@oh-story/dsh 0.1.3 → 0.1.5

package/lib/video-recap/skills/video-understanding/scripts/vlm.py

@@ -0,0 +1,705 @@
+import base64
+import json
+import mimetypes
+import re
+from collections import OrderedDict
+from pathlib import Path
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from threading import Lock
+
+from extract import (
+    FRAME_TIME_CONVENTION_VERSION,
+    frame_time,
+    parse_frame_number,
+)
+from lib import CONFIG
+from lib import log, api_call, load_prompt, mimo_video_api_call, run_cmd, file_fingerprint, stable_hash
+
+# ── Step 4: VLM 视觉分析 ─────────────────────────────────────────────
+
+def _parse_vlm_depth_response(raw_text):
+    """解析 VLM 深度分析响应，提取【描述】、【帧标签】和【深层分析】"""
+    if not raw_text or not raw_text.strip():
+        return "(VLM 无法识别此场景画面)", "", {}
+
+    # 提取【描述】
+    desc_match = re.search(r'【描述】\s*\n?(.*?)(?=【帧标签】|【深层分析】|$)', raw_text, re.DOTALL)
+    if desc_match:
+        description = desc_match.group(1).strip()
+    else:
+        description = raw_text.strip()
+
+    # 提取【帧标签】
+    frame_facts = {}
+    facts_match = re.search(r'【帧标签】\s*\n?(.*?)(?=【深层分析】|$)', raw_text, re.DOTALL)
+    if facts_match:
+        for line in facts_match.group(1).strip().split("\n"):
+            line = line.strip()
+            if not line:
+                continue
+            # 格式: "12.0s | 男子拿起茶壶对嘴喝, 满脸疲惫"
+            m = re.match(r'([\d.]+)\s*s?\s*\|\s*(.+)', line)
+            if m:
+                ts = m.group(1)
+                actions = [a.strip() for a in re.split(r"[，,；;、]+", m.group(2)) if a.strip()]
+                if actions:
+                    frame_facts[ts] = actions
+
+    # 提取【深层分析】
+    depth_match = re.search(r'【深层分析】\s*\n?(.*?)$', raw_text, re.DOTALL)
+    depth_analysis = depth_match.group(1).strip() if depth_match else ""
+
+    if not description:
+        description = "(VLM 无法识别此场景画面)"
+
+    return description, depth_analysis, frame_facts
+
+
+def _max_frames_for_duration(duration):
+    """Frames the VLM sees for one scene: ~1 per `vlm_seconds_per_frame`, floor 3, capped by
+    `vlm_max_frames`. Replaces the old hard cap of 6 that starved long/merged scenes."""
+    spf = float(CONFIG.get("vlm_seconds_per_frame", 4.0) or 4.0)
+    ceiling = int(CONFIG.get("vlm_max_frames", 16) or 16)
+    return max(3, min(ceiling, round(max(0.0, float(duration)) / spf)))
+
+
+def _vlm_scene_cache_path(work_dir):
+    return Path(work_dir) / "vlm_scene_cache.json"
+
+
+def _load_vlm_scene_cache(work_dir):
+    """Per-scene VLM resume cache (scene_key -> analysis). Tolerant: {} if absent/corrupt."""
+    path = _vlm_scene_cache_path(work_dir)
+    if not path.exists():
+        return {}
+    try:
+        data = json.loads(path.read_text(encoding="utf-8"))
+        return data if isinstance(data, dict) else {}
+    except (OSError, ValueError):
+        return {}
+
+
+def _flush_vlm_scene_cache(work_dir, cache):
+    """Persist the resume cache atomically (temp + rename) so an abort/crash never corrupts it."""
+    path = _vlm_scene_cache_path(work_dir)
+    try:
+        tmp = path.with_name(path.name + ".tmp")
+        tmp.write_text(json.dumps(cache, ensure_ascii=False), encoding="utf-8")
+        tmp.replace(path)
+    except OSError as exc:
+        log(f"VLM 场景缓存写入失败（忽略）: {exc}")
+
+
+def _looks_rate_limited(message):
+    """Heuristic: was a scene failure a transient rate-limit (worth a low-concurrency retry) rather
+    than a persistent error (empty response / parse failure) that re-running would not fix?"""
