diff --git a/backend/packages/app/src/windup_app/server/orchestrator/model.py b/backend/packages/app/src/windup_app/server/orchestrator/model.py
index 25082871..5fdb62b9 100644
--- a/backend/packages/app/src/windup_app/server/orchestrator/model.py
+++ b/backend/packages/app/src/windup_app/server/orchestrator/model.py
@@ -38,6 +38,24 @@ class ActionType(StrEnum):
     CUSTOM = "custom"
 
 
+# 动作类型 → 帧数(产品口径)。**这是唯一一份约定**:前端提交时不发 num_frames,
+# 从任务的 input_payload 读回来用 —— 两边各写一个数,分叉时任务照跑、没有一处会红。
+# 待机是原地小幅呼吸,32 帧里绝大多数帧之间没有差别,多出来的帧进不了有效循环,
+# 却照样占抽帧、抠图、对齐、上传的工作量与存储。
+ACTION_FRAME_COUNTS: dict[ActionType, int] = {
+    ActionType.WALK: 32,
+    ActionType.IDLE: 12,
+    ActionType.JUMP: 32,
+    ActionType.ATTACK: 32,
+    ActionType.CUSTOM: 32,
+}
+
+
+def frames_for(action_type: ActionType) -> int:
+    """该动作类型约定的帧数。"""
+    return ACTION_FRAME_COUNTS[action_type]
+
+
 class TaskStatus(StrEnum):
     """生成任务状态。"""
 
@@ -73,7 +91,10 @@ class CharacterActionInput:
     custom_prompt: str | None = None
     reference_video_url: str | None = None
     reference_image_urls: list[str] = field(default_factory=list)
-    num_frames: int = 32
+    # ``None`` = 调用方没指定,在 __post_init__ 里按动作类型解析成约定值。解析放在这层
+    # 而不是各个构造点:落库的 input_payload 是产线与前端读帧数的唯一来源,少解析一处
+    # 就多一个自带默认值的构造点(MQ 重建入参就是其中一个)。
+    num_frames: int | None = None
     # ── action_type=custom 才用到的两个(#239)──────────────────────────────
     # 这个动作是否循环播放。``None`` 原样往下传,由编排层兜成一次性:本层替调用方填默认值
     # 的话,"没给"和"明确给了 False"从这里起就再也分不开了。
@@ -97,6 +118,10 @@ class CharacterActionInput:
     stance: CharacterStance | None = None
     direction: ActionDirection = ActionDirection.EAST
 
+    def __post_init__(self) -> None:
+        if self.num_frames is None:
+            self.num_frames = frames_for(self.action_type)
+
 
