) }
-            : {}),
+          ...(args?.messageId ? { messageId: String(args.messageId) } : {}),
           updatedAt:
             typeof message.updatedAt === 'number'
               ? message.updatedAt
@@ -517,6 +515,7 @@ function createProjectSupervisorRuntimeHarness({
         currentRuntime = runtimeState({
           runId,
           runProfile: args?.runProfile,
+          source: args?.source,
           status: 'running',
           phase: 'planning',
           currentTask: String(args?.task ?? ''),
diff --git a/apps/ai-game-creator-shell/tests/appSurface/project-development.suite.ts b/apps/ai-game-creator-shell/tests/appSurface/project-development.suite.ts
index 2a508e90f..41d5254b6 100644
--- a/apps/ai-game-creator-shell/tests/appSurface/project-development.suite.ts
+++ b/apps/ai-game-creator-shell/tests/appSurface/project-development.suite.ts
@@ -92,7 +92,7 @@ function gameChatRuntimeEvent({
     runId,
     source:
       agentId === 'project-supervisor'
-        ? 'project-supervisor'
+        ? 'project-supervisor-game-chat'
         : 'agent-delegate',
     eventId,
     eventType,
@@ -114,7 +114,7 @@ function gameChatRuntimeState(
     taskId: 'project-supervisor',
     sessionId: 'game-chat-supervisor-session',
     runId: 'game-chat-supervisor-run',
-    source: 'project-supervisor',
+    source: 'project-supervisor-game-chat',
     status: 'running',
     phase: 'execution',
     currentTask: '生成首版游戏',
diff --git a/docs/project-memory/shared-memory/pitfalls.md b/docs/project-memory/shared-memory/pitfalls.md
index 3c6a256d2..aa64acfac 100644
--- a/docs/project-memory/shared-memory/pitfalls.md
+++ b/docs/project-memory/shared-memory/pitfalls.md
@@ -3772,6 +3772,7 @@
 - Provider action 安全持久化补充：pending / provider action 的泄密检测不能因裸自然语言短语 `api key` 直接拒绝，否则 `agent.delegate` 中“不要暴露 External Editor API Key”等安全约束会被误报并阻断首批协作。赋值形式只允许完整匹配受控的“未配置 / 不可用 / 禁止读取”等状态或固定无密钥降级说明，不能用 `starts_with` 放行 `none-but-secret`、`not configured; actual value ...` 等安全前缀后的凭据；`**API Key**:`、`` `API Key`: ``、`API Key（生产）:` 等装饰或限定标签也必须识别为赋值。结构化字段标记 `apiKey / api_key`、`Authorization / Cookie`、`token / Bearer` 以及已知 secret token 形状仍必须检测并失败关闭。
 - Windows retry 扫描补充：`Path::strip_prefix(root)` 在 Windows 上得到的相对 `Path` 转字符串后使用反斜杠，不能直接传给只接受 portable `/` 的 Runtime JSON sidecar 读取器；否则 Runner 重启或显式 `--agent-resume` 扫描已到期 retry 时会报“项目文件路径不能包含反斜杠”，任务持续停在 `waiting-for-provider-retry`。目录扫描应按路径组件重组成 `/` 分隔的 UTF-8 相对路径，不要放宽全局路径校验。
 - 恢复交互：`needs-reconciliation` 即使没有 `pendingToolAction`，也必须提供显式“已核对，结束旧任务”；它只取消旧 run，不直接 retry。若取消后仍有 pending task，由 Runner 自动继续；只有队列为空且旧 run 已取消时，才允许创建新的 retry run，避免重复执行同一用户输入。自主构建 Supervisor 的 retry 不能改写为普通 `agent-background-task` source，必须从已验证的原 Run Profile 绑定恢复 `project-supervisor-gui / project-supervisor-cli` 可信来源；不得只信可追加的 task journal。
+- Steer source：前端选择可 steer Runtime 时不能只比较 Agent、Session 和 Run Profile，还必须在调用方声明 source 时精确比较持久 `source`。例如 game-chat 只能 steer `project-supervisor-game-chat`，不能把同 Session/Profile 的 `project-supervisor-cli` run 当成目标；source 不一致时应按当前入口新建或排队自己的 run，不能先调用后端再把“steer source 与当前 Run 不一致”暴露给用户。
 - 验证：前端回归同时覆盖零历史、无 Session 的初始空态、无 active Session 索引但存在持久 `needs-reconciliation` 总控 Runtime 的恢复展示，以及“先取消、队列为空后才重试”；真实 Windows 运行全部 tool-plan handoff 测试，确保相对句柄 rename、覆盖安装、回读和清理均通过。Responses 回归覆盖 system / user / assistant 文本分别序列化，并保留 user `input_text + input_image`；Runtime 回归覆盖“无效计划 → repair transport 等待 → steer → 新 cursor 再修复”，断言 cursor `0 / 1` 各有一条审计且不冲突。
 
 ## 固定画布产物返工不能变成任意覆盖，design-foundation 不能越权修程序