v0.1.6-preview · 设置与 API 兼容优化、自然转头

本次更新

• Windows 解压后可直接双击「启动游戏.cmd」，自动选择空闲端口，并改善配置重启和退出体验。
• 重新整理 AI 配置与高级设置，完善获取模型、测试连接、思考控制和实际对话失败的提示；输入框支持右键编辑。
• 自定义 OpenAI 兼容接口支持基于已确认能力的 JSON 格式和请求参数适配，继续保留严格的结构化决策、预算与行动校验。
• 玩家交谈时平滑转头看向 NPC，结束后恢复，兼容默认角色和导入 VRM。
• 改善玩家设定、NPC 人设、世界知识、记忆查找和操作说明；修复部分连续回复的 JSON 解析问题。
• 包内新增离线配图「使用说明.html」，介绍 DeepSeek 官方配置、基础操作和常见问题。

开始试玩

完整解压 ZIP → 双击「启动游戏.cmd」→ 设置中打开「AI 配置」→ 保存并按提示重启 → 进入咖啡馆。

无需另装 Python 或 Unity。已内置 EmbeddingGemma 记忆查找模型；主聊天模型和个人 API Key 需自行配置。DeepSeek 官方接口已有定向真实验证，其他接口及中转模型的兼容性取决于具体服务和模型；简短连接测试不代表完整 NPC 对话兼容。

发布验收使用隔离数据与 Mock，未额外发起付费模型请求。详细验证范围及已知测试限制见附件 RELEASE_VALIDATION.md，文件校验见 SHA256SUMS.txt。

English: Improved portable startup, AI settings and API compatibility, clearer error feedback, natural player head gaze, and an offline illustrated Chinese player guide. The package includes Python and the local embedding model; bring your own chat model credentials.

EmbodiedNPC v0.1.5-preview

EmbodiedNPC v0.1.5-preview

Windows x64 preview of the single-player, four-NPC café Demo. Source, tag and Windows build use the same public commit: 15bad1bb98796e6c155d7c91cd3115ff7cd28ce3.

What's changed since v0.1.4-preview

• Local ambient activity: Vivi walks, pauses, observes and wipes inside the counter area. Wen and Jiang share the seated drinking/cup-follow flow with independent timing. Ambient presentations stay in Unity and do not repeatedly invoke an LLM or create cleaning/memory facts.
• Presentation handoff: existing DailyBehavior/mode admission and MotionChannelCoordinator own revocable Ambient instances. Formal work preempts them; conversation can finish a cup return or pause the selected counter destination. Disable and cancellation restore borrowed props and animation state.
• Player facial expression: a subtle base smile, with blended wave, cheer and proud expressions that recover to the base expression; blinking remains separate.
• Snacks: croissant / 牛角包, macaron / 马卡龙, cupcake / 小蛋糕 and pudding / 布丁. One type, 1–4 portions per request. Vivi uses the configured display-case or food-rack approach and a short reach animation. A successful authoritative result triggers the existing service feedback. Deferred snack requests reuse the current typed task flow. Food service is logical: no food pickup entity, inventory or simulated stock is created.
• Left Alt cursor toggle: release/recapture the pointer while respecting dialogue/settings focus and the existing ESC controls.
• Memory and relationship refinements: attributed shared-conversation memories remain distinct from verified world events. Episode-level social interpretation and bounded persistent deltas are separate, with source checks and atomic updates. Recent dialogue gains source-clock distances; the shared built-in embedding session warms before Backend readiness.
• Release preparation fixes: the existing portable assembler now includes the byte-identical shared snack catalog. Public-only scene cleanup excludes orphan bones from previously removed assets, and test screenshot paths no longer depend on a developer username. Public first-import model recovery and existing license/resource boundaries are preserved.

These changes are in the public source diff from 49c6fde through the candidate; the candidate has public 910485a as its sole parent. Private development history, InternalDocs, WIP and test data are not published.