+    m = str(message).lower()
+    return "429" in m or "too many requests" in m or "rate limit" in m or "限流" in m
+
+
+def analyze_scenes(scenes, frames, work_dir, *, resume=True):
+    """对每个场景的关键帧调用 VLM 进行视觉分析（并行）"""
+    if not scenes:
+        analyses = []
+        vlm_file = work_dir / "vlm_analysis.json"
+        vlm_file.write_text(json.dumps(analyses, ensure_ascii=False, indent=2), encoding="utf-8")
+        log("VLM 分析完成: 0 个场景")
+        return analyses
+
+    if not frames:
+        raise RuntimeError("VLM 分析需要先提取至少一帧；frames 为空")
+
+    fps = CONFIG["fps"]
+    if fps <= 0:
+        raise ValueError("CONFIG['fps'] 必须大于 0；请先运行完整 pipeline 或指定 --fps")
+
+    vlm_prompt = load_prompt("VLM_DEPTH_PROMPT")
+    if not vlm_prompt:
+        vlm_prompt = "仔细观察这些视频帧。分两部分输出：\n【描述】不超过80字，描述画面中正在发生什么。\n【深层分析】不超过120字，分析角色情绪、关系动态、潜台词。"
+
+    ctx = CONFIG.get("context_info", "")
+    if ctx:
+        vlm_prompt = f"已知信息：{ctx}\n\n{vlm_prompt}"
+
+    # 构建帧时间映射 (frame_NNNNN.jpg -> time in seconds)。换算规则由 extract.py 独家定义：
+    # ffmpeg 首帧在 t=0 而文件编号从 1 起，所以是 (n-1)/fps，不是 n/fps。
+    frame_times = {}
+    for f in frames:
+        num = parse_frame_number(f)
+        if num is None:
+            continue
+        frame_times[f] = frame_time(num, fps)
+
+    # base64 编码缓存（LRU）。原来是无上界的 dict：整个 VLM 阶段会把用到的每一帧的 base64
+    # 都常驻内存，而 base64 比原图还大 1/3。40 分钟视频 fps=1 约 2400 帧 → 数百 MB 常驻。
+    # 相邻场景才会复用同一帧，容量取「并发数 × 单场景最大帧数」就够，再多也命中不了。
+    b64_capacity = max(
+        8,
+        int(CONFIG.get("vlm_max_frames", 16) or 16) * max(1, int(CONFIG.get("vlm_workers", 4) or 4)),
+    )
+    b64_cache = OrderedDict()
+    b64_lock = Lock()
+
+    def _get_b64(frame_path):
+        with b64_lock:
+            cached = b64_cache.get(frame_path)
+            if cached is not None:
+                b64_cache.move_to_end(frame_path)
+                return cached
+        encoded = base64.b64encode(frame_path.read_bytes()).decode()
+        with b64_lock:
+            b64_cache[frame_path] = encoded
+            b64_cache.move_to_end(frame_path)
+            while len(b64_cache) > b64_capacity:
+                b64_cache.popitem(last=False)
+        return encoded
+
+    def _analyze_single_scene(i, scene):
+        """分析单个场景，返回 (scene_id, result_dict)"""
+        scene_frames = [f for f, t in frame_times.items()
+                        if scene["start"] <= t <= scene["end"]]
+        if not scene_frames:
+            mid = (scene["start"] + scene["end"]) / 2
+            scene_frames = [min(frames, key=lambda f: abs(frame_times.get(f, 999) - mid))]
+
+        duration = scene["end"] - scene["start"]
+        max_frames = _max_frames_for_duration(duration)
+
+        if len(scene_frames) > max_frames:
+            step = len(scene_frames) / max_frames
+            scene_frames = [scene_frames[int(j * step)] for j in range(max_frames)]
+        else:
+            scene_frames = scene_frames[:max_frames]
+
+        content_parts = []
+        for f in scene_frames:
+            b64 = _get_b64(f)
+            content_parts.append({
+                "type": "image_url",
+                "image_url": {"url": f"data:image/jpeg;base64,{b64}"}
+            })
+
+        # 帧级事实标签：将帧时间注入prompt
+        frame_ts_list = [f"{frame_times[f]:.1f}s" for f in scene_frames]
+        frame_ts_text = "帧时间点: " + ", ".join(frame_ts_list)
+        content_parts.append({"type": "text", "text": frame_ts_text + "\n\n" + vlm_prompt})
+
+        payload = {
+            "model": CONFIG["vlm_model"],
+            "messages": [{"role": "user", "content": content_parts}],
+            "max_tokens": int(CONFIG.get("vlm_max_tokens", 1500) or 1500),
+        }
+
+        log(f"VLM 分析场景 {i+1}/{len(scenes)} ({len(scene_frames)} 帧)...")
