SketchNarrator v0.1.6

Changes

• Keep the annotation panel usable on large projects. Four hot paths that scaled with project size were reworked after measuring the panel on real multi-hundred-caption projects:
• Sample the board paper colour by copying the four corner pixels into a single 4x1 scratch canvas and reading them back once, instead of drawing the whole board into an image-sized canvas and issuing four separate 1x1 readbacks that each force a GPU-to-CPU sync. The sampled colour is unchanged.
• Build the element editor's subtitle option list once per caption list and afterwards only move the selection, instead of re-parsing and re-laying out hundreds of <option> nodes on every element selection.
• Assign canvas.width / canvas.height only when the scene canvas size actually changed, instead of reallocating the multi-megapixel backing store and resetting the 2D context on every repaint.
• Throttle canvas drags to at most one repaint per animation frame and defer change diffing, draft persistence, and timeline rebuilding to pointer release. A click that only selects a region without moving it now performs no bookkeeping at all.
• Treat a filesystem that accepts os.symlink() without reporting the result as a link as unsupported in the release-manifest test, so the suite passes on Windows drives without reparse-point support.
• Publish GitHub releases automatically: pushing a v* tag now verifies the release manifest, builds the deterministic public ZIP, and attaches it to a release named after the tag, using the matching CHANGELOG.md section as the release notes.

SketchNarrator v0.1.5

Changes

• Reserve the shared title card area in the visual layout plan: scenes with a card record layout_plan.title_card_region, preset native layouts start at y=0.16, and board generation stays out of that area. Custom regions that intrude into the reservation fail Gate 2 with an explicit conflict instead of silently moving approved objects.
• Fit the approved card into the same reserved area when rendering: the overlay is scaled proportionally, never cropped and never upscaled, so the card text stays inside its reserved space at every output size.
• Reject an empty, non-numeric, or non-positive title_card_profile.max_text_chars with an actionable message instead of an unhandled exception, and report card text, style, and position problems together.
• Pin the final AAC track to 48 kHz so the delivered mix keeps a single sample rate.

Verification

• Renderer tests: 84 passed.
• Workflow tests: 244 passed, 1 skipped, 1 sandbox-only failure (test_symbolic_links_are_never_release_files cannot observe a real symlink where os.symlink is stubbed; the file is untouched by this release).
• Release manifest check: OK.
• Deterministic release package: sketch-narrator-0.1.5.zip.
• Package SHA256: 60910d8b84856252e84aa7d26fdd8a45b2fbdb234025b564dbb6c5bae74ddca2.

SketchNarrator v0.1.4

Changes

• Fix compact ink stroke erosion during Zhang-Suen skeletonization: recover isolated compact ink components (such as eye pupils, dots, and small facial features) via _recover_compact_ink_strokes and promote core short strokes to short_identity in semantic stroke planning.
• Add ink_color_mode (default "source", fallback "monochrome"): sample original artwork RGB colors for line drawing in light whiteboard mode, faithfully preserving golden sparks, red warnings, green neural paths, and blue accents instead of forcing monochrome thresholding.
• Integrate ink_color_mode across stream_render.py, render_stream_whiteboard.py, and workflow.py render pipelines with deterministic cache invalidation.
• Align documentation and project contract: update references/project-contract.md with renderer_profile.ink_color_mode schema and clarify binary threshold edge boundaries vs. original RGB line color sampling.

Verification

• Renderer tests: 82 passed.
• Workflow tests: 240 passed, 1 skipped.
• Release manifest check: OK.
• Deterministic release package: sketch-narrator-0.1.4.zip.
• Package SHA256: b994aee4a587a89b26c924149615b506a463094da1b6f865f80d4436ce1f363a.

SketchNarrator v0.1.3

Changes

• Add an explicit opening-anchor beat to every shot so the first drawing starts about 100ms after the scene begins.
• Recommend opening strokes within 200ms and reject plans or annotations whose first visible element starts later than 500ms.
• Mirror the check in board QA: advisory between 200ms and 500ms, hard error beyond 500ms.
• Update the skill contract, visual-production guide, final QA guide, and project file schema.