 # -- 出参（按任务类型细化，前端可直接回填 character 模块）------------------
 
diff --git a/backend/packages/app/src/windup_app/web/api/generation.py b/backend/packages/app/src/windup_app/web/api/generation.py
index 04339291..939abd90 100644
--- a/backend/packages/app/src/windup_app/web/api/generation.py
+++ b/backend/packages/app/src/windup_app/web/api/generation.py
@@ -192,7 +192,9 @@ class CharacterActionGenerateRequest(BaseModel):
     reference_video_url: str | None = None
     reference_image_urls: list[str] = Field(default_factory=list)
     # 同上:帧数决定抽帧与逐帧抠图的工作量,上界 64 已远超引擎能出的有效周期长度。
-    num_frames: int = Field(default=32, ge=1, le=64)
+    # 不给则按动作类型取约定值(ACTION_FRAME_COUNTS)——写死一个默认值就等于替所有动作
+    # 都答了同一个数,而待机与走路要的帧数本来就不同。
+    num_frames: int | None = Field(default=None, ge=1, le=64)
     # ── action_type=custom 才用到(#239)───────────────────────────────────
     # 这个动作是否循环播放。不给则编排层兜成一次性,也不按描述文字猜 —— 两个方向的代价
     # 不对称:一次性动作被当成循环会让末帧接回首帧抽搐、产物不可用,反之只是不无缝闭环、
diff --git a/backend/packages/app/src/windup_app/worker/handlers.py b/backend/packages/app/src/windup_app/worker/handlers.py
index 01eedb4a..53668cd7 100644
--- a/backend/packages/app/src/windup_app/worker/handlers.py
+++ b/backend/packages/app/src/windup_app/worker/handlers.py
@@ -76,13 +76,16 @@ def _image_input(payload: dict) -> CharacterImageInput:
 def _action_input(payload: dict) -> CharacterActionInput:
     raw_type = payload.get("action_type")
     action_type = raw_type if isinstance(raw_type, ActionType) else ActionType(raw_type)
+    # 帧数缺失时原样传 None,交给入参按动作类型解析:在这里兜一个数就是第二份约定,
+    # 它与真正的约定分叉时任务照跑、帧数照出,没有一处会红。
+    raw_frames = payload.get("num_frames")
     return CharacterActionInput(
         character_id=int(payload["character_id"]),
         action_type=action_type,
         custom_prompt=payload.get("custom_prompt"),
         reference_video_url=payload.get("reference_video_url"),
         reference_image_urls=list(payload.get("reference_image_urls") or []),
-        num_frames=int(payload.get("num_frames") or 16),
+        num_frames=int(raw_frames) if raw_frames is not None else None,
         loop=payload.get("loop"),
         video_model=payload.get("video_model"),
         outfit_id=payload.get("outfit_id"),
diff --git a/backend/tests/test_generation_api.py b/backend/tests/test_generation_api.py
index 9428db8b..b360b2e9 100644
--- a/backend/tests/test_generation_api.py
+++ b/backend/tests/test_generation_api.py
@@ -475,3 +475,65 @@ def test_illegal_stance_is_rejected_at_the_entrance(auth_client):
         json=_action_payload(project["id"], character["id"], stance="octopod"),
     ).json()
     assert body["code"] == 400, body
+
+
+# ── 帧数按动作类型取,且只有一份约定 ────────────────────────────────────────
+#
+# 32 曾同时是后端默认值和前端"这是完整动画任务"的判据,于是改一个动作的帧数会让
+# 前端认不出这类任务。现在约定只在 ACTION_FRAME_COUNTS 一处,前端提交时不发帧数。
+
+
+def test_idle_task_frames_come_from_the_convention(auth_client):
+    """待机任务落库的帧数取该动作类型的约定值。
+
+    断言的是任务 input_payload —— 前端与 MQ 重建都从它读帧数,只断言函数返回值的话,
+    请求层没接上也照样绿。
+    """
+    from windup_app.server.orchestrator.model import ActionType, frames_for
+
+    project = _create_project(auth_client)
+    character = _create_character(auth_client, project["id"])
+
+    body = auth_client.post(
+        "/generation/action",
+        json=_action_payload(project["id"], character["id"], action_type="idle"),
+    ).json()
+
+    assert body["data"] is not None, body
+    stored = body["data"]["input_payload"]["num_frames"]
+    assert stored == frames_for(ActionType.IDLE)
+    assert stored != 32, "待机还在按 32 帧生成"
+
+
+def test_walk_task_keeps_thirty_two_frames(auth_client):
+    """本改动只动待机;走路的帧数不变。"""
+    project = _create_project(auth_client)
+    character = _create_character(auth_client, project["id"])
+
+    body = auth_client.post(
+        "/generation/action",
+        json=_action_payload(project["id"], character["id"], action_type="walk"),
+    ).json()
+
+    assert body["data"]["input_payload"]["num_frames"] == 32
+
+
+def test_explicit_num_frames_wins_over_the_convention(auth_client):
+    project = _create_project(auth_client)
+    character = _create_character(auth_client, project["id"])
+
+    body = auth_client.post(
+        "/generation/action",
+        json=_action_payload(
+            project["id"], character["id"], action_type="idle", num_frames=20,
+        ),
+    ).json()
+
+    assert body["data"]["input_payload"]["num_frames"] == 20
+
+
+def test_frame_convention_covers_every_action_type():
+    """新增动作类型必须同时给出帧数 —— 漏了的话取帧数会在请求里抛 KeyError。"""
+    from windup_app.server.orchestrator.model import ACTION_FRAME_COUNTS, ActionType
+
+    assert set(ACTION_FRAME_COUNTS) == set(ActionType)
diff --git a/backend/tests/test_mq_worker.py b/backend/tests/test_mq_worker.py
index 59013450..8ccae794 100644
--- a/backend/tests/test_mq_worker.py
+++ b/backend/tests/test_mq_worker.py
@@ -915,3 +915,26 @@ def test_consumer_acquires_generation_semaphore(engine, worker_session, monkeypa
 
     redis_mock.xack.assert_called_once()
 