+
+        raw_response = ""
+        for attempt in range(3):
+            resp = api_call(payload)
+            try:
+                msg = resp["choices"][0]["message"]
+                raw_response = (msg.get("content") or msg.get("reasoning_content") or "")
+            except (KeyError, IndexError):
+                log(f"VLM 返回异常: {json.dumps(resp, ensure_ascii=False)[:200]}")
+
+            if raw_response.strip():
+                break
+
+            if attempt < 2:
+                log(f"  场景 {i+1} VLM 返回空，重试 ({attempt+2}/3)...")
+                retry_parts = content_parts[:-1]
+                retry_parts.append({"type": "text", "text": frame_ts_text + "\n\n" + vlm_prompt + "\n请务必按格式输出，不要留空。"})
+                payload = {
+                    "model": CONFIG["vlm_model"],
+                    "messages": [{"role": "user", "content": retry_parts}],
+                    "max_tokens": int(CONFIG.get("vlm_max_tokens", 1500) or 1500),
+                }
+
+        if not raw_response.strip():
+            raise RuntimeError("VLM 连续 3 次返回空内容")
+
+        # 解析 【描述】、【帧标签】和【深层分析】
+        description, depth_analysis, frame_facts = _parse_vlm_depth_response(raw_response)
+
+        result = {
+            "scene_id": i,
+            "start": scene["start"],
+            "end": scene["end"],
+            "description": description,
+            "depth_analysis": depth_analysis,
+        }
+        if frame_facts:
+            result["frame_facts"] = frame_facts
+
+        return i, result
+
+    # 断点续传：逐场景持久化分析结果，避免少数场景失败（多为 429 限流）时整批画面理解全部作废。
+    cache = _load_vlm_scene_cache(work_dir) if resume else {}
+    prompt_fp = stable_hash(vlm_prompt)
+
+    def _scene_cache_key(i, scene):
+        # Must invalidate on the SAME output-affecting settings the outer stage gate tracks
+        # (understand.py:_vlm_cache_payload). prompt_fp already covers vlm_prompt + context_info
+        # (which folds in background_research). The three below change the request/endpoint but
+        # are NOT in vlm_prompt, so they must be keyed explicitly — otherwise a partial-failure
+        # cache reused after flipping e.g. mimo_disable_thinking yields a stale/mixed analysis.
