DeepBlend Studio 0.6.4 — Smaller recipe context for modeling tasks

Selecting a product recipe now returns the creator-facing metadata and parameter choices without hundreds of internal assembly binding rows. Author, declared license, source, version, digest, preview and parameter IDs/types/defaults/ranges are retained. Host and author SDK catalogs, project validation and saved source locks keep the full maps.

With the same four built-in recipes, the actual model tool result is 17,131 bytes instead of 134,845 bytes (about 87.3% less). Real Host integration verifies creation through the returned digest, continued resizing, invalid-range refusal and original file protection. Four deliberate regressions are rejected.

A same-bai/DeepBlend Studio/natural-prompt pair produced matching saved geometry, materials, cameras, animation and render profiles. Hero renders at frame 24, 8 samples and 768 × 576 have identical decoded pixels. The observed steps were 9 to 8 and model+tool totals 219.168 to 163.866 seconds; one pair with variable load/cache/model randomness is not a general performance or reliability claim. Developer-mode premature observation and final-reply timeout cases remain failures.

Candidate packaging checks load seven modules and 18 tools, verify the compact model catalog and full Host maps, retain old embedded locks and decode PNG/JPEG. Installed bundle provenance and complete browser bytes are checked separately from native modeling. Cross-platform payload checks do not establish executing every target.

The corrected frozen candidate completed one continuous 39-suite acceptance run, including all 132 unit/contract files; all 688 tracked files remained unchanged. Candidate CI and exact merged-main CI each passed three jobs. Candidate CI's synthetic merge checkout tree was independently matched to the frozen source. The original misrouted batch-baseline failures are retained; explicit batch selection and a killed regression preserve the strict speed assertion.

Built from clean pushed main 9eebafb. Every one of the 341 bundle files matches staging, the extracted SDK and unchanged isolated DSH installation passed, and all seven npm archives match final main. All immutable core assets match GitHub server sizes/SHA256. All seven npm packages are publicly served; anonymous complete-archive SHA512/SHA1/SHA256 and every file match the final-main preparation. Public tag, binary, empty-cache CLI and whole three-route readbacks are collected after publication and recorded separately.

Public readback completed: actual tag v0.6.4 points to 9eebafb. Anonymous latest bundle bytes match the clean build; the public CLI in an initially empty cache outside the repository reports installation, image and video readiness. One complete source/npm/tarball installation comparison agrees on 0.6.4, workbench HTTP 200 and both exact presets, with 13/13/1 packages fetched and zero issues. 417 Host product module reads belong to the three actual installations. This web-only observation loads six Host namespaces; separate extracted SDK and real agent checks verify seven modules and 18 tools. The initial postprocessing expectation failure is retained.

The 58 immutable core assets remain unchanged. Twenty append-only publication readback assets independently match GitHub sizes/SHA256, bringing the verified total to 78. Core index · SHA256SUMS · Public readbacks.

Community submission #6868 is updated with the stable release and verified canonical entry bytes. The entry uses version-free URLs and required no empty content commit. Upstream merge and storefront refresh remain unverified; this records the submission update only. Next: source-backed thick-wall cup and curved-lamp tasks with real native dimensions, assembly, fixed renders, actions/time/rework and preserved old files.

DeepBlend Studio 0.6.3 — Smooth woven speaker detail

New speaker projects use recipe 2.1.0 with smoother woven warp bends. The 0.32 mm wire radius, circular grille trim, 111 editable entities and three millimetre assembly controls are retained. Saved 2.0.0 projects keep their embedded original geometry and continue resizing.

Native Blender checks at 264 × 138 × 162 mm and 340 × 180 × 220 mm found zero weave surface intersection candidates after the change, while the other 98 visible product meshes retained identical geometry. Fixed-condition renders at 264 × 138 × 162 mm and the saved 340 × 138 × 200 mm state preserve the overall product appearance and show smoother bends in closeups. Triangles increase about 2.42 times. Surface checks are not manufacturing certification.