Start and configure

1. Extract the complete Windows ZIP to a new writable folder and run Start.cmd. No Unity Editor, separate Python install or pip setup is needed.
2. Configure your own Main LLM/API key, or run your own supported local model service. Main chat-model weights and credentials are not bundled. The package includes the pinned EmbeddingGemma-300M Q4 ONNX resources and their separate license/terms.
3. Follow the in-game save/apply/restart flow and readiness prompts. Starting the application does not mean the AI service is configured or reachable.
4. WASD moves, mouse looks, E interacts, G opens player actions, Left Alt releases/recaptures the cursor, and ESC retains the existing menu/back behavior.

Fresh preview data: this release is validated with new UserData. Do not copy old Preview UserData into it; cross-preview save compatibility/migration is outside this release. Keep old installations separately if you want to retain them.

Validation and limits

• Backend comparison: 605 tests, 568 passed, 12 failures and 25 errors. All 37 nonpassing test IDs also fail on the previous public baseline; no new nonpassing IDs were found. The focused new-feature suite passed 64/64. This is not an all-green full suite.
• Unity: 106 of 111 unique cases demonstrated passing across isolated runs; the remaining five older UI-test failures were reproduced with the same messages on v0.1.4-preview. The initial combined run was 74/111, followed by 30/30 isolated Task cases and 2/2 Ambient evidence reruns. No new unresolved failure remains in that comparison.
• Windows x64 build: 0 errors, 58 warnings. Public first-import model recovery and scene/resource tests were exercised.
• Final ZIP: CRC passed, all 3,141 extracted file hashes matched, and package privacy scans found no bundled UserData, live databases, credentials or developer logs/caches.
• Actual packaged cold start: extracted to a new folder containing spaces and Chinese characters; launched the extracted Start.cmd, started its own packaged Backend, checked health/runtime status and 161 knowledge entries, and visually checked the standalone title/configuration UI. With no saved model config, the UI correctly requested AI configuration. Closing the Player exited with code 0 and stopped its Backend; original development services remained untouched.
• At the user's request, packaged acceptance stopped at startup: no additional standalone café playthrough, control/animation exercise or configured embedding-index run was performed. Café/animation/control coverage above comes from Unity tests and observations, not a full packaged gameplay pass.

Real external-provider dialogue, natural-language snack ordering and model-to-delivery end-to-end acceptance are NOT RUN for this release. Scripted/Mock checks prove only their stated interfaces and local behavior, not real-model semantics. No paid provider load test was performed. Model availability, fees and output quality depend on your selected service.

Default knowledge now contains 161 entries. Old/custom knowledge is not silently overwritten. The NPCs have limited autonomy and legal actions; this remains an early prototype. This is an extracted Windows check on the development machine, not certification on a new physical PC or every GPU/provider.

Original code is Apache-2.0; imported and project-scoped art/model licenses remain separate. The complete Demo is not an unrestricted commercial asset bundle. See THIRD_PARTY_NOTICES.md and the package's license/terms files.

v0.1.4-preview

v0.1.4-preview

A Windows x64 preview focused on everyday conversation, consistent reply presentation and bounded foreground/background concurrency.

• Expanded the default world-knowledge catalog from 108 to 156 entries: 12 new private everyday backgrounds for each of the four NPCs. Ordinary-chat guidance reduces repetitive follow-ups and unrelated ordering prompts without changing action authority.
• Unified overhead/history paragraph splitting, with at most three bubbles. Overhead segments arrive two seconds apart and remain for 20 seconds; history renders immediately and retains one underlying dialogue record.
• Retained order dialogue can carry its validated historical GoalOutcome into the actual model context, including after the ordinary 900-second result window. Dialogue and outcome are budgeted together. Historical completion does not replace current task state or prove an object's present location.
• Shared, configurable bounded model scheduling: cloud 6/6/2; Ollama 1/1/1. Independent Episodes, distinct relationship pairs and independent maintenance scopes can run concurrently, with active-attempt protection, ordered watermarks and drained shutdown.
• Settings > Developer tools > Concurrency settings reuses profile validation, saving and restart activation.