+
+
+def test_action_input_takes_frames_from_the_convention():
+    """MQ 重建入参时缺帧数就按动作类型取,不在这层兜一个自己的数。
+
+    生产走的就是这条重建路径:这里兜的数与约定分叉时,任务照跑、帧照出,没有一处会红。
+    """
+    from windup_app.server.orchestrator.model import frames_for
+    from windup_app.worker.handlers import _action_input
+
+    rebuilt = _action_input({"character_id": 1, "action_type": "idle"})
+
+    assert rebuilt.num_frames == frames_for(ActionType.IDLE)
+    assert rebuilt.num_frames != 32
+
+
+def test_action_input_keeps_the_stored_frame_count():
+    """落库的帧数是产线唯一来源,重建时原样取,不拿约定覆盖它。"""
+    from windup_app.worker.handlers import _action_input
+
+    rebuilt = _action_input({"character_id": 1, "action_type": "idle", "num_frames": 20})
+
+    assert rebuilt.num_frames == 20
diff --git a/frontend/src/entities/generation/api.test.ts b/frontend/src/entities/generation/api.test.ts
index f6b46ec1..191a8fec 100644
--- a/frontend/src/entities/generation/api.test.ts
+++ b/frontend/src/entities/generation/api.test.ts
@@ -298,7 +298,7 @@ describe('createGenerationApis', () => {
     ).rejects.toThrow('动作首帧生成必须提供已确认的角色母版')
   })
 
-  it('以首帧请求完整动画并按后端 index 排序，当前合同固定为三十二帧', async () => {
+  it('以首帧请求完整动画并按后端 index 排序', async () => {
     const request = vi.fn(async (_url: string, _init?: RequestInit) =>
       success(
         taskData({
@@ -337,7 +337,6 @@ describe('createGenerationApis', () => {
       custom_prompt: 'move forward',
       reference_video_url: null,
       reference_image_urls: ['https://cdn.test/frame-1.png', 'https://cdn.test/extra.png'],
-      num_frames: 32,
       outfit_id: 'default',
       direction: 'east',
     })
@@ -351,6 +350,46 @@ describe('createGenerationApis', () => {
     })
   })
 
+  it('提交动作不发帧数，待机任务的十二帧照样识别得出阶段并映射结果', async () => {
+    const request = vi.fn(async (_url: string, _init?: RequestInit) =>
+      success(
+        taskData({
+          task_type: 'character_action',
+          input_payload: { num_frames: 12, action_type: 'idle' },
+          result: {
+            type: 'character_action',
+            action_type: 'idle',
+            frames: actionFrames(12),
+          },
+        }),
+      ),
+    )
+    const apis = createGenerationApis({
+      transport: { request, stream: vi.fn(() => vi.fn()) },
+    })
+
+    const created = await apis.create({
+      type: 'complete_animation',
+      projectId: '42',
+      characterId: '5',
+      outfitId: 'default',
+      method: 'video-cropping',
+      actionType: 'idle',
+      firstFrameUrl: 'https://cdn.test/frame-1.png',
+      prompt: null,
+      referenceMedia: [],
+    })
+    // 不带 expectation 查一次，走的是 inferExpectation：它必须只认 task_type 与
+    // action_type，认帧数的话待机任务在这里就抛"无法映射到前端阶段"。
+    const recovered = await apis.get('42', '91')
+
+    expect(JSON.parse(String(request.mock.calls[0]?.[1]?.body))).not.toHaveProperty('num_frames')
+    expect(created.result).toMatchObject({ type: 'complete_animation' })
+    expect((created.result as { frames: readonly unknown[] }).frames).toHaveLength(12)
+    expect(recovered.type).toBe('complete_animation')
+    expect((recovered.result as { frames: readonly unknown[] }).frames).toHaveLength(12)
+  })
+
   it('选视频裁剪时不发 outfit_id——后端拿它在场与否当三渲二的唯一判据', async () => {
     const request = vi.fn(async (_url: string, _init?: RequestInit) =>
       success(
diff --git a/frontend/src/entities/generation/api.ts b/frontend/src/entities/generation/api.ts
index ec2932b0..2d39d9dc 100644
--- a/frontend/src/entities/generation/api.ts
+++ b/frontend/src/entities/generation/api.ts
@@ -246,9 +246,26 @@ function mapImageResult(
     : { type: expectation.type, direction: expectation.direction, images }
 }
 
