feat(video): 接入 Agnes Video 2.5 与 Token Plan Flash

.env.example

@@ -49,6 +49,9 @@ AI_VIDEO_FALLBACKS=kling-v2-6
 # veo3.1 只对逗号分隔的这些用户 id 开放(组内做高质量素材用,不面向客户)。
 # 空 = 谁都不能用。它按秒计费且比 kling 贵,且**永远不进兜底链**,只能由请求显式指定。
 AI_VIDEO_VEO_USER_IDS=
+# Agnes Video 2.5 走独立国际站；密钥仅配置在服务端环境变量中。
+AI_VIDEO_AGNES_BASE_URL=https://apihub.agnes-ai.com/v1
+AI_VIDEO_AGNES_API_KEY=
 # 留空即按型号查上游牌价（美元）；配了它就整条链共用这一个价，跨型号时会算错。
 AI_IMAGE_UNIT_COST=
 AI_VIDEO_UNIT_COST_PER_SECOND=

backend/packages/framework/src/windup_framework/config/provider.py

@@ -49,6 +49,10 @@ class AIProviderSettings(BaseSettings):
     # 用途是组内做高质量素材,不面向客户。空值 = 没有任何人可用(默认)。
     # 逗号分隔的用户 id,如 "1,7,12"。
     video_veo_user_ids: str = ""
+    # Agnes 使用独立国际站凭证，不复用 Modelink 的全局 base_url/api_key。
+    # Key 只从部署环境读取；为空时 provider 会在上传首帧和发请求之前拒绝建单。
+    video_agnes_base_url: str = "https://apihub.agnes-ai.com/v1"
+    video_agnes_api_key: str = ""
 
     chat_fallbacks: str = ""
     image_fallbacks: str = "gemini-3.1-flash-image-preview"

backend/packages/framework/src/windup_framework/gateway/registry.py

@@ -12,6 +12,8 @@
     "kling-v2-6": Family.VIDEO_INPUT_REFERENCE,
     "kling-video-o1": Family.VIDEO_IMAGE_LIST,  # 登记但不允许进 chain
     "veo3.1": Family.VIDEO_FAL_QUEUE,  # 登记≠放行,还要在 USER_GATED_MODELS 的白名单里
+    "agnes-video-2.5": Family.VIDEO_AGNES,
+    "agnes-video-2.5-flash": Family.VIDEO_AGNES,
 }
 
 

backend/packages/framework/src/windup_framework/gateway/types.py

@@ -20,6 +20,7 @@ class Family(str, Enum):
     VIDEO_INPUT_REFERENCE = "video.input_reference"
     VIDEO_IMAGE_LIST = "video.image_list"
     VIDEO_FAL_QUEUE = "video.fal_queue"
+    VIDEO_AGNES = "video.agnes"
 
 
 #: 每个 scene 允许出现哪些 family。链上混不同 family 是合法的 —— 兜底型号与主型号
@@ -34,6 +35,7 @@ class Family(str, Enum):
     Scene.CHARACTER_ACTION: frozenset({
         Family.VIDEO_INPUT_REFERENCE,
         Family.VIDEO_FAL_QUEUE,
+        Family.VIDEO_AGNES,
     }),
 }
 

backend/packages/framework/src/windup_framework/gateway/video.py

@@ -29,7 +29,7 @@
     hash_bytes,
     hash_image_input,
 )
-from windup_framework.gateway.types import AdapterResult, NextStep, Scene
+from windup_framework.gateway.types import AdapterResult, Family, NextStep, Scene
 
 _DEFAULT_RETRY_AFTER_S = 2.0
 _SLEEP_CAP_S = 30.0
@@ -60,6 +60,20 @@ def __init__(self, registry, adapter, circuit, settings, route_adapters=None) ->
     def _adapter_for(self, route: GatewayRoute):
         return lookup_adapter(self._route_adapters, route, self._adapter)
 
+    def _open_circuit_scope(self, route: GatewayRoute, model: str) -> str | None:
+        """返回会阻止本次型号提交的熔断层；Agnes 不受 Modelink 共享层影响。"""
+        isolated = self._registry.family_of(model) is Family.VIDEO_AGNES
+        if not isolated:
+            if self._circuit.is_open("aggregator"):
+                return "aggregator"
+            if self._circuit.is_open("base_url:" + route.base_url_id):
+                return "base_url"
+            if self._circuit.is_open(key_circuit_id(route)):
+                return "key"
+        if self._circuit.is_open("model:" + model):
+            return "model"
+        return None
+
     def start_i2v(
         self,
         first_frame: bytes,
@@ -95,7 +109,9 @@ def poll_i2v(
             snap = adapter.inspect_job(job_id, model=model)
             if snap.ok and snap.job_status == "completed" and snap.edge_fingerprint:
                 if hasattr(adapter, "download_completed"):
-                    return adapter.download_completed(job_id, snap.edge_fingerprint)
+                    return adapter.download_completed(
+                        job_id, snap.edge_fingerprint, model=model
+                    )
             return snap
         return adapter.follow_job(job_id, model=model)
 
