v1.2.2 — optional scene guides and distribution packages

中文说明（本页全文） · 中文版（独立页面） · English notes

v1.2.2 — optional scene guidance and distribution packages

2026-11-02. This maintenance release packages the optional guides and distribution tools already reviewed in the repository. Downloading it does not update an existing installation.

Changes

• Include the merged action-icon card and optional scene-space and scene-revision guides. They cover viewing direction, connected forms, object placement, surface-bounded fills, readable patterns and related changes. Read the relevant section when useful; user references take priority.
• Include the reviewed full-size Pixel Anchor logo kit as a separate download: six appearances at 32/64/128/256/512px, editable specifications, previews, license and checksums. The 256/512px badges have enhanced detail; the smaller badges and transparent marks remain simple. The Skill keeps its existing 32/128px icons.
• Include the merged local-link checking fix in the package checks. Add a separate SkillHub packaging tool that adds platform metadata to the complete Skill without changing its generic copy.

Behavior and dependencies

The only runtime-source change from v1.2.1 is the version constant. Rendering, PNG import, named-cell replacement, exact RGBA verification, assembly, JSON v1, Python APIs, recipes and example pixels remain unchanged.

Python 3.10+ and Pillow >=10.4,<13 remain the declared range; requirements.txt pins Pillow 12.3.0 for reproduction. Windows compatibility sampling reached Python 3.15.2 with that baseline. Pillow 13 passed a relaxed-guard experiment, but remains outside the supported range. The NPX installer requires its own Node runtime, installs files from the repository branch and still requires Python/Pillow setup; use the complete Skill subdirectory in INSTALL.md.

Packages

The GitHub attachments are deterministic-pixel-art-v1.2.2.zip, deterministic-pixel-art-logos.zip, package-manifest.json, verification.json and one SHA256SUMS.txt covering the other four files. English and Chinese notes are readable on the Release page. The SkillHub tool creates a separate platform upload package with root SKILL.md, the same runtime version and a recorded source commit; platform login, slug ownership and review are separate from local packaging. This GitHub release does not upload to SkillHub.

The frozen Skill ZIP passed 106 isolated execution checks; the platform ZIP passed the same 106 checks. All 30 editable logo specifications rerendered to identical RGBA. The official SkillHub collector matched all 104 platform files; its dry-run checked metadata only. Exact-candidate CI passed all eight Windows/Ubuntu jobs, with 178 source tests in each compatibility job. The attached verification.json records the final release commit, its main CI and the unchanged package hashes. These checks are mechanical acceptance by the coordinating model, without a new independent model review or creative trial.

Limits

The guides and packaging checks do not establish general improvements in artwork quality, autonomous correction, time or token use. Existing external trial failures and evidence limits remain recorded. Native512 creative validation and other unselected features remain outside this maintenance release.

中文说明

v1.2.2 — 可选场景指南与分发包

2026-11-02。本次维护版交付仓库中已审阅的可选指南与分发工具。下载版本包不会自动更新已有安装。

变化

• 纳入已合并的动作图标卡、场景空间与场景修订选读指南，涉及观察方向、相连体块、对象位置、限定在表面内的填色、可读纹理及相关部件的同步修改。需要时读相关段落，用户参考优先。
• 将已审阅的全尺寸 Pixel Anchor Logo 素材作为独立下载：六种外观、32/64/128/256/512px，包含可编辑规格、预览、许可与摘要。256/512px 徽章采用增强细节，较小徽章与透明标记保持简洁；Skill 仍使用现有 32/128px 图标。
• 纳入已合并的包内链接检查修复。增加独立的 SkillHub 打包工具，从完整 Skill 生成平台元数据，不改通用副本。

行为与依赖

相对 v1.2.1，运行时源码仅改变版本常量。绘图、PNG 导入、命名单格替换、RGBA 精确核对、拼接、JSON v1、Python API、配方和示例像素均保持。