+/**
+ * 任务声明的帧数。哪种动作出多少帧由后端定，前端读回来当结果帧数的判据——
+ * 在前端也写一个数就是第二份约定，与后端分叉时两边都不会报错。
+ */
+function declaredFrameCount(
+  inputPayload: Record<string, unknown> | null,
+  expectation: GenerationExpectation,
+): number | undefined {
+  if (expectation.type !== 'complete_animation' || inputPayload === null) return undefined
+  const value = inputPayload.num_frames
+  if (!Number.isSafeInteger(value) || (value as number) <= 0) {
+    throw new GenerationApiError('动作任务 input_payload.num_frames 无效', 200)
+  }
+  return value as number
+}
+
 function mapActionResult(
   result: Record<string, unknown>,
   expectation: Extract<GenerationExpectation, { type: 'complete_animation' }>,
+  frameCount: number | undefined,
 ): GenerationResult {
   if (result.type !== 'character_action') {
     throw new GenerationApiError('完整动画结果 type 无效', 200)
@@ -293,7 +310,8 @@ function mapActionResult(
   })
 
   const orderedFrames = frames.sort((left, right) => left.index - right.index)
-  const expectedFrameCount = 32
+  // 事件没带 input_payload 时无从比对帧数，退回只查连续性——不能拿一个前端猜的数当判据。
+  const expectedFrameCount = frameCount ?? orderedFrames.length
   if (orderedFrames.length !== expectedFrameCount) {
     throw new GenerationApiError(`完整动画结果必须包含 ${expectedFrameCount} 帧`, 200)
   }
@@ -316,6 +334,7 @@ function mapResult(
   status: TaskStatus,
   expectation: GenerationExpectation,
   expectedCandidateCount?: ImageCandidateCount,
+  frameCount?: number,
 ): GenerationResult | null {
   if (status !== 'completed') {
     if (result !== null) {
@@ -325,7 +344,7 @@ function mapResult(
   }
   if (result === null) throw new GenerationApiError('完成任务缺少 result', 200)
   return expectation.type === 'complete_animation'
-    ? mapActionResult(result, expectation)
+    ? mapActionResult(result, expectation, frameCount)
     : mapImageResult(result, expectation, expectedCandidateCount)
 }
 
@@ -365,13 +384,6 @@ function validateInputPayload(
     }
     return candidateCount
   }
-  const expectedFrameCount = 32
-  if (inputPayload.num_frames !== expectedFrameCount) {
-    throw new GenerationApiError(
-      `动作任务 input_payload.num_frames 必须为 ${expectedFrameCount}`,
-      200,
-    )
-  }
   if (inputPayload.action_type !== expectation.actionType) {
     throw new GenerationApiError('动作任务 input_payload.action_type 与请求不一致', 200)
   }
@@ -400,12 +412,11 @@ function inferExpectation(dto: GenerationTaskDto): GenerationExpectation {
   if (typeof actionType !== 'string' || !ACTION_TYPES.has(actionType)) {
     throw new GenerationApiError('动作任务 input_payload.action_type 无效', 200)
   }
-  if (dto.inputPayload.num_frames === 32) {
-    return direction === undefined
-      ? { type: 'complete_animation', actionType }
-      : { type: 'complete_animation', actionType, direction }
-  }
-  throw new GenerationApiError('动作任务 input_payload.num_frames 无法映射到前端阶段', 200)
+  // 阶段由 task_type 与 action_type 定：character_action 在前端只对应"完整动画"这一个阶段。
+  // 帧数是产物参数不是阶段判据——拿它判，改一个动作的帧数就会让这类任务整个认不出来。
+  return direction === undefined
+    ? { type: 'complete_animation', actionType }
+    : { type: 'complete_animation', actionType, direction }
 }
 
