feat(video): DashScope video specialist with correct i2v request bodies

- Add video_generation_client: async video-synthesis, t2v vs i2v (Wan 2.7 input.media first_frame vs legacy img_url).

- Coerce *-t2v* to *-i2v* when first frame present; AIA_VIDEO_I2V_INPUT_STYLE / I2V_MODEL overrides.

- Gateway/direct_loop early return for video specialist; workspace video + factory allowlist.

- Normalize WS and admin chat attachments (image_ref parity); mp4 attachment store roundtrip tests.

Co-authored-by: Cursor <cursoragent@cursor.com>

runtime/agents/factory.py

@@ -276,7 +276,7 @@ def build_ops_agent(
 
 
 # `default_registry` treats empty allow_tags + empty allow_tools as "no filter". Use an impossible
-# tool name so the image specialist gets an empty tool surface (vision-only turns).
+# tool name so image/video specialists get an empty tool surface (dedicated HTTP lanes).
 _IMAGE_SPECIALIST_TOOL_ALLOWLIST: tuple[str, ...] = ("__oclaw_image_specialist_no_tools__",)
 
 
@@ -334,7 +334,7 @@ def build_gateway_executor(
         "path_policy_user_id": path_policy_user_id,
         "store": store,
     }
-    if prof.name == "image":
+    if prof.name in {"image", "video"}:
         reg_kw["allow_tools"] = list(_IMAGE_SPECIALIST_TOOL_ALLOWLIST)
     tools = default_registry(**reg_kw)
     return Agent(

runtime/agents/specialist_agent.py

@@ -16,10 +16,17 @@ from oclaw.platform.persistence.sqlite_store import SqliteStore
 from oclaw.platform.llm.image_legacy_client import (
     IMAGE_SPECIALIST_DEFAULT_PROMPT_ZH,
     collect_legacy_lane_images_from_attachments,
+    collect_legacy_lane_images_with_session_fallback,
     legacy_image_assistant_body_with_placeholder,
     legacy_image_turn_bundle,
     send_legacy_image_messages,
 )
+from oclaw.platform.llm.video_generation_client import (
+    VIDEO_SPECIALIST_DEFAULT_PROMPT_ZH,
+    legacy_video_assistant_body_with_placeholder,
+    legacy_video_turn_bundle,
+    send_video_generation_request,
+)
 from oclaw.runtime.tools import default_registry
 from oclaw.runtime.agents.specialists import expert_name_for_specialist
 
@@ -301,6 +308,51 @@ class SpecialistAgentRunner:
                     tool_traces=(),
                     notes="image_pipeline",
                 )
+            elif step.specialist == "video":
+                image_protocol = "video_generation.http"
+                _, chosen_model, _ = self._resolve_profile_and_model(step.specialist)
+                text_parts = [
+                    str(x).strip()
+                    for x in (step.objective, step.input_text, parent_task.user_text)
+                    if str(x or "").strip()
+                ]
+                user_text_v = "\n".join(text_parts) if text_parts else VIDEO_SPECIALIST_DEFAULT_PROMPT_ZH
+                policy_sid = str(policy_session_id or parent_task.session_id or session_id or "").strip()
+                v_frames, _v_src = collect_legacy_lane_images_with_session_fallback(
+                    store=self.store,
+                    session_id=policy_sid,
+                    attachments=list(parent_task.attachments or []),
+                    max_images=1,
+                )
+                v_frame = str(v_frames[0]).strip() if v_frames else None
+                resp_v = send_video_generation_request(
+                    prompt=user_text_v,
+                    model=str(getattr(chosen_model, "model", "") or "").strip() or None,
+                    api_key=str(getattr(chosen_model, "api_key", "") or "").strip() or None,
+                    base_url=str(getattr(chosen_model, "base_url", "") or "").strip() or None,
+                    img_url=v_frame,
+                    on_progress=on_progress,
+                    should_stop=should_stop,
+                )
+                ok, output, produced_attachments = legacy_video_turn_bundle(resp_v)
+                output = legacy_video_assistant_body_with_placeholder(
+                    lang=self.lang,
+                    body_text=output,
+                    produced=produced_attachments if ok else None,
+                )
+                self.store.add_message(
+                    session_id=session_id,
+                    role="assistant",
+                    content=output,
+                    attachments=(produced_attachments or None) if ok else None,
+                )
+                specialist_delivery = SpecialistDelivery(
+                    specialist=step.specialist,
+                    step_id=step.step_id,
+                    answer_text=str(output or ""),
+                    tool_traces=(),
+                    notes="video_pipeline",
+                )
             else:
                 agent = self._build_agent_for(
                     step.specialist,