声明范围仍为 Python 3.10+、Pillow >=10.4,<13；requirements.txt 固定 Pillow 12.3.0 用于复现。Windows 抽样已验证 Python 3.15.2 配合该基准；Pillow 13 在放宽检查的实验副本中通过，仍在支持范围之外。NPX 安装器需要自己的 Node 环境，从仓库分支复制文件，之后仍要准备 Python/Pillow；使用 INSTALL.md 中的完整 Skill 子目录入口。

包文件

GitHub 附件为 deterministic-pixel-art-v1.2.2.zip、deterministic-pixel-art-logos.zip、package-manifest.json、verification.json，以及覆盖其余四份文件的一份 SHA256SUMS.txt。双语说明在 Release 页面直接阅读。SkillHub 工具 另行生成平台上传包，根目录是 SKILL.md，运行时版号相同并记录源提交；账号登录、slug 所有权与平台审核另行完成。本次 GitHub 发布不包含 SkillHub 上架。

固定 Skill ZIP 的 106 项隔离执行检查通过，平台 ZIP 的同组 106 项检查也通过；30 份 Logo 可编辑规格重绘后的 RGBA 一致。官方 SkillHub 文件收集器核对了全部 104 份平台文件，其 dry-run 仅检查元数据。精确候选 CI 的 Windows/Ubuntu 共八项任务全部通过，各兼容任务运行 178 项源码测试。附件 verification.json 记录最终发布提交、对应 main CI 与保持不变的包摘要。这些是主协调模型完成的机械验收，本轮没有新增独立模型评审或创作试验。

限制

指南和打包检查不证明普遍画质、自主纠错、时间或 token 收益。原外部试验失败和证据局限继续保留。原生 512 创作验证及其他未选功能仍在本维护版范围之外。

v1.2.1 — clearer guidance and Skill icons

中文说明（本页全文） · 中文版（独立页面） · English notes

v1.2.1 — clearer guidance and Skill icons in the download

This maintenance update packages the wording corrections and Pixel Anchor icons previously reviewed in the repository.

Changes

• Clarify Chinese and English descriptions of interrupted writes, report-size limits and image checks in the target game or application. Commands and functional requirements are unchanged.
• Include 32×32 and 128×128 Skill icons, with the updated display name, description and icon paths. README headers and fonts remain repository-only.
• Keep rendering, PNG import, scoped edits, atlas replacement, validation, assembly and example behavior unchanged. The only runtime-source change from v1.2.0 is the version constant; JSON v1, Python APIs and example pixels remain compatible.

Installation and files

Follow the AI installation guide, download this release's ZIP and check it against SHA256SUMS.txt before extraction. Python 3.10+ and Pillow >=10.4,<13 are supported; the packaged requirements.txt pins Pillow 12.3.0. Keep update backups outside the folders the agent scans for Skills.

Download attachments are deterministic-pixel-art-v1.2.1.zip, package-manifest.json, verification.json and SHA256SUMS.txt. English and Chinese notes are readable on the Release page, without separate Markdown downloads. verification.json records the release commit, CI results, package hash and checks.

The original v1.2.0 ZIP and tag remain available and unchanged. Downloading the new package does not update an existing installation. Whether an AI application displays the Skill icons still needs to be checked after that application loads the package.

Verification scope

Release checks cover package files and hashes, icons and display metadata, execution after extraction, existing commands and examples, and repeatable packaging. Exact results and the checked commit are recorded in verification.json.

External trials do not establish general improvements in image quality, task time or token use. LT04's original direction error and LT08's incomplete early history remain recorded. Minecraft game integration, other fonts, artwork drawn directly at 512×512 and drawing by multiple agents still need separate validation.