@@ -139,50 +155,13 @@ def fail(http_status: int | None) -> None:
                 f"http_status={http_status} error_type={err}"
             )
 
-        if self._circuit.is_open("aggregator"):
-            model = models[0] if models else ""
-            route = routes[0]
-            self._emit(
-                AttemptTrace(
-                    request_id=request_id,
-                    scene=Scene.CHARACTER_ACTION,
-                    model=model,
-                    route=route,
-                    attempt_index=seq.next_index(),
-                    retry_count=0,
-                    route_reason="skip_circuit_open",
-                    outcome="failed",
-                    circuit_scope="aggregator",
-                    total_latency_ms=total_ms(),
-                    maybe_billed=False,
-                    detail=AttemptDetail(
-                        input_hash=input_hash,
-                        policy_next_step="fail",
-                        upstream_reached="false",
-                    ),
-                )
-            )
-            fail(None)
-
         for route_index, route in enumerate(routes):
-            if self._circuit.is_open("base_url:" + route.base_url_id):
-                if route_index + 1 < len(routes):
-                    fallback_used = True
-                    route_reason_override = "base_url_unreached"
-                    continue
-                fail(last_http_status)
-            if self._circuit.is_open(key_circuit_id(route)):
-                if route_index + 1 < len(routes):
-                    fallback_used = True
-                    route_reason_override = "key_rate_limit"
-                    continue
-                fail(last_http_status)
-
             adapter = self._adapter_for(route)
             switch_to_next_route = False
             for i, model in enumerate(models):
                 model_index = start_i + i
-                if self._circuit.is_open("model:" + model):
+                open_scope = self._open_circuit_scope(route, model)
+                if open_scope is not None:
                     fallback_used = True
                     fallback_reason = "skip"
                     self._emit(
@@ -195,13 +174,15 @@ def fail(http_status: int | None) -> None:
                             retry_count=0,
                             route_reason="skip_circuit_open",
                             outcome="failed",
-                            circuit_scope="model",
+                            circuit_scope=open_scope,
                             fallback_used=fallback_used,
                             total_latency_ms=total_ms(),
                             maybe_billed=False,
                             detail=AttemptDetail(
                                 input_hash=input_hash,
-                                policy_next_step="fallback",
+                                policy_next_step=(
+                                    "fallback" if open_scope == "model" else "fail"
+                                ),
                                 upstream_reached="false",
                                 model_index=model_index,
                             ),
@@ -331,6 +312,17 @@ def fail(http_status: int | None) -> None:
                         has_job_id=has_job_id,
                     )
                     has_next_route = route_index + 1 < len(routes)
+                    has_next_model = i + 1 < len(models)
+                    uses_isolated_model_credentials = (
+                        self._registry.family_of(model) is Family.VIDEO_AGNES
+                    )
+                    if (
+                        uses_isolated_model_credentials
+                        and bound_job_id is None
+                        and has_next_model
+                        and step in (NextStep.OPEN_AGGREGATOR, NextStep.FALLBACK_KEY)
+                    ):
+                        step = NextStep.FALLBACK
                     if step is NextStep.FAIL and bound_job_id is None:
                         tier_step = budget.tier_b_escalation(
                             error_type,
@@ -418,13 +410,15 @@ def fail(http_status: int | None) -> None:
                             and error_type is not ModelErrorType.UPSTREAM_FAILED
                         ):
                             fail(last_http_status)
+                        allowed_pre_submit_fallbacks = {
+                            ModelErrorType.RATE_LIMIT,
+                            ModelErrorType.MODEL_NOT_FOUND,
+                        }
+                        if uses_isolated_model_credentials:
+                            allowed_pre_submit_fallbacks.add(ModelErrorType.UNREACHED)
                         if (
                             bound_job_id is None
-                            and error_type
-                            not in (
-                                ModelErrorType.RATE_LIMIT,
-                                ModelErrorType.MODEL_NOT_FOUND,
-                            )
+                            and error_type not in allowed_pre_submit_fallbacks
                         ):
                             fail(last_http_status)
                         fallback_used = True

backend/packages/framework/src/windup_framework/providers/protocol/__init__.py

@@ -1,3 +1,9 @@
+from .agnes_video import (
+    AGNES_VIDEO_25,
+    AGNES_VIDEO_25_FLASH,
+    AGNES_VIDEO_MODELS,
+    AgnesVideoProtocol,
+)
 from .fal_queue import (
     FAL_I2V_ENDPOINTS,
     FalQueueVideoProtocol,
@@ -15,6 +21,10 @@
 from .types import HttpCall, JobProtocol, VideoRequest
 
 __all__ = [
+    "AGNES_VIDEO_25",
+    "AGNES_VIDEO_25_FLASH",
+    "AGNES_VIDEO_MODELS",
+    "AgnesVideoProtocol",
     "FAL_I2V_ENDPOINTS",
     "FAL_IMAGE_ENDPOINTS",
     "IMAGE_LIST_MODELS",