Real bai / deepseek-v4.1-flash recipe tasks created a saved 300 × 138 × 162 mm speaker and, in DeepBlend Studio, resized an existing copy to 340 × 180 × 220 mm with a new native checkpoint and hero preview. Independent reopening verified dimensions and weave checks; original revision files remain unchanged. These are maintainer recipe cases, not arbitrary modeling or external adoption.

The Studio edit finished after 14 model steps and substantial context use; no efficiency claim. The development-mode edit exceeded a five-minute observation bound and remains a failed run. Image attachments were not accepted by the selected text model and the secondary visual-review route failed, so no model artistic approval is claimed. Actual renders were independently viewed.

Validation completed before this draft:

• One full acceptance invocation passed all 39 suites at d07b8a5. All 686 tracked files remained unchanged; unit coverage includes 132 files and 1,589 node:test cases.
• All three candidate CI jobs passed. Merged main 9caa704 has an identical source tree; all three main CI jobs passed.
• Clean merged main built the bundle; all 341 regular archive files match staging. The extracted package loaded seven modules and 18 tools, retained old embedded speaker locks and decoded PNG/JPEG.
• Installation in an independent unchanged Harness deployment selected the installed bundle. All 139 observed profile product module reads stayed inside it; served browser bytes and both presets match.

Public closeout:

• All seven npm 0.6.3 archives were independently fetched without credentials. SHA512/SHA1/SHA256 and every archive file match the exact merged-main preparation; each version was accepted once.
• The public tag points to 9caa704. The anonymous complete latest binary is 72,170,603 bytes and its SHA256 matches the clean main build.
• An external empty-cache CLI with separate empty user/global npm configuration passed installation, image and video readiness.
• One complete public source/npm/tarball parity invocation passed: all version 0.6.3, workbench HTTP 200, both presets identical, fetched package counts 13/13/1, zero issues. All 417 observed product module reads belong to the three actual profile homes; tracked main source stayed unchanged.
• Community submission PR #6868 has the exact updated 0.6.3 description and unchanged canonical YAML projection. It remains open and unmerged; no storefront refresh is claimed.

The 48 immutable core evidence assets are retained. Eleven supplemental public readbacks bring the release to 59 independently checked assets with matching sizes and server SHA256 digests.

DeepBlend Studio 0.6.2 — Preview the model you just saved

Saving a model and immediately clicking Preview now renders the current saved revision. Pending preview reads are retired when you switch projects, leave and return to the same project, or close the workbench. Duplicate clicks and previews during scene writes are blocked.

The standalone and embedded workbenches share the fix. Existing speaker dimensions, materials, photography settings and saved history are preserved.

Validation:

• One complete acceptance invocation passed all 39 suites at 8eec315; all 677 tracked files remained unchanged.
• The merged source 1ab82cc has an identical source tree. Full units covered 132 files and 1,586 node:test cases; the self-counted total depends on whether presets are installed.
• The workbench contract passed 71 checks; five deliberate mutations failed explicit regression assertions. Real Chrome/Blender photography passed 36 checks with reviewed desktop and narrow screenshots.
• The exact merged source built the bundle. All 341 regular archive files matched staging. The extracted package loaded seven modules and 18 tools, preserved recipe edit/retry history, and decoded PNG/JPEG.
• Installation in an independent, unchanged Harness deployment selected the installed bundle. Every observed product-module resolution stayed inside that bundle; the served browser client matched its full bytes, and both presets matched.

This patch changes preview workflow, not modeling geometry. Native payloads for macOS arm64, Linux x64, Linux musl x64 and Windows x64 have file, license and architecture checks; this is not a claim of executing every platform. Natural-language modeling reliability, artistic approval and external adoption remain separate measurements.

See the attached raw acceptance log, source hashes, native evidence and file proofs, package integrity metadata and checksums. Community submission is tracked separately from upstream acceptance and storefront refresh.