中文说明

v1.2.1 — 说明更清楚，正式包加入 Skill 图标

这是 v1.2.0 的维护更新，将此前已在仓库完成的说明修正和 Pixel Anchor 图标纳入正式下载包。

本次变化

• 改写难懂的中英文说明，直接解释文件写入中断、报告长度限制，以及在实际游戏或应用中检查图片的方法；原有命令和功能要求不变。
• 包内加入 32×32、128×128 两张 Skill 图标，并更新展示名称、简介和图标路径。README 标题图和字体仍只放在仓库，不进入 Skill ZIP。
• 渲染器、PNG 导入、局部修改、图集替换、校验、组件拼接和示例脚本的行为均未改变。运行时代码相对 v1.2.0 只修改版本号；JSON v1、Python API 和示例图片保持兼容。

安装和文件

按 AI 安装指南 下载本版 ZIP，核对 SHA256SUMS.txt 后再解压。运行需要 Python 3.10+ 和 Pillow >=10.4,<13；包内 requirements.txt 指定 Pillow 12.3.0。更新时请保留旧版备份，并将备份放到 AI 不会扫描加载 Skill 的目录。

下载附件为 deterministic-pixel-art-v1.2.1.zip、package-manifest.json、verification.json 和 SHA256SUMS.txt。中英文说明在 Release 页面直接阅读，不另附需要下载的 Markdown 文件。verification.json 记录本次发布对应的提交、CI、包摘要和检查结果。

v1.2.0 的原 ZIP 和 tag 保留不变。下载新包不会自动更新已有安装；Skill 图标能否在某个 AI 应用中显示，还需由该应用加载后确认。

验证范围

发布检查核对包内文件和摘要、图标与展示配置、独立解压运行、已有命令和示例，以及重复打包结果。实际结果和精确提交以 verification.json 为准。

外部试用不能证明工具能普遍改善画质、缩短任务时间或节省 token。原有 LT04 方向错误和 LT08 早期过程记录不全的限制继续保留。Minecraft 游戏内效果、其他字体、512×512 直接绘制和多个 AI 协作绘图仍需分别验证。

v1.2.0 — editable pixels and scoped AI workflows

中文说明（本页全文） · 中文版（独立页面） · English notes

v1.2.0 — PNG import, precise edits and component assembly

Create and edit pixel-art icons, sprites and UI from text specifications or short Python examples. Give individual images names in an atlas, or assemble separately drawn components into a scene. Output is PNG; Minecraft styling is optional.

Changes since v1.1.1

• Inspect image dimensions, colors, transparency, connected pixel regions and symmetry, or display a small region as a symbol grid. Reports cap the number of listed items to keep output manageable. Choose review backgrounds and reuse existing atlas layouts.
• Import static 8-bit PNGs without reducing colors, preserving RGB values even in fully transparent pixels. Import a whole image or select one named atlas cell after checking its layout. Replace that cell while keeping the other cells and layout unchanged.
• Compare exact RGBA values before and after an edit to check that changes stay inside allowed regions or masks. Assemble PNG components with explicit crops, permitted overlaps and required coverage; inspect seams in coordinate-labeled previews.
• Use editable Python examples for icons, button states, a 128×128 sprite and component scenes. A bridge example draws the left and right chunks separately, with plank patterns aligned across the join. Guides cover small edits, text inspection, visual review and checks in the actual game or application.
• If the filesystem cannot create hard links, set PIXELART_PORTABLE_OUTPUT=1 once in the terminal for subsequent commands and examples. Per-command options take priority. Overwriting existing files still requires explicit --force.

Compatibility and setup

Python 3.10+ and Pillow >=10.4,<13 are supported. requirements.txt pins Pillow 12.3.0 for reproducibility; the reference environment is Python 3.14.2 / Pillow 12.3.0. Runtime dependencies remain Python and Pillow.

JSON v1 and existing API/command defaults remain compatible. Drawing primitives overwrite RGBA; blits use source-over composition unless an explicit supported alternative is chosen. PNG round trips preserve decoded 8-bit RGBA, not necessarily compressed PNG bytes, layers or original drawing instructions. Animation and unsupported high-bit-depth inputs are rejected.