+        return "|".join(str(x) for x in (
+            i, round(float(scene["start"]), 3), round(float(scene["end"]), 3),
+            CONFIG.get("vlm_model"), prompt_fp, CONFIG.get("vlm_max_tokens"),
+            CONFIG.get("vlm_seconds_per_frame"), CONFIG.get("vlm_max_frames"),
+            round(float(fps), 3), FRAME_TIME_CONVENTION_VERSION,
+            CONFIG.get("api_url"), CONFIG.get("mimo_disable_thinking", True),
+            CONFIG.get("mimo_media_resolution"),
+        ))
+
+    analyses = [None] * len(scenes)
+    todo = []
+    for i, s in enumerate(scenes):
+        cached = cache.get(_scene_cache_key(i, s))
+        analyses[i] = cached if isinstance(cached, dict) else None
+        if analyses[i] is None:
+            todo.append(i)
+    if todo and len(todo) < len(scenes):
+        log(f"VLM 复用 {len(scenes) - len(todo)} 个已缓存场景，待分析 {len(todo)} 个")
+
+    def _run_pass(indices, workers):
+        """分析给定场景索引；结果按批持久化到续传缓存（在主线程，无需加锁）。返回 (i, err) 失败列表。
+
+        续传缓存每次都是整份重写，原来每完成一个场景就刷一次 → 写入量随场景数平方增长
+        （300 个场景约 75MB 冗余写），而且序列化发生在收结果的主循环上。改成按批刷：
+        崩溃最多丢失不到一批的进度，而这些场景本来就会在重跑时被重新分析。
+        """
+        if not indices:
+            return []
+        failures = []
+        pending = 0
+        # 与并发度对齐：一批大致就是「同时在飞的那几个场景」。
+        flush_every = max(1, min(16, workers))
+        with ThreadPoolExecutor(max_workers=max(1, workers)) as executor:
+            futures = {executor.submit(_analyze_single_scene, i, scenes[i]): i for i in indices}
+            try:
+                for future in as_completed(futures):
+                    i = futures[future]
+                    try:
+                        idx, result = future.result()
+                        analyses[idx] = result
+                        cache[_scene_cache_key(idx, scenes[idx])] = result
+                        pending += 1
+                        if pending >= flush_every:
+                            _flush_vlm_scene_cache(work_dir, cache)
+                            pending = 0
+                    except Exception as e:  # noqa: BLE001 - 单个场景失败不能拖垮其余已完成的
+                        log(f"VLM 场景 {i+1} 分析失败: {e}")
+                        failures.append((i, str(e)))
+            finally:
+                # 无论正常结束、失败还是中断，都把这一轮已完成的场景落盘。
+                if pending:
+                    _flush_vlm_scene_cache(work_dir, cache)
+        return failures
+
+    base_workers = min(len(todo) or 1, int(CONFIG.get("vlm_workers", 4) or 4))
+    log(f"VLM 并行分析 {len(todo)} 个场景 (workers={base_workers})...")
+    failures = _run_pass(todo, base_workers)
+    if failures:
+        # 限流(429)多是并发瞬时拥塞，降并发重试一轮（已成功的从缓存跳过）；其它错误（空响应/解析失败）
+        # 重试无益，直接保留为失败，不浪费一轮调用。
+        rate_limited = [i for i, msg in failures if _looks_rate_limited(msg)]
+        persistent = [(i, msg) for i, msg in failures if not _looks_rate_limited(msg)]
+        if rate_limited:
+            retry_workers = max(1, base_workers // 4)
+            log(f"VLM {len(rate_limited)} 个场景疑似限流(429)，降并发到 {retry_workers} 重试...")
+            persistent += _run_pass(rate_limited, retry_workers)
+        failures = persistent
+
+    if failures:
+        sample = "; ".join(f"场景 {i+1}: {msg}" for i, msg in failures[:3])
+        raise RuntimeError(
+            f"VLM 分析失败 {len(failures)}/{len(scenes)} 个场景。其余已缓存到 vlm_scene_cache.json，"
+            f"重跑可断点续传（只重试失败场景）。示例: {sample}"
+        )
+
+    vlm_file = work_dir / "vlm_analysis.json"
+    vlm_file.write_text(json.dumps(analyses, ensure_ascii=False, indent=2), encoding="utf-8")
+    _vlm_scene_cache_path(work_dir).unlink(missing_ok=True)  # 完整理解已生成，续传缓存可清理
+    log(f"VLM 分析完成: {len(analyses)} 个场景")
+    return analyses
+
+
+def _video_data_url(video_path):
+    """Return a MiMo-compatible data URL for a local video chunk, or None when too large."""
+    max_bytes = int(float(CONFIG.get("mimo_video_base64_max_mb", 45.0)) * 1024 * 1024)
+    encoded_size = int(video_path.stat().st_size * 4 / 3) + 128
+    if encoded_size > max_bytes:
+        log(
+            "MiMo 视频分片超过 base64 上限: "
+            f"编码后约 {encoded_size / 1024 / 1024:.1f}MB，超过限制 {max_bytes / 1024 / 1024:.1f}MB；"
+            "请降低 MIMO_VIDEO_CHUNK_MAX_SECONDS 或 MIMO_VIDEO_FPS"
+        )
+        return None
+    mime_type = mimetypes.guess_type(str(video_path))[0] or "video/mp4"
+    encoded = base64.b64encode(video_path.read_bytes()).decode("ascii")
+    return f"data:{mime_type};base64,{encoded}"
+
+
+def _mimo_video_chunks(scenes):
+    """Build local MiMo video-understanding chunks from ffmpeg scene boundaries."""