Publication readback: All seven npm 0.6.2 archives were freshly downloaded without credentials; SHA512 integrity, SHA1/SHA256 and every archive file matched the prepared main-source packages. All28 core GitHub assets have been independently matched by size and server-provided SHA256 digest. MainCI38204951463 completed all three jobs successfully. One whole public source/npm/tarball parity run passed: all0.6.2, HTTP200 workbenches, byte-identical two presets, fetched packages13/13/1, zero problems, tracked source unchanged. Existing community PR6868 was updated and independently read back; it remains OPEN, not an upstream merge or verified storefront refresh. Supplemental public readback proofs are attached.

DeepBlend 0.6.1 — verified installation ownership

DeepBlend 0.6.1 加固安装一致性：完整插件通过自己的入口加载运行器、Host、预设和模型工具；工作台加载属于当前安装的客户端。单独挂载 UI 组件仍可打开全屏工作台。

安装验证现在会核对 DSH 实际选择的 bundle 配置路径。若 DSH 部署位于另一个旧 DeepBlend 工作区内，Harness 的安装优先规则可能覆盖刚安装的 profile 包；检查会明确指出这一情况。请使用独立于其它 DeepBlend 工作区的 DSH 部署。

已有音箱继续使用三个毫米尺寸输入调整 多部件音箱场景（111 个实体，含辅助工具和布景），保留独立外观、摄影和历史。参数保存、版本恢复、PNG 原图与 MP4 交付保持原有操作方式。

发布校验直接读取归档的受检文件，减少约 150 MB 的额外展开空间。仍校验四个目标平台的包版本、许可、二进制头和完整字节；不代表已经在所有平台执行。对应 49 项检查包括临时目录不可创建、符号链接、缺失文件和字节变化。

验证范围：冻结产品来源 d25fa15 的整轮本地运行为 33/39，六项在磁盘耗尽期间失败；在不改源码的情况下逐项复验后，创建恢复29、摄影36、尺寸创建11、已有装配24、预览历史71、固定视角23均通过。39 套流程分别有通过证据，未把原失败运行写成一次整轮通过。676 个源文件前后不变。之后 d2e1354 只修改发布校验工具和测试，七个产品包字节相同。

候选 d2e1354 的三项 CI 全部成功，原始单元日志为 132/132 文件、1586 个 node:test 用例。合并主线为 17fa874。确切主线构建的 341 个打包文件与暂存字节全部一致；真实隔离 DSH 安装、组件路径、完整客户端和两套预设核验通过。合并主线的三项 CI 也全部成功；确切安装包的 SDK 七模块、18 个工具、参数保存/重试/原始保护和 PNG/JPEG 解码通过。公开发布读回已完成：七个 npm 包独立匿名 SHA512/SHA1/SHA256 与每个文件匹配；仓库外空缓存 CLI 正常，公开安装包完整 SHA256 匹配。一次完整公开三路安装（源码、npm、Release tarball）全部通过，工作台与两套预设一致，108条实际profile组件路径都指向隔离安装，源文件前后不变。社区现有投稿 #6868 标题与说明已更新并独立读取核对，规范 YAML 与源码投影完全一致；仍等待上游审核，未声称已收录。社区投稿状态以上游审核结果为准；脚本化工具检查不计为真实 DeepSeek 自主建模或外部用户采用。

DeepBlend 0.6.0 — 继续调整已有整套模型

已有音箱项目现在可以直接调整整套模型的宽、深、高，一次保存生成同一项目的新版本和真实预览。无需重新创建模型：外壳、内腔、开口和部件装配位置联动，保留独立修改的材质、倒角、修饰器与摄影设置。恢复历史也会恢复该版本的整体尺寸。

在场景树的“调整整套模型”中修改三个毫米输入，然后点击“保存整套尺寸并预览”。音箱 2.0.0 支持宽 264–340、深 138–180、高 162–220 mm。自定义摄影保持原样，保存后可检查新的构图。手动修改了受控几何时，会列出相关部件并拒绝覆盖；旧配方继续使用部件编辑。