Getting started / 使用

Extract the complete ZIP, run Start.cmd and configure your own Main LLM. No separate Python or Unity installation is needed. Built-in EmbeddingGemma Q4 and its complete terms are bundled; Main LLM weights and credentials are not.

完整解压后运行 Start.cmd，配置自己的主模型。设置 → 开发者工具 → 并发调整可修改并发上限；保存后按提示重启。运行数据保存在包内 UserData，更新前请正常退出并备份。

Existing knowledge catalogs are preserved rather than merged with defaults. To use the expanded catalog, export your custom knowledge first and explicitly restore defaults; this replaces the catalog.

Scope

Single-player CafeDemoV2 with four NPCs. Model wording, recall and reflection can still be imperfect. Historical outcome linking requires retained source dialogue and does not guarantee cross-session order recovery. Asset/model licenses remain separate; see LICENSES.md and the Licenses directory. Local extracted-Windows validation does not certify a second physical machine or every Provider.

EmbodiedNPC v0.1.3-preview · 世界知识、混合检索与体验导出

EmbodiedNPC v0.1.3-preview

本版是 Windows x64 Pre-release，相比 v0.1.2-preview 更新：

• 完整可编辑世界知识库、108 条默认知识，支持公共/指定 NPC 权限、常驻/按需模式、JSON 完整替换导入、差异预览、导出和恢复默认。
• 世界知识与长期记忆混合检索，按需补查，最多两次认知推理；改善计划与已发生事件、生成文本与实际呈现对话的区分。
• 十种热饮：拿铁、卡布奇诺、美式、浓缩、馥芮白、摩卡、焦糖玛奇朵、热巧克力、抹茶拿铁、红茶。
• 更有变化的 NPC 对话动作和面部表情，以及偶尔出现的开心、疑惑、惊讶、思考辅助图标。
• 玩家设定、角色外观与知识管理体验改进；新增“导出全部体验数据”，在咖啡馆设置中生成只读、当前玩家范围的分析 ZIP。
• 修复标题页尚未绑定玩家身份时，完整体验导出显示异常的问题；现在会提示先进入咖啡馆。

下载与启动

下载 EmbodiedNPC-Windows-x64-v0.1.3-preview.zip，完整解压后双击 Start.cmd。无需安装 Unity Editor 或另外安装 Python；已附 Portable Backend、EmbeddingGemma-300M Q4 ONNX 及其独立许可。主 LLM 需要自行配置，包内没有开发者的 API Key、玩家存档或模型服务凭据。

个人设置、记忆和关系保存在包内 UserData。正常退出后再备份/移动整个 UserData；已有知识目录不会自动合并新版默认内容，可自行导入或恢复默认。不要将含凭据的 UserData 当作发行包分享。

验证与限制

正式版本 Windows x64 非 Development 构建成功。已实际验证干净解压、Start.cmd/标题页、模型配置页面、CafeDemoV2、Backend 连接、知识/Memory 初始化、Windows Player 完整体验 ZIP 及正常退出/重开。具体边界见随附 RELEASE_VALIDATION.md。

这仍是早期预览：模型可能出错，记忆和检索不是事实正确性的保证，不同 VRM 的表情能力也不同。此次接线验证使用本机受控 Provider，未重复付费模型 A/B，也不是全新物理 Windows 机器的认证。

完整体验 ZIP 包含玩家主动输入的聊天、人设、知识和记忆等文本，可能涉及个人信息。请自行检查并决定是否分享；游戏不会自动上传，API 凭据不在体验导出范围。

原创代码 Apache-2.0；模型、美术、动画、字体、音频按独立许可。星芒维持项目/Fork 范围内非商业许可，EmbeddingGemma 遵守 Gemma 条款。详见包内 LICENSES.md、Licenses 和源码 THIRD_PARTY_NOTICES.md。

---

Windows x64 pre-release with editable world knowledge, hybrid memory/knowledge retrieval, ten hot drinks, richer dialogue presentation, four occasional emotion icons and complete player-scoped experience export. Extract everything and run Start.cmd; no Unity Editor or separate Python install is required. Configure your own Main LLM. Built-in Embedding and its terms are included; credentials and player saves are not. Back up UserData after normal shutdown. Review authored text in experience exports before voluntarily sharing it. Model errors and VRM expression differences remain possible; see the validation note for the checks actually completed.