+    if not scenes:
+        raise RuntimeError("MiMo 视频分片理解需要 scenes；请先运行 ffmpeg scene/scdet 场景检测")
+
+    max_seconds = float(CONFIG.get("mimo_video_chunk_max_seconds", 20.0) or 20.0)
+    min_seconds = float(CONFIG.get("mimo_video_chunk_min_seconds", 1.0) or 1.0)
+    chunks = []
+    for scene_index, scene in enumerate(scenes):
+        try:
+            start = float(scene.get("start", 0.0))
+            end = float(scene.get("end", start))
+        except (TypeError, ValueError, AttributeError):
+            continue
+        if end <= start:
+            continue
+        scene_id = scene.get("scene_id", scene_index) if isinstance(scene, dict) else scene_index
+        cursor = start
+        while cursor < end:
+            chunk_end = min(end, cursor + max_seconds)
+            if end - chunk_end < min_seconds and chunk_end < end:
+                chunk_end = end
+            if chunk_end > cursor:
+                chunks.append({
+                    "chunk_id": len(chunks),
+                    "scene_id": scene_id,
+                    "start": round(cursor, 3),
+                    "end": round(chunk_end, 3),
+                })
+            cursor = chunk_end
+    if not chunks:
+        raise RuntimeError("MiMo 视频分片理解没有可用分片；请检查 scenes.json")
+    return chunks
+
+
+def _extract_video_chunk(video_path, chunk, output_path):
+    """Cut one scene-based chunk into a compact local MP4 for MiMo video_url data URL."""
+    start = float(chunk["start"])
+    duration = max(0.1, float(chunk["end"]) - start)
+    fps = float(CONFIG.get("mimo_video_fps", 2.0) or 2.0)
+    cmd = [
+        "ffmpeg", "-y",
+        "-ss", f"{start:.3f}",
+        "-t", f"{duration:.3f}",
+        "-i", str(video_path),
+        "-map", "0:v:0",
+        "-an",
+        "-vf", f"fps={fps:g}",
+        "-c:v", "libx264",
+        "-preset", "veryfast",
+        "-crf", "30",
+        "-pix_fmt", "yuv420p",
+        "-movflags", "+faststart",
+        str(output_path),
+    ]
+    result = run_cmd(cmd, timeout=CONFIG.get("mimo_video_chunk_timeout", 180))
+    if result.returncode != 0:
+        raise RuntimeError(f"MiMo 视频分片裁剪失败: {result.stderr[-500:]}")
+    return output_path
+
+
+def _mimo_chunk_prompt(chunk):
+    return (
+        f"这是原视频 {chunk['start']:.1f}s-{chunk['end']:.1f}s 的场景分片，"
+        f"scene_id={chunk['scene_id']}。"
+        f"{CONFIG.get('mimo_video_prompt', '请用中文概括这个视频分片。')}"
+    )
+
+
+def _mimo_video_model():
+    """MiMo 视频理解使用的模型：优先 mimo_video_model，回退 mimo_model，再回退 vlm_model。"""
+    return CONFIG.get("mimo_video_model") or CONFIG.get("mimo_model") or CONFIG["vlm_model"]
+
+
+def mimo_video_settings_fingerprint():
+    """Return non-secret MiMo video-overview settings that affect generated content."""
+    return {
+        "model": _mimo_video_model(),
+        "mimo_video_api_url": CONFIG.get("mimo_video_api_url"),
+        "mimo_video_fps": CONFIG.get("mimo_video_fps", 2.0),
+        "mimo_media_resolution": CONFIG.get("mimo_media_resolution", "default"),
+        "mimo_video_chunk_max_seconds": CONFIG.get("mimo_video_chunk_max_seconds", 20.0),
+        "mimo_video_chunk_min_seconds": CONFIG.get("mimo_video_chunk_min_seconds", 1.0),
+        "mimo_video_base64_max_mb": CONFIG.get("mimo_video_base64_max_mb", 45.0),
+        "mimo_video_prompt": CONFIG.get("mimo_video_prompt", ""),
+        "mimo_disable_thinking": CONFIG.get("mimo_disable_thinking", True),
+    }
+
+
+def _mimo_chunk_cache_key(chunk):
+    """Stable identifier for a MiMo chunk (index + scene span) for partial-cache reuse."""
