完成极速成品和脚本优化

core/backend/apps/ai/video_digest.py

@@ -46,8 +46,8 @@ MIN_FRAMES = 8
 MAX_FRAMES = 36
 FRAME_WIDTH = 768
 FRAME_QUALITY = 3
-DIGEST_MAX_TOKENS = 8192
-_SHOT_MARK = re.compile(r"【第\s*\d+\s*镜】")
+DIGEST_MAX_TOKENS = 12288
+_SHOT_MARK = re.compile(r"【(?:镜头\s*\d+|第\s*\d+\s*镜)】")
 
 _FFMPEG_TIMEOUT = 60
 
@@ -93,9 +93,10 @@ def load_digest_skill() -> str:
         return main.read_text(encoding="utf-8")
     # 兜底:skill 丢了也别整条链路挂掉,退化成一句话提示词(产出会明显变差,交接文档已注明须带 skills 目录)
     return (
-        "你是分镜拆解 agent。输入是一条电商短视频的完整文件,含画面和口播。"
-        "必须覆盖全片,从 0 秒写到片尾,有几镜写几镜,不要概括成几大段。"
-        "每镜写画面和解说词,解说词按听到的口播逐字写。输出中文纯文本。"
+        "你是分镜拆解 agent。输入是一条短视频的完整文件,含画面和音轨。"
+        "必须覆盖全片,从 00:00 写到片尾,有几镜写几镜,不要概括成几大段。"
+        "每镜按【镜头 01】写出时间、时长、景别、机位、运镜、画面、人物动作、人物表情、"
+        "台词/旁白、音效、背景音乐、字幕、备注。没有就写无。输出中文纯文本。"
     )
 
 
@@ -130,6 +131,73 @@ def probe_duration(path: str | Path) -> float:
     return duration
 
 
+def probe_video_size(path: str | Path) -> tuple[int, int]:
+    """ffprobe 读画面宽高。读不到返回 0,0,不挡拆解。"""
+    try:
+        out = subprocess.run(
+            [
+                _binary("ffprobe"), "-v", "error",
+                "-select_streams", "v:0",
+                "-show_entries", "stream=width,height",
+                "-of", "csv=p=0:s=x",
+                str(path),
+            ],
+            capture_output=True, timeout=_FFMPEG_TIMEOUT, check=True,
+        ).stdout.decode("utf-8", errors="replace").strip()
+        width_s, height_s = out.split("x", 1)
+        return max(0, int(width_s)), max(0, int(height_s))
+    except Exception:  # noqa: BLE001
+        return 0, 0
+
+
+def extract_cover_jpeg(path: str | Path, duration: float) -> bytes:
+    """抽一帧作历史封面。失败返回空字节,不挡拆解。"""
+    try:
+        ffmpeg = _binary("ffmpeg")
+    except VideoDigestError:
+        return b""
+    at = max(0.0, min(float(duration) * 0.35, max(0.0, float(duration) - 0.15)))
+    try:
+        done = subprocess.run(
+            [
+                ffmpeg, "-v", "error", "-ss", f"{at:.2f}", "-i", str(path),
+                "-frames:v", "1", "-vf", "scale=640:-2",
+                "-q:v", "4", "-f", "image2", "-",
+            ],
+            capture_output=True, timeout=_FFMPEG_TIMEOUT, check=True,
+        )
+    except Exception:  # noqa: BLE001
+        return b""
+    return done.stdout or b""
+
+
+def ratio_label(width: int, height: int) -> str:
+    if not width or not height:
+        return ""
+    ratio = width / height
+    if abs(ratio - 9 / 16) < 0.08:
+        return "9:16 竖屏"
+    if abs(ratio - 16 / 9) < 0.08:
+        return "16:9 横屏"
+    if abs(ratio - 1) < 0.08:
+        return "1:1"
+    return "横屏" if width > height else "竖屏"
+
+
+def title_from_filename(name: str) -> str:
+    stem = Path(name or "").stem.strip()
+    return (stem or "参考视频")[:80]
+
+
+def duration_clock(seconds: float) -> str:
+    total = max(0, int(round(float(seconds or 0))))
+    return f"{total // 60:02d}:{total % 60:02d}"
+
+
+def shot_count(text: str) -> int:
+    return len(_SHOT_MARK.findall(text or ""))
+
+
 def plan_frame_times(duration: float) -> list[int]:
     """均匀采样时间点。取每段的**中点**,避开首尾黑场与片尾卡片。"""
     count = max(MIN_FRAMES, min(MAX_FRAMES, math.ceil(duration / SECONDS_PER_FRAME)))
@@ -223,6 +291,60 @@ def _compress_video(path: str) -> bytes | None:
         Path(out.name).unlink(missing_ok=True)
 