 function validateTaskIdentity(
@@ -450,7 +461,13 @@ function mapTask(
       projectId: String(dto.projectId),
       type: resolvedExpectation.type,
       status: dto.status,
-      result: mapResult(dto.result, dto.status, resolvedExpectation, candidateCount),
+      result: mapResult(
+        dto.result,
+        dto.status,
+        resolvedExpectation,
+        candidateCount,
+        declaredFrameCount(dto.inputPayload, resolvedExpectation),
+      ),
       error: dto.errorMessage,
     },
     ...(candidateCount === undefined ? {} : { candidateCount }),
@@ -530,14 +547,14 @@ function mapEvent<TType extends GenerationType>(
   ) {
     throw new GenerationApiError('task_update 不属于当前项目', 200)
   }
-  const candidateCount =
+  const inputPayload =
     value.input_payload === undefined
+      ? undefined
+      : dtoNullableRecord(value.input_payload, 'input_payload')
+  const candidateCount =
+    inputPayload === undefined
       ? expectedCandidateCount
-      : validateInputPayload(
-          dtoNullableRecord(value.input_payload, 'input_payload'),
-          expectation,
-          expectedCandidateCount,
-        )
+      : validateInputPayload(inputPayload, expectation, expectedCandidateCount)
   const status = eventStatus(value, eventName)
   const result = value.result === undefined ? null : dtoNullableRecord(value.result, 'result')
   const error =
@@ -549,7 +566,13 @@ function mapEvent<TType extends GenerationType>(
     taskId: String(taskId),
     type: expectation.type,
     status,
-    result: mapResult(result, status, expectation, candidateCount),
+    result: mapResult(
+      result,
+      status,
+      expectation,
+      candidateCount,
+      inputPayload === undefined ? undefined : declaredFrameCount(inputPayload, expectation),
+    ),
     error,
   }
 }
@@ -619,7 +642,9 @@ export function createGenerationApis(config: GenerationApiConfig): GenerationApi
           ...(input.actionType === 'custom' ? { loop: input.loop ?? false } : {}),
           reference_video_url: null,
           reference_image_urls: referenceImageUrls,
-          num_frames: 32,
+          // 不发帧数：哪种动作出多少帧是后端按 action_type 定的约定，前端发一个数就是
+          // 第二份约定，两边分叉时任务照跑、帧照出，没有一处会红。
+
           // 后端拿 outfit_id 在场与否当三渲二的唯一判据（#122），所以它同时是"路线选择"
           // 本身，不只是一个标识。无条件发送会让建过 3D 资产的造型点"视频裁剪"也走三渲二，
           // 画风、成本、生成语义全被静默改掉，故只在用户真选了三渲二时发。
diff --git a/frontend/src/features/workflow-controller/README.md b/frontend/src/features/workflow-controller/README.md
index 84c75ab9..b17a1a70 100644
--- a/frontend/src/features/workflow-controller/README.md
+++ b/frontend/src/features/workflow-controller/README.md
@@ -34,7 +34,7 @@ async function generateCharacter() {
 - `entities/workflow-run` 定义纯数据和异步 CRUD，不包含推进方法。
 - Controller 根据 `dependsOnNodeIds` 解锁节点，允许同一依赖下的多个 Action 并行。
 - 新增 Action 一次创建动作首帧、动作生成方式、完整动画和审核四个 node，不会遗漏路线选择或用数组位置猜关系。
-- 动作首帧按项目的真实源方向分别调用图片 Generation，每个方向默认生成三张候选，也可按调用场景在 1–4 张间调整；用户为全部真实方向各确认一张后，完整动画节点才使用对应首帧调用 32 帧动作 Generation。可水平镜像的方向只保存关系，不创建重复任务。
+- 动作首帧按项目的真实源方向分别调用图片 Generation，每个方向默认生成三张候选，也可按调用场景在 1–4 张间调整；用户为全部真实方向各确认一张后，完整动画节点才使用对应首帧调用动作 Generation，帧数由后端按动作类型定，前端不传。可水平镜像的方向只保存关系，不创建重复任务。
 - Controller 方法与后端 Generation、WorkflowRun node 使用同一概念名：`characterTemplate`、`firstFrame`、`completeAnimation` 和 `review`，不再为同一概念保留另一套叫法。
 - 当前视频裁剪路线继续调用既有 Generation；3D 转 2D 选择会随 WorkflowRun 落库，但接口提供前明确阻止生成。
 - Generation 通过 `nodeId + taskId` 写回；节点重做后，旧任务的迟到结果会被丢弃。
diff --git a/frontend/src/features/workflow-controller/controller.test.ts b/frontend/src/features/workflow-controller/controller.test.ts
index 6704a548..c7583670 100644
--- a/frontend/src/features/workflow-controller/controller.test.ts
+++ b/frontend/src/features/workflow-controller/controller.test.ts
@@ -3049,7 +3049,7 @@ describe('WorkflowController', () => {
     expect(generation.apis.create).toHaveBeenCalledTimes(1)
   })
 