Extract the ZIP, install its requirements.txt in the task's Python environment, then run from the extracted parent directory:

python deterministic-pixel-art/scripts/pixel_art.py render --spec deterministic-pixel-art/assets/gem-16.json --out output/gem.png --preview output/gem-preview.png
python deterministic-pixel-art/scripts/pixel_art.py inspect --image output/gem.png --grid

When updating a Skill, keep backups outside the folders the agent scans for Skills, so it does not load both copies. In portable-output mode, another program may read a new file before writing has finished. In any output mode, a program crash or power loss can leave a group of output files only partly updated. After an interruption, check that all expected files are present and complete before using them.

Evidence and limits

The v1.2.0 release passed 165 source tests, 106 isolated package tests and 43 reference-image checks. Repeated builds produced identical packages, and the Skill passed the official format validation. Of its 98 files, 97 match the accepted development package byte for byte; the remaining module changes only its version constant. verification.json records the release commit, its Windows/Ubuntu CI results and the package's source.

External trials cover eight task categories. Records retain LT04's original reversed direction and missing early revisions, and LT08's incomplete early history. These trials do not establish general improvements in image quality, task time or token use. Minecraft game integration, other fonts, artwork drawn directly at 512×512 and drawing by multiple agents remain unverified. Passing automated checks still leaves the image's meaning and appearance to be reviewed.

Distribution

Download attachments are deterministic-pixel-art-v1.2.0.zip, package-manifest.json, verification.json and SHA256SUMS.txt. English and Chinese notes are readable on the Release page. Check the ZIP's SHA-256 against SHA256SUMS.txt before extracting it. The ZIP contains the self-contained Skill, without project coordination records, Python virtual environments or the repository's demo font. MIT covers the code, original artwork and Skill; the repository-only Silkscreen font retains its separate SIL OFL 1.1 license.

2026-10-31 notes update: English and Chinese notes are now readable online, and their wording has been clarified. The two note attachments were removed. SHA256SUMS.txt lists only the ZIP, package manifest and verification record; those three files are unchanged.

2026-10-31 说明更新：中英文说明改为在线阅读，并修正文案。两份说明附件已移除；SHA256SUMS.txt 只列出 ZIP、包清单和验证记录，这三个文件没有变化。

中文说明

v1.2.0 — PNG 导入、局部修改与组件拼接

用 JSON 等文本规格文件或简短的 Python 示例脚本制作、修改像素图标、精灵和 UI；也可以将多张小图放进带名称的图集，或将独立绘制的组件拼成场景。输出为 PNG，Minecraft 风格可按需选用。

相对 v1.1.1 的变化

• 分析图片尺寸、用色、透明度、相连的像素区域和镜像差异，也可将局部像素显示为字符网格。报告限制逐项列出的数量，避免输出过长；检查图片时可指定背景，并复用已有图集布局。
• 无损导入静态 8 位 PNG：不减少颜色，保留全透明像素里的 RGB 值。可以导入整图，也可以按已核对的布局选取图集中某一格；替换这一格时，其余小图和布局不变。
• 核对图片修改前后的 RGBA 值，确认变化都在指定区域或蒙版内。拼接 PNG 组件时，可明确指定裁切范围、允许的重叠和必须填满的区域，并用带坐标的预览检查接缝。
• 附带图标、按钮状态、128×128 像素精灵和组件场景的示例脚本；另有左右两块独立绘制、木板纹理在接缝处对齐的桥梁示例。文档分别说明局部修改、文本分析、看图检查，以及在实际游戏或应用中检查的方法。
• 如果文件系统不支持创建硬链接，可在终端设置一次 PIXELART_PORTABLE_OUTPUT=1，让后续命令和示例脚本使用便携输出模式；单条命令也可单独指定。已有文件仍需明确使用 --force 才能覆盖。

兼容性与运行

支持 Python 3.10+、Pillow >=10.4,<13。requirements.txt 指定 Pillow 12.3.0，便于得到一致结果；发布时使用的环境是 Python 3.14.2 / Pillow 12.3.0。运行只需 Python 和 Pillow。