 
+def _is_browser_playable(path: str) -> bool:
+    """Chrome 播不了微信常见的 HEVC。H.264 + AAC/无音轨才直接存。"""
+    try:
+        out = subprocess.run(
+            [
+                _binary("ffprobe"), "-v", "error",
+                "-show_entries", "stream=codec_name,codec_type",
+                "-of", "json",
+                str(path),
+            ],
+            capture_output=True, timeout=_FFMPEG_TIMEOUT, check=True,
+        )
+        streams = json.loads(out.stdout).get("streams") or []
+    except Exception:  # noqa: BLE001
+        return False
+    video_ok = False
+    audio_ok = True
+    for stream in streams:
+        kind = stream.get("codec_type")
+        codec = str(stream.get("codec_name") or "").lower()
+        if kind == "video":
+            video_ok = codec in {"h264", "vp8", "vp9", "av1"}
+        elif kind == "audio":
+            audio_ok = codec in {"aac", "mp3", "opus", "vorbis"}
+    return video_ok and audio_ok
+
+
+def _prepare_browser_video(path: str, suffix: str) -> tuple[str, str, bool]:
+    """转成浏览器能播的 H.264 AAC。失败退回原片。第三项表示调用方要删临时文件。"""
+    ext = suffix if str(suffix).startswith(".") else f".{suffix or 'mp4'}"
+    if _is_browser_playable(path):
+        return path, ext, False
+    ffmpeg = _binary("ffmpeg")
+    out = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False)
+    out.close()
+    try:
+        done = subprocess.run(
+            [
+                ffmpeg, "-v", "error", "-y", "-i", path,
+                "-c:v", "libx264", "-preset", "veryfast", "-crf", "23",
+                "-c:a", "aac", "-b:a", "128k",
+                "-movflags", "+faststart", out.name,
+            ],
+            capture_output=True, timeout=_FFMPEG_TIMEOUT * 3,
+        )
+        if done.returncode != 0 or not Path(out.name).is_file() or Path(out.name).stat().st_size <= 0:
+            Path(out.name).unlink(missing_ok=True)
+            return path, ext, False
+        return out.name, ".mp4", True
+    except Exception:  # noqa: BLE001 — 转码失败仍存原片,总比历史卡不能播好
+        Path(out.name).unlink(missing_ok=True)
+        return path, ext, False
+
+
 def _native_video(path: str, size: int, suffix: str) -> DigestVideo | None:
     mime = _SUFFIX_MIME.get(suffix.lower(), "video/mp4")
     if size <= INLINE_VIDEO_MAX_BYTES:
@@ -233,17 +355,37 @@ def _native_video(path: str, size: int, suffix: str) -> DigestVideo | None:
     return None
 
 
-def digest_input_from_upload(upload) -> tuple[DigestVideo | None, list[VideoFrame], float]:
-    """优先整段视频(含音轨);塞不进请求才抽帧。"""
+def digest_input_from_upload(
+    upload,
+    *,
+    keep_source: bool = False,
+) -> tuple[DigestVideo | None, list[VideoFrame], float, dict]:
+    """优先整段视频(含音轨);塞不进请求才抽帧。顺带抽出历史卡要用的封面/宽高。
+
+    keep_source=True 时不删临时文件,调用方上传原片后再删。
+    """
     path = ""
+    keep = False
+    original_name = Path(getattr(upload, "name", "") or "参考视频.mp4").name or "参考视频.mp4"
     try:
         path, suffix, size, duration = _materialize_upload(upload)
+        width, height = probe_video_size(path)
+        cover_jpeg = extract_cover_jpeg(path, duration)
         video = _native_video(path, size, suffix)
-        if video is not None:
-            return video, [], duration
-        return None, extract_frames(path, plan_frame_times(duration)), duration
+        frames = [] if video is not None else extract_frames(path, plan_frame_times(duration))
+        extras = {
+            "file_name": original_name,
+            "file_size": size,
+            "width": width,
+            "height": height,
+            "cover_jpeg": cover_jpeg,
+            "suffix": suffix,
+            "source_path": path if keep_source else "",
+        }
+        keep = keep_source
+        return video, frames, duration, extras
     finally:
-        if path:
+        if path and not keep:
             Path(path).unlink(missing_ok=True)
 