runtime/agents/specialists.py

@@ -43,10 +43,15 @@ SPECIALISTS: dict[SpecialistId, SpecialistConfig] = {
         expert_name="image",
         default_tool_tags=None,
     ),
+    "video": SpecialistConfig(
+        specialist_id="video",
+        expert_name="video",
+        default_tool_tags=None,
+    ),
 }
 
 def discover_specialist_ids() -> tuple[SpecialistId, ...]:
-    rows = specialist_registry_snapshot(base_order=("generalist", "ops", "memory", "image"))
+    rows = specialist_registry_snapshot(base_order=("generalist", "ops", "memory", "image", "video"))
     return tuple(str(x.get("id") or "").strip().lower() for x in rows if str(x.get("id") or "").strip())
 
 

runtime/direct_loop.py

@@ -1206,6 +1206,97 @@ def _maybe_image_specialist_legacy_gateway_turn(
     )
 
 
+def _maybe_video_specialist_legacy_gateway_turn(
+    *,
+    store: Any,
+    session_id: str,
+    turn_uuid: str,
+    lang: str,
+    model: ChatModel,
+    user_text: str,
+    attachments: list[dict[str, Any]] | None,
+    skill_binding_role: str | None,
+    on_token: Optional[Callable[[str], None]],
+    on_progress: Optional[Callable[[str], None]],
+    should_stop: Optional[Callable[[], bool]] = None,
+) -> TurnRunOutcome | None:
+    """When the UI selects **video** specialist, skip Responses/chat-model transports.
+
+    Uses DashScope async ``video-synthesis`` (see :mod:`oclaw.platform.llm.video_generation_client`).
+    With a user image (or session image fallback), sends ``input.img_url`` for **image-to-video**;
+    otherwise **text-to-video**. Disable with ``AIA_VIDEO_SPECIALIST_DISABLE_LEGACY_GATEWAY_LANE=1``.
+    """
+    if str(os.getenv("AIA_VIDEO_SPECIALIST_DISABLE_LEGACY_GATEWAY_LANE") or "").strip().lower() in (
+        "1",
+        "true",
+        "yes",
+        "on",
+    ):
+        return None
+    if str(skill_binding_role or "").strip().lower() != "video":
+        return None
+
+    from oclaw.platform.llm.image_legacy_client import collect_legacy_lane_images_with_session_fallback
+    from oclaw.platform.llm.video_generation_client import (
+        VIDEO_SPECIALIST_DEFAULT_PROMPT_ZH,
+        legacy_video_assistant_body_with_placeholder,
+        legacy_video_turn_bundle,
+        send_video_generation_request,
+    )
+
+    frames, frame_src = collect_legacy_lane_images_with_session_fallback(
+        store=store,
+        session_id=session_id,
+        attachments=attachments,
+        max_images=1,
+    )
+    frame_url = str(frames[0]).strip() if frames else None
+
+    if on_progress:
+        if frame_url:
+            if frame_src.endswith("_history"):
+                if str(lang or "").startswith("en"):
+                    on_progress("oclaw: reusing an earlier session image as first frame…")
+                else:
+                    on_progress("oclaw: 使用会话中较早的图片作为图生视频首帧…")
+            on_progress("oclaw: video specialist (DashScope image-to-video)…")
+        else:
+            on_progress("oclaw: video specialist (DashScope text-to-video)…")
+
+    prompt_plain = str(user_text or "").strip() or VIDEO_SPECIALIST_DEFAULT_PROMPT_ZH
+    resp = send_video_generation_request(
+        prompt=prompt_plain,
+        model=str(getattr(model, "model", "") or "").strip() or None,
+        api_key=str(getattr(model, "api_key", "") or "").strip() or None,
+        base_url=str(getattr(model, "base_url", "") or "").strip() or None,
+        img_url=frame_url,
+        on_progress=on_progress,
+        should_stop=should_stop,
+    )
+    ok, body_text, produced = legacy_video_turn_bundle(resp)
+    body_text = legacy_video_assistant_body_with_placeholder(
+        lang=lang,
+        body_text=body_text,
+        produced=produced if ok else None,
+    )
+    store.add_message(
+        session_id=session_id,
+        role="assistant",
+        content=body_text,
+        turn_uuid=turn_uuid,
+        event_type="assistant_text",
+        attachments=(produced or None) if ok else None,
+    )
+    if ok and on_token and body_text:
+        on_token(body_text)
+    return TurnRunOutcome(
+        final_text=body_text,
+        tool_traces=tuple(),
+        handoff_note="video_specialist_legacy_http" if ok else "video_specialist_legacy_upstream_failed",
+        turn_uuid=turn_uuid,
+    )
+
+
 def run_oclaw_direct_loop(
     *,
     store: Any,
@@ -1267,6 +1358,21 @@ def run_oclaw_direct_loop(
         )
         if legacy_early is not None:
             return legacy_early
+        video_early = _maybe_video_specialist_legacy_gateway_turn(
+            store=store,
+            session_id=session_id,
+            turn_uuid=turn_uuid,
+            lang=lang,
+            model=model,
+            user_text=str(user_text or ""),
+            attachments=attachments,
+            skill_binding_role=skill_binding_role,
+            on_token=on_token,
+            on_progress=on_progress,
+            should_stop=should_stop,
+        )
+        if video_early is not None:
+            return video_early
 