继续支持 JSON v1，原有 API 和命令的默认行为不变。基本绘图操作直接替换像素的 RGBA 值；blit 默认按透明度将新图叠加到原图上（source-over），其他支持的合成方式需单独指定。导入 PNG 后再导出，解码后的 8 位 RGBA 像素值保持一致，但压缩后的文件字节可能不同，也无法恢复原图层或绘制步骤。不接受动画和不支持的高位深图片。

解压 ZIP，在任务实际使用的 Python 环境安装包内 requirements.txt，然后从解压后的父目录运行：

python deterministic-pixel-art/scripts/pixel_art.py render --spec deterministic-pixel-art/assets/gem-16.json --out output/gem.png --preview output/gem-preview.png
python deterministic-pixel-art/scripts/pixel_art.py inspect --image output/gem.png --grid

更新 Skill 时，请把旧版备份移到 AI 不会扫描加载 Skill 的目录，避免同时加载两份。使用便携输出模式时，其他程序可能读到尚未写完的新文件。无论采用哪种输出模式，如果写入过程中程序崩溃或电脑断电，一组输出文件都可能只完成一部分；中断后请先检查文件是否齐全、内容是否完整，再继续使用。

验证结果与使用限制

v1.2.0 正式版已通过 165 项源码测试、106 项独立安装包测试和 43 张基准图片检查，并验证了重复打包结果一致、Skill 格式合规。包内 98 个文件中，97 个与已验收的开发包逐字节相同，另一个模块仅修改了版本号。verification.json 记录发布对应的提交、Windows/Ubuntu CI 结果和打包来源。

外部试用覆盖八类任务。记录中仍保留 LT04 最初画反方向、缺少早期版本，以及 LT08 早期过程记录不全的问题。这些试用不足以证明工具能普遍改善画质、缩短任务时间或节省 token。Minecraft 游戏内效果、其他字体的适配、512×512 像素的直接绘制和多个 AI 协作绘图仍未验证。自动检查通过后，仍需人工看图，确认图案含义和视觉效果符合要求。

分发内容

下载附件包括 deterministic-pixel-art-v1.2.0.zip、package-manifest.json、verification.json 和 SHA256SUMS.txt；中英文版本说明可在 Release 页面直接阅读。解压前请用 SHA256SUMS.txt 核对 ZIP 的 SHA-256。ZIP 内包含独立 Skill，不含项目协调记录、Python 虚拟环境或仓库演示字体。代码、原创图像和 Skill 使用 MIT 许可；仓库演示使用的 Silkscreen 字体另遵循 SIL OFL 1.1 许可。

v1.2.0-rc.2 — precise edits and focused platform trials

v1.2.0-rc.2 — precise edits and native asset workflows

Prerelease. Stable and installed local Skill remain v1.1.1.

• Visible silhouettes on light primary backgrounds, native target-background views, actionable output errors and explicit portable publication.
• Lossless named-atlas import, single-cell replacement and raw-RGBA allowed-region/mask verification. Existing source-over defaults and JSON v1 remain compatible.
• Bounded PNG component assembly, explicit coverage/overlap/crop contracts, coordinate-labeled seams, original native128 sprite and256 scene controls.
• Three small focused follow-ups: P1-W → WorkBuddy/DeepSeek, P1-C → Qwen office assistant, P1-V → Gemini task mode. Each contains this frozen Skill, original inputs, START_HERE and a command/failure recorder.

See the simple assignment/download guide. Qwen can download its task ZIP directly; no ZIP upload is required.

Local scripted rehearsals are packaging/contract checks, not external-agent results. Three new external runs remain pending. The earlier eight exploratory tasks are not a controlled A/B comparison; no universal aesthetic or token-saving claim is made. The256 tile control is crop/reassembly, not independently painted seamless chunks. Portable new files may be visible while writing.

Skill ZIP SHA-256: a34e5c58e420fd43702fe9b892fbbae2a32a57a47311e58f242dad140e9ded22.

Verified locally:80 source tests,43 unchanged stable images,61 isolated ZIP cases, official Skill validation and three scripted focused rehearsals. All eight CI jobs passed on 8932e7a95424cebfbc55dd047b18fe6acc8f46fb, the release source commit. The Skill ZIP exactly matches the previously frozen artifact.