 
@@ -267,20 +409,28 @@ def build_digest_messages(
     *,
     product_hint: str = "",
     video: DigestVideo | None = None,
+    aspect_ratio: str = "",
+    file_title: str = "",
 ) -> list[dict]:
     """system = 拆解 skill;user = 完整视频(优先)或抽帧。"""
     frames = frames or []
+    measured = f"实测时长约 {round(duration)} 秒"
+    if aspect_ratio:
+        measured += f",画面比例 {aspect_ratio}"
+    title_hint = f"文件名可参考片名:《{file_title}》。" if file_title else ""
     if video is not None:
         head = [
-            f"这是一条时长约 {round(duration)} 秒的电商带货短视频的**完整文件**(含画面和口播音轨)。",
-            "请按技能还原全片分镜:从 0 秒写到片尾,有几镜写几镜。",
-            "解说词按听到的口播逐字写;画面上的花字一并写进画面。",
+            f"这是一条{measured}的短视频**完整文件**(含画面和口播音轨)。{title_hint}",
+            "请按技能还原全片导演分镜稿:从 00:00 写到片尾,有几镜写几镜。",
+            "每镜必须写齐时间、时长、景别、机位、运镜、画面、人物动作、人物表情、台词/旁白、音效、背景音乐、字幕、备注。",
+            "台词/旁白按听到的口播逐字写;画面上的花字写进字幕。",
         ]
     else:
         head = [
-            f"这是一条时长约 {round(duration)} 秒的电商带货短视频,",
-            f"按时间顺序均匀抽了 {len(frames)} 帧。每帧图前面标了它在原片中的时间点。",
-            "请按技能还原**全片**分镜:从 0 秒写到片尾,有几镜写几镜。",
+            f"这是一条{measured}的短视频,",
+            f"按时间顺序均匀抽了 {len(frames)} 帧。每帧图前面标了它在原片中的时间点。{title_hint}",
+            "请按技能还原**全片**导演分镜稿:从 00:00 写到片尾,有几镜写几镜。",
+            "每镜必须写齐时间、时长、景别、机位、运镜、画面、人物动作、人物表情、台词/旁白、音效、背景音乐、字幕、备注。",
         ]
     if product_hint:
         head.append(f"用户接下来想用这条片子的结构去拍自己的商品:{product_hint}。")
@@ -429,6 +579,127 @@ def digest_team_video(*, team, user, upload, model_config_id=None) -> dict:
     )
 