+    return (
+        f"{chunk['chunk_id']}|{chunk['scene_id']}|"
+        f"{float(chunk['start']):.3f}-{float(chunk['end']):.3f}"
+    )
+
+
+def _mimo_cached_chunks_fingerprint(done):
+    return stable_hash(done)
+
+
+def _mimo_overview_payload_fingerprint(overview):
+    payload = dict(overview)
+    payload.pop("overview_fingerprint", None)
+    return stable_hash(payload)
+
+
+def _mimo_partial_provenance(video_path, scenes):
+    return {
+        "source_video_fingerprint": file_fingerprint(video_path),
+        "chunks": [_mimo_chunk_cache_key(chunk) for chunk in _mimo_video_chunks(scenes)],
+    }
+
+
+def _load_mimo_partial(partial_path, video_path=None, scenes=None):
+    """Load the internal partial chunk cache, keyed by chunk identifier.
+
+    Returns {} when missing/unreadable or when settings/source/chunk provenance differs,
+    so a changed source video or scene plan cannot reuse paid chunks from another run.
+    """
+    if not partial_path.exists():
+        return {}
+    try:
+        partial = json.loads(partial_path.read_text(encoding="utf-8"))
+    except (json.JSONDecodeError, OSError):
+        return {}
+    if not isinstance(partial, dict):
+        return {}
+    if partial.get("settings") != mimo_video_settings_fingerprint():
+        return {}
+    if video_path is not None or scenes is not None:
+        try:
+            if partial.get("provenance") != _mimo_partial_provenance(video_path, scenes):
+                return {}
+        except (OSError, RuntimeError, TypeError, ValueError):
+            return {}
+    done = partial.get("chunks")
+    if not isinstance(done, dict):
+        return {}
+    recorded = partial.get("chunks_fingerprint")
+    if not recorded or recorded != _mimo_cached_chunks_fingerprint(done):
+        return {}
+    return done
+
+
+def _save_mimo_partial(partial_path, done, video_path=None, scenes=None):
+    """Persist completed chunk results incrementally so paid chunks survive a mid-loop failure."""
+    payload = {
+        "settings": mimo_video_settings_fingerprint(),
+        "chunks": done,
+        "chunks_fingerprint": _mimo_cached_chunks_fingerprint(done),
+    }
+    if video_path is not None or scenes is not None:
+        payload["provenance"] = _mimo_partial_provenance(video_path, scenes)
+    partial_path.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
+
+
+def _mimo_chunks_match(cached_chunks, expected_chunks):
+    if not isinstance(cached_chunks, list) or len(cached_chunks) != len(expected_chunks):
+        return False
+    try:
+        cached_keys = [_mimo_chunk_cache_key(chunk) for chunk in cached_chunks]
+        expected_keys = [_mimo_chunk_cache_key(chunk) for chunk in expected_chunks]
+    except (KeyError, TypeError, ValueError):
+        return False
+    return cached_keys == expected_keys
+
+
+def mimo_video_overview_cache_fresh(overview_path, video_path, scenes):
+    """Return True only when the final MiMo overview matches current inputs/settings."""