Downloads include Skill/task ZIPs, a combined checksum manifest and detailed workflow/preparation reports. MIT is unchanged. Candidate checks and release-download verification are recorded in the preparation report.

中文：本候选新增精确小改、可见审阅和原生128/256范例。三个新会话按上表分工即可；失败也交回submission和会话。新外测尚未取得，不自动升正式版，也不替换本地稳定安装。

v1.2.0-rc.1 — text inspection and exact PNG editing

v1.2.0-rc.1 — text inspection and exact PNG editing

This is a prerelease. The latest stable release and local Codex installation
remain v1.1.1, which contains the external-feedback reliability fixes.

New capabilities

• inspect / inspect_image: bounded pixel geometry, palette, alpha coordinates,
  4/8-neighbor components, symmetry, named cells, exact local grids and comparisons.
• import / image_to_spec: static 8-bit PNG to editable JSON v1, preserving all
  decoded RGBA pixels, including partial alpha and hidden RGB. No quantization.
• Shared review analysis; versioned content-bound sidecars, --metadata reuse and
  optional --embed-metadata layout records. Existing JSON v1 and render semantics
  remain compatible.
• Short skill paths for small edits, text-only work and visual review; specialist
  guidance remains optional.

Verification and measured costs

61 source tests, 43 unchanged stable outputs, eight exact import/inspection tasks,
45 isolated Skill ZIP scenarios, the official skill-creator validator and 15
evaluation-bundle smoke checks pass locally. Windows/Ubuntu CI must pass on the
tagged commit. Across Pillow versions, compare decoded RGBA rather than PNG bytes.

One Windows/Python 3.14.2/Pillow 12.3.0 comparison with v1.1.0 found pure rendering
essentially unchanged. Full render/review medians rose by 15.0/16.7 ms (about 9%);
sample review JSON grew from 1,237 to 6,999 bytes. These are real runtime/context
costs. Agent productivity, correction-round savings, token savings and general
aesthetic improvement remain unproven.

External evaluation remains pending

The agent bundle contains A/B runtimes, eight fixed local-repair tasks, inputs,
requests and the command recorder. The separate evaluator bundle contains
answers, version mapping, verification runtime and an independent visual-review
template. Do not give that bundle to the tested agents.

Use fresh conversations for all eight A/B tasks in both a text-only and a
multimodal environment. Keep each environment's model, prompt and budget fixed;
save real commands, localization, specs, PNGs, transcripts and failure records.
Unknown token usage stays null. See README.md inside the agent ZIP.

Engineering checks alone cannot promote this candidate. The agreed gate also
requires complete paired results, no critical semantic/content regression, and
concrete localization or revision-count improvement in at least six task IDs
without opposing regression. There are currently zero new external runs;
the attached pending-gate report correctly returns false. Scripted smoke checks
and historical v1.0→v1.1 reports do not satisfy this gate.

Downloads include the Skill ZIP, per-file manifest, both evaluation ZIPs,
capability/performance/pending reports and SHA-256 sums. Keep the evaluator
download with the person coordinating the trials. No font or in-game acceptance
is claimed. MIT licensing is unchanged.

Full report: https://github.com/Chenamel/deterministic-pixel-art/blob/v1.2.0-rc.1/docs/audit-v1.2.md

中文说明

这是候选版，新增文本分析、PNG 无损导入和布局复用。61 项源码测试、43 张稳定
产物兼容、八项精确往返、45 项独立 ZIP 场景、官方 Skill 格式校验及 15 项测评包
运行检查通过。新增分析也有代价：完整命令约慢 15–17 ms，样例审阅 JSON 约为
原先的 5.66 倍；未证明节省 Agent 修改轮次、token 或普遍提高画质。

测评包和验收包分开下载，验收答案不交给被测模型。新版真实配对测评尚缺，
因此保留 RC，正式发布及本地 Codex Skill 均保持 v1.1.1。达到既定门槛后再转正式版。