 
+def _store_digest_cover(*, team, jpeg: bytes) -> tuple[str, str]:
+    """封面传到 TOS。失败返回空,历史卡走占位底。"""
+    from io import BytesIO
+
+    from apps.assets.storage import TosStorage
+
+    if not jpeg:
+        return "", ""
+    key = f"teams/{team.id}/video-digest/{uuid.uuid4()}.jpg"
+    storage = TosStorage()
+    stored = storage.upload_fileobj(fileobj=BytesIO(jpeg), object_key=key, content_type="image/jpeg")
+    return stored.object_key, storage.public_url(object_key=stored.object_key)
+
+
+def _store_digest_video(*, team, path: str, suffix: str) -> tuple[str, str]:
+    """原片(优先转成 H.264)传到 TOS,历史封面点击才能播。失败返回空,不挡拆解。"""
+    from apps.assets.storage import TosStorage
+
+    if not path or not Path(path).is_file():
+        return "", ""
+    upload_path, ext, ephemeral = _prepare_browser_video(path, suffix)
+    try:
+        if ext.lower() not in ALLOWED_SUFFIXES:
+            ext = ".mp4"
+        mime = _SUFFIX_MIME.get(ext.lower(), "video/mp4")
+        key = f"teams/{team.id}/video-digest/{uuid.uuid4()}{ext.lower()}"
+        storage = TosStorage()
+        with Path(upload_path).open("rb") as fileobj:
+            stored = storage.upload_fileobj(fileobj=fileobj, object_key=key, content_type=mime)
+        return stored.object_key, storage.public_url(object_key=stored.object_key)
+    finally:
+        if ephemeral:
+            Path(upload_path).unlink(missing_ok=True)
+
+
+def _media_url_from_payload(req: dict, url_key: str, object_key_name: str, *, signed: bool = False) -> str:
+    stored = str(req.get(url_key) or "").strip()
+    key = str(req.get(object_key_name) or "").strip()
+    if not key:
+        return stored
+    try:
+        from apps.assets.storage import TosStorage
+
+        storage = TosStorage()
+        if signed:
+            return storage.presigned_get_url(object_key=key, expires_in=6 * 3600)
+        return storage.public_url(object_key=key)
+    except Exception:  # noqa: BLE001
+        return stored
+
+
+def serialize_digest_history(task) -> dict:
+    from django.utils import timezone
+
+    from apps.ai.models import AITask
+
+    req = task.request_payload or {}
+    resp = task.response_payload or {}
+    prompt = str(resp.get("prompt") or resp.get("digest") or "").strip()
+    duration = float(req.get("duration_seconds") or 0)
+    file_name = str(req.get("file_name") or "") or "参考视频.mp4"
+    title = str(req.get("title") or "").strip() or title_from_filename(file_name)
+    shots = int(req.get("shot_count") or 0) or shot_count(prompt)
+    created = timezone.localtime(task.created_at) if task.created_at else timezone.now()
+    return {
+        "id": str(task.id),
+        "title": title,
+        "status": "已完成" if task.status == AITask.Status.SUCCEEDED else "失败",
+        "duration": int(round(duration)),
+        "duration_label": duration_clock(duration),
+        "ratio": str(req.get("ratio") or ""),
+        "shots": shots,
+        "file_name": file_name,
+        "cover_url": _media_url_from_payload(req, "cover_url", "cover_key"),
+        "video_url": _media_url_from_payload(req, "video_url", "video_key", signed=True),
+        "prompt": prompt,
+        "created_at": task.created_at.isoformat() if task.created_at else "",
+        "created_date": created.strftime("%Y-%m-%d"),
+    }
+
+
+def list_team_digest_history(*, team, limit: int = 50) -> list[dict]:
+    from apps.ai.models import AITask
+
+    qs = (
+        AITask.objects.filter(
+            team=team,
+            task_type=AITask.Type.VIDEO_DIGEST,
+            project__isnull=True,
+            status=AITask.Status.SUCCEEDED,
+            is_deleted=False,
+            purged_at__isnull=True,
+        )
+        .order_by("-created_at")[: max(1, min(int(limit), 50))]
+    )
+    return [serialize_digest_history(task) for task in qs]
+
+
+def save_digest_prompt(*, team, task_id, prompt: str) -> dict | None:
+    from apps.ai.models import AITask
+
+    cleaned = (prompt or "").strip()
+    if not cleaned:
+        raise VideoDigestError("提示词不能为空")
+    task = AITask.objects.filter(
+        id=task_id,
+        team=team,
+        task_type=AITask.Type.VIDEO_DIGEST,
+        project__isnull=True,
+        is_deleted=False,
+        purged_at__isnull=True,
+    ).first()
+    if task is None:
+        return None
+    payload = dict(task.response_payload or {})
+    payload["prompt"] = cleaned[:32000]
+    task.response_payload = payload
+    task.save(update_fields=["response_payload", "updated_at"])
+    return serialize_digest_history(task)
+
+
 def _digest_video(*, team, user, upload, project=None, product_hint="", model_config_id=None) -> dict:
     """上传视频 → 分镜稿。抽帧在建任务之前做,文件不合格不占积分。"""
     from django.db import transaction
@@ -439,10 +710,20 @@ def _digest_video(*, team, user, upload, project=None, product_hint="", model_co
     from apps.billing.pricing import quote_video_digest
     from apps.billing.services.ledger import charge_reserved_credit, reserve_credit
 
-    video, frames, duration = digest_input_from_upload(upload)
+    video, frames, duration, extras = digest_input_from_upload(
+        upload, keep_source=(project is None)
+    )
+    extras = extras or {}
+    source_path = str(extras.get("source_path") or "")
+    file_name = str(extras.get("file_name") or getattr(upload, "name", "") or "参考视频.mp4")
+    width = int(extras.get("width") or 0)
+    height = int(extras.get("height") or 0)
+    file_size = int(extras.get("file_size") or getattr(upload, "size", 0) or 0)
 