智能体新增 blender_recipe_parameters，先读当前版本和范围，再条件保存。来源锁与参数状态一起提交，相同请求重试不会重复创建版本。作者 SDK 新增保存来源解析与受约束的普通场景补丁规划。

发布维护修复 --dry-run：完整验证七个 npm 包并明确报告未发布，不再把缺少公开版本误报为失败。真实发布仍要求公开 registry 读回。

验证：冻结实现 ee7d833 的完整入口 39/39，669 个源码文件运行前后相同。最终 ae62936 仅增加发布脚本修复、对应检查和文档，七个产品包与配方字节相同；该源码三项 CI 全部成功，原始单元日志核对 132/132 个文件、1582 个测试用例。Linux 已有装配编辑报告 24/24，实际 CI checkout tree 已核对。摄影 36/36、仓库外严格 TypeScript/JavaScript SDK 消费者、七模块/18工具安装包与每个打包文件检查通过。合并后的主线来源为 35f3d1b，完整证据随 Release 附件保存。

用法：音箱尺寸说明。组件：npm bundle 0.6.0。

公开安装读回：七个 npm 包的匿名完整性与全部文件检查、仓库外空缓存 CLI、公开 Release 安装包字节核对通过。一次完整三路安装（源码、npm、Release tarball）在确切 35f3d1b 来源及隔离的原版 DSH 0.1.5-rc.2 部署下通过，实际 profile 组件路径核对 114 条无旧工作区回退；原始日志、模块路径和校验值已补充附件。社区现有投稿 #6868 已更新，等待上游审核。

早期 tarball 验证失败来自本机 DSH 嵌在旧 DeepBlend 工作区内，Harness 优先加载其安装位置的旧 bundle 配置。隔离部署对照中原 0.6.0 包也通过，不能把此环境问题归因于安装包。请将 DSH 独立于其它 DeepBlend 工作区安装；后续来源探针会直接检测这种覆盖。原失败记录保留。

DeepBlend 0.5.0 — linked dimensions for complete speaker assemblies

DeepBlend 0.5.0 — 三项尺寸创建完整桌面音箱

在图库选择音箱，填写箱体宽、深、高，即可创建含 111 个部件的可编辑 Blender 模型。只把宽度设为 300 mm，保持默认深、高，就得到 300×138×162 mm 的箱体。外壳、内腔、开口和盖板联动；格栅、旋钮和脚垫跟随装配位置，保留零件尺寸、底面高度和按钮按压行程。

支持宽 264–340、深 138–180、高 162–220 mm。无效输入会在创建前被拒绝。摄影机随尺寸和对应部件调整，最大尺寸下的旋钮仍保留在细节画面内。原版配方输入完整保留，图库升级不会改写旧项目。

作者接口增加局部三轴盒体尺寸，以及受约束的尺寸、位置、摄影机和动画位置偏移绑定。运行组件不支持新能力时，在编译前明确提示，并保护当前版本。

尺寸输入用于创建新模型。已有音箱的整套尺寸修改是下一轮重点，当前仍可编辑部件、材质和摄影设置。

验证：运行实现的冻结完整验收 38/38；0.5.0 来源的完整单元检查 132/132；干净包内七模块加载、PNG/JPEG 处理及完整配方解析通过。最终6894453的三项CI全部成功，Linux全屏112/112及真实父进程核对通过；对应干净构建及安装包读回均通过。保存文件重开、尺寸边界、按钮行程、部件与材质编辑、历史恢复及并发冲突均有实际 Blender/浏览器证据。

真实 DeepSeek 请求返回密钥无效，尚未产生建模工具调用，在线 Agent 验收未完成。边界差检查不等于最小壁厚或制造认证；外部艺术审阅和用户采用数据尚未测量。