+    overview_path = Path(overview_path) if not hasattr(overview_path, "read_text") else overview_path
+    if not overview_path.exists():
+        return False
+    try:
+        overview = json.loads(overview_path.read_text(encoding="utf-8"))
+    except (json.JSONDecodeError, OSError):
+        return False
+    if not isinstance(overview, dict) or overview.get("input") != "scene_chunks":
+        return False
+    if overview.get("settings") != mimo_video_settings_fingerprint():
+        return False
+    overview_fingerprint = overview.get("overview_fingerprint")
+    if not overview_fingerprint or overview_fingerprint != _mimo_overview_payload_fingerprint(overview):
+        return False
+    recorded = overview.get("chunks_fingerprint")
+    if not recorded:
+        return False
+    chunks = overview.get("chunks")
+    if not isinstance(chunks, list) or not all(
+        isinstance(chunk, dict) and _is_mimo_chunk_usable(chunk.get("content"))
+        for chunk in chunks
+    ):
+        return False
+    if recorded != _mimo_cached_chunks_fingerprint(chunks):
+        return False
+    try:
+        if overview.get("source_video_fingerprint") != file_fingerprint(video_path):
+            return False
+        expected_chunks = _mimo_video_chunks(scenes)
+    except (OSError, RuntimeError, TypeError, ValueError):
+        return False
+    return _mimo_chunks_match(chunks, expected_chunks)
+
+
+def _analyze_mimo_video_chunk(chunk_path, chunk):
+    video_url = _video_data_url(chunk_path)
+    if not video_url:
+        raise RuntimeError(f"MiMo 视频分片 {chunk['chunk_id'] + 1} 超过 data URL 上限")
+
+    content_parts = [
+        {
+            "type": "video_url",
+            "video_url": {"url": video_url},
+            "fps": CONFIG.get("mimo_video_fps", 2.0),
+            "media_resolution": CONFIG.get("mimo_media_resolution", "default"),
+        },
+        {"type": "text", "text": _mimo_chunk_prompt(chunk)},
+    ]
+    model = _mimo_video_model()
+    payload = {
+        "model": model,
+        "messages": [{"role": "user", "content": content_parts}],
+        "max_tokens": 1200,
+    }
+    resp = mimo_video_api_call(payload)
+    try:
+        msg = resp["choices"][0]["message"]
+    except (KeyError, IndexError, TypeError) as exc:
+        raise RuntimeError("MiMo 视频分片理解响应缺少 choices[0].message") from exc
+    return {
+        "chunk_id": chunk["chunk_id"],
+        "scene_id": chunk["scene_id"],
+        "start": chunk["start"],
+        "end": chunk["end"],
+        "model": resp.get("model", model),
+        "content": msg.get("content", ""),
+        "reasoning_content": msg.get("reasoning_content", ""),
+        "usage": resp.get("usage", {}),
+        "clip_path": f"mimo_video_chunks/{chunk_path.name}",
+    }
+
+
+_MIMO_REJECTION_MARKERS = (
+    "request was rejected", "considered high risk", "high risk",
+    "content policy", "cannot process", "无法处理", "内容审核", "违规",
+)
+
+
+def _is_mimo_chunk_usable(content):
+    """A chunk is usable only if MiMo returned real analysis (not empty / a moderation refusal)."""
+    text = str(content or "").strip()
+    if not text:
+        return False
+    low = text.lower()
+    return not any(marker in low for marker in _MIMO_REJECTION_MARKERS)
+
+
+def analyze_video_overview(video_path, work_dir, scenes=None):
+    """Use MiMo video understanding over local ffmpeg scene chunks."""
+    if not CONFIG.get("mimo_video_overview", False):
+        return None
+    if not CONFIG.get("mimo_video_api_key"):
+        log("MiMo 视频概览已启用，但未设置 MIMO_VIDEO_API_KEY/MIMO_API_KEY，跳过")
+        return None
+
+    chunks = _mimo_video_chunks(scenes)
+    chunks_dir = work_dir / "mimo_video_chunks"
+    chunks_dir.mkdir(exist_ok=True)
+
+    # 增量缓存：已完成的分片先落盘，避免一段失败时丢弃所有已付费分片
+    partial_path = work_dir / "mimo_video_overview.partial.json"
+    done = _load_mimo_partial(partial_path, video_path, scenes)
+
+    log(f"MiMo 视频理解：按 ffmpeg scene 分片分析 {len(chunks)} 段...")