Verification

• Renderer tests: 77 passed.
• Workflow tests: 240 passed, 1 skipped.
• Release manifest check: OK.
• Deterministic release package: sketch-narrator-0.1.3.zip.
• Package SHA256: ed2484ef86368736589644a31f4c894718fe235324d0b28bc90be4f47a21a731.

SketchNarrator v0.1.2

新增与改进 (Features & Improvements)

• 自适应手绘节奏求解器 (Adaptive Duration Pacing)：
  • 新增 scripts/pacing.py 及 workflow.py pace-annotations 命令。
  • 根据分镜词级时间戳智能拉长元素绘制时间（默认预算比例 92%，留出 250~500ms 观察留白），彻底消除 1.2 秒过快闪电画完与数秒死板静止的问题。
• 分镜空窗期质检告警 (Idle Pacing QA Advisory)：
  • 在 board_qa.py 中增加对元素绘制时长偏短（可用窗口 ≥ 3 秒但绘制占比 < 55%）的质检预警。
• 场景标题卡支持 (Scene Title Cards)：
  • 将可选的场景标题卡提升进入分镜、Gate 2 确认、视觉计划、渲染缓存与面板预览。

验证 (Verification)

• 单元测试套件 236 项全量通过（0 failure, 0 error）。
• 在真实医学科普长视频项目（6 场景，131.77 秒）中全流程实机验证通过，笔画动作平缓饱满延伸至句尾。
• 发布清单全量校验通过，包含 153 个确定性打包文件。

Assets & Checksum

• Package: sketch-narrator-v0.1.2.zip
• SHA256: 3cb3f066a72531decf96eeda2891e958f70922f9cdac0725c27023f9cfc10383
• Size: 30,244,169 bytes

SketchNarrator v0.1.1

修复

• 修复多场景项目在 Gate 2 被错误阻断的问题。
• 单幕诊断现在保留完整 words.json 与全局词锚点，只把全片覆盖检查交给最终全片编译。
• 正式编译仍严格检查词条遗漏、重复和交叉。

验证

• 使用真实失败项目只读复核：6 幕、474 个词、6 个镜头成功编译。
• 上层 228 项测试通过，1 项按设计跳过。
• 渲染器 76 项测试通过。
• 发布清单检查通过，公开 ZIP 未包含私人资产或本机路径。

Fix

Fixes a Gate 2 validation failure that blocked every multi-scene project. Per-scene diagnostics now preserve global word anchors, while final compilation continues to enforce complete, ordered narration coverage.

SHA256: 9282a05114b2afc62a943e6be18425e0f0c8b7927677c835305610e598a586c2

SketchNarrator v0.1.0

SketchNarrator v0.1.0

SketchNarrator 的首个公开版本。它把中文主题、文案或参考视频整理为按真实旁白时间驱动的手绘讲解视频，并把分镜、画板、标注、动画计划、渲染和成片 QA 串成可复核的本地工作流。

主要能力

• 三次确认流程：口播、视觉分镜、工作台画面与执行计划。
• 工作台支持逐幕拖拽、缩放、新增、删除、排序和撤回；用户确认拥有最终执行权。
• 同一语义对象 ID 贯穿分镜、标注、动画计划和渲染，执行直接读取已批准版本。
• 时间计划按真实音频形成，并确定性编译到目标帧率。
• 10 种内置手绘风格参考图，以及可替换的公开 CC0 轻量音效预设。
• 使用宿主 Agent 自带的生图能力，无需另接图片生成 API。
• 面向 Codex、Antigravity、WorkBuddy、豆包办公等具备文件操作、命令执行和生图能力的 Agent。
• 私人角色、字体、音效与片头片尾通过仓库外的本地配置接入，不包含在公开包中。

安装

下载本 Release 的 sketch-narrator-0.1.0.zip，解压后按照 README 的对应平台说明安装。也可以直接克隆仓库。

发布校验

• 公开清单：150 个源文件。
• Release ZIP：151 个文件（含 release-manifest.json）。
• ZIP SHA256：f9248af5d634ac41cb2b742f40bb216ac162e8890b5323996c06b7b74e02a048。
• 发布清单校验、公开边界检查、本地验证和 GitHub Actions 均已通过。