Verified commit: a693d51de83a36d3cf8dacbfd3f53e16aa6c739f. All eight Windows/Ubuntu jobs passed: https://github.com/Chenamel/deterministic-pixel-art/actions/runs/39577610334

v1.1.1 — Reliability fixes

v1.1.1 — reliability fixes from external feedback

• Null mask outlines now equal omission; explicitly transparent outlines retain their replacement semantics.
• Native, grayscale and cell review cards use the selected primary background.
• Wide sources use adaptive sheets; very large views expose labeled native crops and an overview.
• Output groups are staged and reserved before single-file atomic publication, with rollback on ordinary errors.
• Bounds, DSL, missing dependencies and deep JSON have actionable diagnostics.
• Pillow 10.4/11.3/12.1 compatibility and POSIX venv invocation are covered alongside the 12.3 baseline.

Local verification: 42 source tests, all 43 stable migration pixels unchanged,
37 isolated ZIP execution cases and the official skill-creator validator pass.
Windows/Ubuntu CI results must also pass before publishing this release.
PNG bytes across different Pillow/zlib environments are not promised identical.

Details: feedback report. These are reliability changes,
not a new claim of general aesthetic improvement. Crash-safe multi-file
transactions, fonts and in-game verification remain outside this patch.

Downloads include the standalone Skill ZIP, per-file manifest and SHA-256 sums.
Keep the existing editable specifications; no artwork redesign is required.

All eight Windows/Ubuntu CI jobs passed: https://github.com/Chenamel/deterministic-pixel-art/actions/runs/39567484126

v1.1.0 — Pixel art direction and contextual review

Editable art direction and contextual review for deterministic pixel assets.

• Add review contact sheets and advisory JSON: native/enlarged views, backgrounds, grayscale, silhouettes, named atlas cells, declared contrast checks and visible differences.
• Add eight original editable studies and task-specific pixel-art/UI/Minecraft guidance.
• Author self-review: six preferred, two unchanged; small unblinded sample, no independent or in-game evaluation. More detail can increase colors and drawing operations.
• Preserve all 43 v1.0.0 migration outputs, JSON v1 and existing public interfaces/commands.
• Pass 33 unit tests, 37 isolated package scenarios and the official Skill validator. Runtime remains Python + Pillow.

Download deterministic-pixel-art.zip, verify it against SHA256SUMS.txt, and extract the skill directory into your local Codex skills directory. Keep an old installation backup outside that directory. No tutorial images or Minecraft vanilla assets are bundled; MIT remains unchanged.

English report · 中文报告 · Source and paired images

v1.1.0-rc.1 — Visual guidance and contextual review

Editable art direction and contextual review for deterministic pixel assets.

• Add review contact sheets and advisory JSON: native/enlarged views, backgrounds, grayscale, silhouettes, named atlas cells, declared contrast checks and visible differences.
• Add eight original editable studies and task-specific pixel-art/UI/Minecraft guidance.
• Author self-review: six preferred, two unchanged; small unblinded sample, no independent or in-game evaluation. More detail can increase colors and drawing operations.
• Preserve all 43 v1.0.0 migration outputs, JSON v1 and existing public interfaces/commands.
• Pass 33 unit tests, 37 isolated package scenarios and the official Skill validator. Runtime remains Python + Pillow.

Download deterministic-pixel-art.zip, verify it against SHA256SUMS.txt, and extract the skill directory into your local Codex skills directory. Keep an old installation backup outside that directory. No tutorial images or Minecraft vanilla assets are bundled; MIT remains unchanged.

English report · 中文报告 · Source and paired images

v1.0.0 — verified portable baseline

The existing verified renderer and portable Codex skill, now pinned as a release.

• Python + Pillow runtime, JSON v1 and compact Minecraft GUI DSL.
• 24 tests, 43 historical regression outputs and 19 packaged execution checks.
• This release records the existing commit; it does not include the later visual-quality workflow.
• MIT license. Download the standalone skill ZIP and verify it with SHA256SUMS.txt.