0.4.0 — 复用已保存作品的摄影设置

从已保存作品选择版本和相机，把相机设置与整组灯光套用到当前作品草稿，减少重复填写。检查构图后，明确保存并预览；保存记录保留来源作品、版本与摘要。

• 保留目标作品的模型、材质、世界、动画、其他相机、传感器/光圈/裁剪范围和交付设置。
• 对象瞄准需要明确映射，已有摄影草稿需要明确选择替换。读取失败可重试；迟到或错误身份的回执不能覆盖后来选择。
• 支持清除旧瞄准、返回自由旋转；相同灯光仅顺序不同不会生成多余版本。
• 坐标按米直接复用，不自动缩放或取景。不同尺寸模型仍须检查构图。更新插件并重启 Host 后使用新能力。

最终冻结连续完整验收通过 37/37 套件，契约 128/128；646 个源文件摘要保持。真实摄影检查 36/36，包含独立重开 Blender checkpoint、保存来源、原文件保护与 390px 界面；11 种错误在指定断言被拒绝。首轮图片来源登记失败和原五分钟 CI 中断证据保留，不计为最终通过。确切候选与合并主线均已通过三项 CI，源码树相同。干净克隆、对应构建、包内模块与 PNG/JPEG 实际解码、相机清除目标和环境自检均通过；原始产物已核对 API 摘要、CRC 和保存文件。公开读回结果在发布记录中逐项登记。

摄影使用指南 · 发布记录 · 产品计划

安装：@deepblend/dsh-blender-bundle@0.4.0，或下载版本无关的 deepblend-bundle.tgz。七个 npm 包按相同版本发布，17 个工具保持。

DeepBlend 0.3.9 — previews stay with your work

Keep preview images and their download addresses with the work you selected. When a preview read fails, show the failure with an explicit refresh instead of presenting the earlier project’s image. Refresh only reads: it keeps unsaved edits, saved versions and original PNGs. Validate response project/address identity and ignore older or stopped reads.

切换作品时不沿用上一项目的图和下载地址；读取失败可明确刷新恢复，未提交修改与保存产物保持。既有版本冲突仍需要明确处理，刷新不会自动创建、保存或渲染。

Validation and source provenance are attached with checksums. Packed execution is verified on macOS arm64; other included native targets are checked for files/licences, and repository rendering/browser behavior is separately verified by Linux CI. No external user adoption is claimed.

0.3.8 — Reliable saved-version comparison

Compare saved versions directly from history. A historical entry opens that version against the current work; the current version chooses its nearest other saved version. A single version has a disabled entry with an explanation.

Keep unsaved edits while viewing. Clear earlier structural differences and reject late success/error replies after later selections or requests; response identities must match the displayed project and pair. Viewing does not restore, create or automatically render. Recorded image conditions remain explicit.

从版本记录直接查看保存版与当前作品，保留未提交修改；旧配对、旧差异和迟到回执不再混入后来查看。查看与明确恢复、保存和渲染分别操作。

Validation, immutable source provenance and public installation evidence are supplied with checksums. No external user adoption or execution on every included native platform is claimed.

0.3.7 — Keep recipe comparison settings

Compare recipes without losing each set of edits. Returning to a recipe restores its parameters; clicking Selected preserves them, and Reset affects only that recipe. Browser drafts retain comparison settings across reload, with explicit restore and separate recipe versions/digests.

Creating uses the selected inputs. Successful creation preserves other choices and later edits, while retrying an earlier request keeps its original captured inputs. Legacy 0.3.6 browser drafts remain readable. Example images continue to show defaults; explicit creation renders the selected settings.

Validation and public installation evidence are attached with checksums. Browser storage limits remain visible; comparison settings are local to this browser and project storage.

对比配方时保留各自的调整：切换回来继续修改，明确重置只影响当前配方。浏览器草稿恢复比较设置，明确创建才使用当前输入渲染，原请求重试不受后来修改影响。