Skip to content

v0.19.0

Choose a tag to compare

@github-actions github-actions released this 28 Aug 04:52
· 12 commits to main since this release

HotClip v0.19.0 — 证据只算一次,下一刀更快更准

这一版把长视频分析从“一次性中间过程”变成跨任务复用、可检查、可安全清理的本地证据,同时增加无需下载新模型的运动信号。

  • 本地多模态证据索引:响度、镜头、运动、TransNetV2 边界与可选视觉扫描结果按源文件指纹、能力版本和模型身份原子落盘。文件、算法或模型变化时只重建失效能力,损坏条目自动丢弃。
  • 九路证据链:轻量运动峰值加入文本、响度、镜头、表情、视觉、弹幕、语气与笑声/掌声。工作台时间轴新增青色运动曲线,让非语言高能片段的依据直接可见。
  • 抽帧覆盖更稳:运动峰值引导代表帧,但始终为全场均匀覆盖保留三分之一预算,避免前段热闹画面挤掉后段安静却重要的场景。
  • 全入口共享:桌面、CLI、MCP、录播监听与 webhook 使用同一份 Tier-0 信号、镜头边界与可选 VLM 扫描结果;视觉 API Key 不会写入索引。
  • 有界且独立:索引默认最多 64MB,按最近使用淘汰;健康检查显示占用,二次确认只清可重复生成的分析证据,不碰项目、素材、模型、转写缓存或 1GB 渲染缓存。
  • 长任务可取消:信号采集与镜头检测响应取消信号,失败或缓存不可用时保持原有透明重算行为。

验证:103 个测试文件全部通过,1,061 项通过、1 项按既有条件跳过;TypeScript 双目标检查、生产构建、真实 FFmpeg 运动元数据烟测,以及中英文工作台/诊断页响应式浏览器验收均通过。


HotClip v0.19.0 — Analyze once, make every next cut faster and sharper

This release turns long-video analysis from disposable intermediate work into reusable, inspectable and safely clearable local evidence, while adding a model-free motion signal.

  • Local multimodal evidence index: loudness, shots, motion, TransNetV2 boundaries and optional vision-scan outcomes are atomically keyed by source fingerprint, capability version and model identity. Changed files, detectors or models rebuild only stale capabilities; corrupt entries are discarded automatically.
  • Nine-channel evidence chain: lightweight motion joins transcript, loudness, shots, facial emotion, vision, live chat, vocal tone and laughter/applause. A cyan motion curve in the workbench makes non-verbal activity evidence visible.
  • Better frame coverage: motion peaks guide representative frames, while one third of the budget remains reserved for uniform full-source coverage so early spectacle cannot hide a later quiet-but-important scene.
  • Shared everywhere: desktop, CLI, MCP, folder watch and webhooks reuse the same Tier-0 signals, shot boundaries and optional VLM scan outcomes. Vision API keys are never persisted in the index.
  • Bounded and independent: the 64MB LRU index has its own diagnostics and two-step cleanup. Clearing it never touches projects, source media, models, transcript cache or the separate 1GB render cache.
  • Cancellable long work: signal collection and shot detection honor cancellation, while missing or invalid cache entries preserve the existing transparent recompute path.

Verification: all 103 test files passed, with 1,061 passing tests and 1 existing conditional skip. Both TypeScript targets, the production build, a real FFmpeg motion-metadata smoke test, and bilingual responsive browser QA for the workbench and diagnostics passed.