-  it('完整动画必须是 32 帧，完成后只解锁自己的审核节点', async () => {
+  it('完整动画完成后只解锁自己的审核节点', async () => {
     const run = createRun([
       ...completedCharacterNodes(),
       firstFrameNode({
@@ -3209,7 +3209,7 @@ describe('WorkflowController', () => {
     })
   })
 
-  it('完整动画节点拒绝帧数不足的结果', async () => {
+  it('完整动画节点拒绝没有帧的结果', async () => {
     const run = createRun([
       ...completedCharacterNodes(),
       firstFrameNode({ status: 'passed', phase: 'completed', selectedFirstFrameUrl: 'first.png' }),
@@ -3233,7 +3233,7 @@ describe('WorkflowController', () => {
         status: 'completed',
         result: {
           type: 'complete_animation',
-          frames: [{ index: 0, url: 'frame.png', durationMs: 80 }],
+          frames: [],
         },
         error: null,
       },
@@ -3241,7 +3241,7 @@ describe('WorkflowController', () => {
 
     expect(controller.getWorkflow().nodes[4]).toMatchObject({
       status: 'failed',
-      error: '完整动画应为 32 帧，实际为 1 帧',
+      error: '完整动画结果没有帧',
     })
   })
 
diff --git a/frontend/src/features/workflow-controller/controller.ts b/frontend/src/features/workflow-controller/controller.ts
index 9d998ef9..f0bbaedc 100644
--- a/frontend/src/features/workflow-controller/controller.ts
+++ b/frontend/src/features/workflow-controller/controller.ts
@@ -35,7 +35,6 @@ import {
   WorkflowRunConflictError,
 } from '@/entities'
 
-const COMPLETE_ANIMATION_FRAME_COUNT = 32
 const PROJECT_NAME_MAX_LENGTH = 20
 const QUICK_START_PROJECT_NAME_ATTEMPTS = 100
 
@@ -1894,9 +1893,9 @@ function generationResultError(node: WorkflowNode, generation: Generation): stri
     ) {
       return '完整动画结果格式无效'
     }
-    return generation.result.frames.length === COMPLETE_ANIMATION_FRAME_COUNT
-      ? null
-      : `完整动画应为 ${COMPLETE_ANIMATION_FRAME_COUNT} 帧，实际为 ${generation.result.frames.length} 帧`
+    // 帧数是否合乎该动作类型的约定，由 generation 适配器按任务声明的 num_frames 判；
+    // 这里再写一个数就是第二份约定，各动作帧数不同时它会把合规结果判成失败。
+    return generation.result.frames.length > 0 ? null : '完整动画结果没有帧'
   }
 
   return '当前节点不能绑定生成结果'
diff --git a/openapi.json b/openapi.json
index 053c9fbe..152b54a5 100644
--- a/openapi.json
+++ b/openapi.json
@@ -232,11 +232,17 @@
             "title": "Loop"
           },
           "num_frames": {
-            "default": 32,
-            "maximum": 64.0,
-            "minimum": 1.0,
-            "title": "Num Frames",
-            "type": "integer"
+            "anyOf": [
+              {
+                "maximum": 64.0,
+                "minimum": 1.0,
+                "type": "integer"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Num Frames"
           },
           "outfit_id": {
             "anyOf": [