diff --git a/apps/ai-game-creator-shell/src-tauri/src/agent/runtime_driver/recovery_scan.rs b/apps/ai-game-creator-shell/src-tauri/src/agent/runtime_driver/recovery_scan.rs
index 775421806..60b4ac18e 100644
--- a/apps/ai-game-creator-shell/src-tauri/src/agent/runtime_driver/recovery_scan.rs
+++ b/apps/ai-game-creator-shell/src-tauri/src/agent/runtime_driver/recovery_scan.rs
@@ -1049,8 +1049,18 @@ pub(in crate::agent) fn resume_game_creator_agent_background_tasks_unredacted_at
         // and reacquire in project -> execution order, re-read the task under
         // the new lock set, project exactly once, then release the project lock
         // before any later path can enter the Session lane.
+        //
+        // 但双锚缺失的 run 要让路。本块修 session 投影，是为了让**将要继续**的
+        // continuation 不带着过期决定去发第一个 Provider 请求；而 `plan.submit_gdd`
+        // 的 pending/batch 恢复锚点同时丢失的 run 根本不会继续——下面
+        // `reconcile_missing_plan_submit_anchors_at` 会把它判失败并要求人工核对。
+        // 若不让路，本块会先落一条泛化的「未命中 planning session 协调器」，把那条
+        // 精确得多的双锚缺失诊断永久挡在后面：两者都写 needs-reconciliation，谁先
+        // 写谁赢，而先写的那条恰恰是信息量更少的。探测器自身很便宜，且对「已按双锚
+        // 缺失收敛过」的状态返回 None，所以这道让路不会反复触发。
         let runtime_lock = if agent_id == GAME_CREATOR_PROJECT_PLANNING_AGENT_ID
             && read_recoverable_runnable_game_creator_agent_runtime_task(root, &agent_id)?.is_some()
+            && missing_plan_submit_anchor_candidate_at(root, &agent_id)?.is_none()
         {
             drop(runtime_lock);
             let project_lock = match acquire_game_creator_agent_runtime_project_write_lock_with_wait(