+    chunk_results = []
+    unusable_chunks = []
+    removed_stale_cache = False
+    for chunk in chunks:
+        cache_key = _mimo_chunk_cache_key(chunk)
+        cached = done.get(cache_key)
+        if cached is not None:
+            if _is_mimo_chunk_usable(cached.get("content")):
+                log(
+                    f"  MiMo 分片 {chunk['chunk_id'] + 1}/{len(chunks)}: "
+                    f"{chunk['start']:.1f}-{chunk['end']:.1f}s（命中增量缓存，跳过）"
+                )
+                chunk_results.append(cached)
+                continue
+            done.pop(cache_key, None)
+            removed_stale_cache = True
+        chunk_path = chunks_dir / (
+            f"chunk_{chunk['chunk_id']:03d}_scene_{chunk['scene_id']}_"
+            f"{chunk['start']:.2f}-{chunk['end']:.2f}.mp4"
+        )
+        _extract_video_chunk(video_path, chunk, chunk_path)
+        log(
+            f"  MiMo 分片 {chunk['chunk_id'] + 1}/{len(chunks)}: "
+            f"{chunk['start']:.1f}-{chunk['end']:.1f}s"
+        )
+        chunk_result = _analyze_mimo_video_chunk(chunk_path, chunk)
+        if _is_mimo_chunk_usable(chunk_result.get("content")):
+            chunk_results.append(chunk_result)
+            done[cache_key] = chunk_result
+            _save_mimo_partial(partial_path, done, video_path, scenes)
+        else:
+            unusable_chunks.append(chunk)
+            log(f"  MiMo 分片 {chunk['chunk_id'] + 1}: 未返回有效内容，保留为待重试")
+
+    if removed_stale_cache:
+        _save_mimo_partial(partial_path, done, video_path, scenes)
+
+    if not chunk_results:
+        log(f"MiMo 视频概览：{len(chunks)} 段均无有效内容（疑似被内容审核拦截），跳过概览")
+        try:
+            partial_path.unlink()
+        except OSError:
+            pass
+        return None
+
+    if unusable_chunks:
+        # Degrade gracefully instead of aborting the whole understanding: some chunks get
+        # moderation-rejected (e.g. a burned-in watermark / violent frames). Write the overview
+        # from the usable chunks; the scenes whose chunks were rejected simply fall back to the
+        # frame-VLM description downstream. (Aborting here would make overview unsafe to enable
+        # by default on any moderated source.)
+        sample = ", ".join(str(chunk["chunk_id"] + 1) for chunk in unusable_chunks[:5])
+        log(
+            f"MiMo 视频概览：{len(unusable_chunks)}/{len(chunks)} 段无有效内容（疑似内容审核），"
+            f"以可用分片降级产出，未覆盖场景回退到逐帧描述。样例分片: {sample}"
+        )
+
+    content = "\n\n".join(
+        f"### 分片 {item['chunk_id'] + 1} "
+        f"(scene {item['scene_id']}, {item['start']:.1f}-{item['end']:.1f}s)\n"
+        f"{item['content'].strip()}"
+        for item in chunk_results
+    )
+
+    overview = {
+        "model": _mimo_video_model(),
+        "content": content,
+        "chunks": chunk_results,
+        "chunk_count": len(chunk_results),
+        "fps": CONFIG.get("mimo_video_fps", 2.0),
+        "media_resolution": CONFIG.get("mimo_media_resolution", "default"),
+        "input": "scene_chunks",
+        "partial": bool(unusable_chunks),
+        "unusable_chunk_count": len(unusable_chunks),
+        "source_video_fingerprint": file_fingerprint(video_path),
+        "chunk_max_seconds": CONFIG.get("mimo_video_chunk_max_seconds", 20.0),
+        "settings": mimo_video_settings_fingerprint(),
+        "chunks_fingerprint": _mimo_cached_chunks_fingerprint(chunk_results),
+    }
+    overview["overview_fingerprint"] = _mimo_overview_payload_fingerprint(overview)
+    overview_path = work_dir / "mimo_video_overview.json"
+    overview_path.write_text(json.dumps(overview, ensure_ascii=False, indent=2), encoding="utf-8")
+    # 所有分片完成后清理增量缓存，保持 work_dir 仅有规范产物
+    try:
+        partial_path.unlink()
+    except OSError:
+        pass
+    log(f"MiMo 分片视频概览完成: {overview_path}")
+    return overview