     skill_exec = SkillExecutor(config=ToolExecutionConfig(max_workers=max(1, min(int(max_tool_workers or 8), 32))))
     tool_traces: list[dict[str, Any]] = []

runtime/gateway.py

@@ -919,7 +919,7 @@ class OclawGateway:
                 attachments=list(msg.attachments or []),
                 metadata=dict(base_metadata),
             )
-            if manager_specialist in {"ops", "generalist", "image", "memory"}:
+            if manager_specialist in {"ops", "generalist", "image", "memory", "video"}:
                 if callable(specialist_executor_factory):
                     try:
                         selected_executor = specialist_executor_factory(manager_specialist)

runtime/skill_role_binding.py

@@ -17,7 +17,7 @@ def _truthy(v: str | None) -> bool:
 
 def ordered_specialist_ids() -> list[str]:
     base = [str(k).strip().lower() for k in discover_specialist_ids() if str(k).strip()]
-    preferred = [x for x in ("generalist", "ops", "memory", "image") if x in set(base)]
+    preferred = [x for x in ("generalist", "ops", "memory", "image", "video") if x in set(base)]
     return preferred + [x for x in base if x not in set(preferred)]
 
 

runtime/workspaces/experts.py

@@ -252,7 +252,7 @@ def build_expert_catalog_block(*, include_main: bool = False, per_field_limit: i
 
 def discover_specialist_ids_from_workspaces(
     *,
-    base_order: tuple[str, ...] = ("generalist", "ops", "memory", "image"),
+    base_order: tuple[str, ...] = ("generalist", "ops", "memory", "image", "video"),
 ) -> tuple[str, ...]:
     cache_key = (expert_workspace_signature_token(), tuple(str(x).strip().lower() for x in base_order if str(x).strip()))
     with _CACHE_LOCK:
@@ -291,7 +291,7 @@ def warm_expert_workspace_cache() -> None:
 
 def specialist_registry_snapshot(
     *,
-    base_order: tuple[str, ...] = ("generalist", "ops", "memory", "image"),
+    base_order: tuple[str, ...] = ("generalist", "ops", "memory", "image", "video"),
 ) -> tuple[dict[str, Any], ...]:
     """Single source of truth for runtime specialist discovery and metadata."""
     ordered = discover_specialist_ids_from_workspaces(base_order=base_order)

runtime/workspaces/video/META.json

@@ -0,0 +1,5 @@
+{
+  "display_name_en": "Video generation",
+  "display_name_zh": "视频生成专家",
+  "role": "expert"
+}

runtime/workspaces/video/ROLE_SYSTEM.md

@@ -0,0 +1 @@
+你是视频生成专家：将用户自然语言 prompt 交给 Wan / 百炼 text-to-video API，返回可下载或可播放的成片附件。不要编造已生成视频的 URL；仅展示接口真实返回结果或明确错误。

runtime/workspaces/video/SOUL.md

@@ -0,0 +1,9 @@
+你是**文生视频 / 图生视频**方向的专家助手：根据用户给出的画面与镜头描述（及可选的**首帧参考图**），调用百炼 / DashScope **异步视频合成**接口生成短视频结果，并将产出以会话附件（`video_ref`）形式返回。用户上传图片时，首帧会作为 `img_url` 提交（需使用支持图生视频的 i2v 模型）。
+
+回答要求：
+- 若用户描述含糊，可基于常识补全合理的镜头语言，但避免与用户明确约束相矛盾。
+- 生成失败时给出可读的上游错误或参数提示（如模型与区域、时长、分辨率不匹配）。
+
+边界：
+- 本专家链路**不调用**通用工具循环；仅走专用 HTTP 视频合成与轮询。
+- 不承诺具体成片内容符合版权素材或真人肖像等合规要求；用户需自行确保 prompt 合规。