     model_config = resolve_digest_model_config(preferred_id=model_config_id)
     if model_config is None:
+        if source_path:
+            Path(source_path).unlink(missing_ok=True)
         raise VideoDigestError("视频提炼需要 Gemini 3.1 Pro(会看图),当前没有启用,请联系管理员")
     logger.info(
         "video digest using %s:%s (%s) input=%s duration=%.1fs bytes=%s frames=%s",
@@ -456,7 +737,12 @@ def _digest_video(*, team, user, upload, project=None, product_hint="", model_co
     )
 
     messages = build_digest_messages(
-        frames, duration, product_hint=product_hint, video=video
+        frames,
+        duration,
+        product_hint=product_hint,
+        video=video,
+        aspect_ratio=ratio_label(width, height),
+        file_title=title_from_filename(file_name),
     )
     request_payload = {
         "model": model_config.name,
@@ -467,6 +753,12 @@ def _digest_video(*, team, user, upload, project=None, product_hint="", model_co
         "frame_count": len(frames),
         "video_bytes": len(video.data) if video is not None else 0,
         "frame_times": [f.at_seconds for f in frames],
+        "file_name": file_name,
+        "title": title_from_filename(file_name),
+        "file_size": file_size,
+        "width": width,
+        "height": height,
+        "ratio": ratio_label(width, height),
     }
 
     quote = quote_video_digest(team=team, model_config=model_config)
@@ -502,6 +794,8 @@ def _digest_video(*, team, user, upload, project=None, product_hint="", model_co
                 task.status = AITask.Status.RESERVED
                 task.save(update_fields=["status", "updated_at"])
         except ValueError as exc:
+            if source_path:
+                Path(source_path).unlink(missing_ok=True)
             if "insufficient credit" in str(exc).lower():
                 raise VideoDigestError("团队余额不足，请充值后重试") from exc
             raise VideoDigestError(str(exc)) from exc
@@ -532,15 +826,56 @@ def _digest_video(*, team, user, upload, project=None, product_hint="", model_co
         )
         _text, _response, digest = routed.value
     except Exception as exc:  # noqa: BLE001
+        if source_path:
+            Path(source_path).unlink(missing_ok=True)
         _fail_digest_task(task, reservation, str(exc))
         raise
 
+    cover_key, cover_url, video_key, video_url = "", "", "", ""
+    if project is None:
+        try:
+            cover_key, cover_url = _store_digest_cover(team=team, jpeg=extras.get("cover_jpeg") or b"")
+        except Exception:  # noqa: BLE001 — 封面失败不挡拆解结果
+            logger.warning("video digest cover upload failed", exc_info=True)
+        try:
+            video_key, video_url = _store_digest_video(
+                team=team,
+                path=source_path,
+                suffix=str(extras.get("suffix") or ".mp4"),
+            )
+        except Exception:  # noqa: BLE001 — 原片失败仍可看提示词,封面不能播
+            logger.warning("video digest source upload failed", exc_info=True)
+        if source_path:
+            Path(source_path).unlink(missing_ok=True)
+            source_path = ""
+
+    shots = shot_count(digest)
+    request_payload = {
+        **(task.request_payload or {}),
+        **request_payload,
+        "shot_count": shots,
+        "cover_key": cover_key,
+        "cover_url": cover_url,
+        "video_key": video_key,
+        "video_url": video_url,
+    }
+
     with transaction.atomic():
         task.status = AITask.Status.SUCCEEDED
-        task.response_payload = {"digest": digest[:32000]}
+        task.request_payload = request_payload
+        task.response_payload = {"digest": digest[:32000], "prompt": digest[:32000]}
         task.actual_cost = task.estimated_cost
         task.completed_at = timezone.now()
-        task.save(update_fields=["status", "response_payload", "actual_cost", "completed_at", "updated_at"])
+        task.save(
+            update_fields=[
+                "status",
+                "request_payload",
+                "response_payload",
+                "actual_cost",
+                "completed_at",
+                "updated_at",
+            ]
+        )
         charge_reserved_credit(reservation=reservation, actual_amount=task.actual_cost)
 
     return {
@@ -551,6 +886,14 @@ def _digest_video(*, team, user, upload, project=None, product_hint="", model_co
         "duration": round(duration, 1),
         "task_id": str(task.id),
         "estimated_cost": str(task.estimated_cost),
+        "title": request_payload.get("title") or title_from_filename(file_name),
+        "file_name": file_name,
+        "ratio": request_payload.get("ratio") or "",
+        "shots": shots,
+        "cover_url": cover_url,
+        "video_url": video_url,
+        "width": width,
+        "height": height,
     }