v0.1.2-preview · 动作 P0 与自然待机修复

本版整合已完成人工验收的动作 P0 修复，仍为早期实验性 Preview。

• 改善 NPC 与玩家基础待机姿态，降低不自然的循环起伏。
• 双臂更自然下垂，手掌放松、轻微虚握；保留薇薇裙摆避让和顾客坐姿/脚部 IK。
• 玩家 Wave 使用手臂专用层；非挥手侧按约 15% 自然跟随，减少额外躯干旋转。Cheer/Proud 保持原有层。
• 保留 CafeDemoV2、对话、世界任务结果与长期记忆流程；不包含运行时 VRM 导入等未来功能。

**运行：**下载 Windows x64 ZIP，完整解压到可写目录，双击根目录 Start.cmd。便携 Backend 自动启动；正常退出游戏后自动停止。主模型仍需自行配置，本包不含主模型或 API Key。

完整内置 EmbeddingGemma-300M Q4 ONNX，首次默认本地 Embedding，无需另下权重。外部 Embedding 与关闭向量选项保留。模型许可证、Gemma 条款/政策、Notice 和转换来源随包提供；其他素材授权及星芒项目内非商业许可继续适用，见 LICENSES.md。

本次主项目及目标发布工程分别通过 76/76 项针对性 Unity 测试；新 Player 构建 0 错误（55 警告）。新 ZIP 已完成 CRC/逐文件哈希校验，并实际验证 Start.cmd、标题进入 CafeDemoV2、本地无密钥配置的 UI 保存与受控重启、内置模型 768 维推理和正常退出清理。

验证边界详见随附 RELEASE_VALIDATION.md：本轮没有重复完整五人多视角实玩，也未确认 Standalone 的自动按键 Wave/交谈冒烟；最终姿态采用用户已验收 R4。Backend 针对性用例 101/104 通过，另 3 项旧夹具不匹配，不称全套回归全绿。未做陌生物理电脑或所有云端模型验收。

English: Refines quiet idles, relaxed arms/hands and an arms-only player wave with restrained support-arm motion. A freshly built Windows player and portable backend include the complete pinned local EmbeddingGemma Q4 resources and licenses. Configure your own Main LLM. See the attached validation scope; this is an early preview, not a clean-machine/full-gameplay certification.

EmbodiedNPC v0.1.1-preview — Early Prototype

更新 / Changes

• 世界知识从 25 条扩充至 85 条，保留已有设定和原有访问范围。
• 补充四名 NPC 的工作、学习、兴趣、独立生活及共同经历，保持原有身份与关系。
• 用现有 KayKit CC0 Idle_A 改善站姿；对话与挥手层只控制手臂，避免躯干后仰。
• 关闭薇薇 Locomotion 的 Foot IK，修正待机和步态失真；保留 Walk 片段、播放速度和移动逻辑。

World knowledge expands from 25 to 85 entries, with richer everyday backgrounds for all four NPCs. This release also improves idle poses, isolates dialogue/wave animation to the arms, and disables Vivi's distorting Foot IK without replacing the existing walk clips or changing movement speeds. The broad talking-arm animation uses a relaxed idle fallback; facial expressions, nods and waving remain available.

下载与启动 / Getting started

下载 Windows x64 ZIP，完整解压后运行 Start.cmd。内置 EmbeddingGemma-300M Q4 ONNX，正常情况下无需下载或配置 Embedding；仍需配置自己的主模型 API 或本地推理服务。外部 Embedding 和关闭向量能力选项保留。

Download and extract the complete Windows x64 ZIP, then run Start.cmd. EmbeddingGemma-300M Q4 ONNX is bundled for memory vectors. You must still configure your own Main LLM API or local inference service. External Embedding and disabled-vector options remain available. See README_zh-CN.md / README_en.md inside the ZIP.

这是仍在优化架构和交互体验的 Early Prototype，模型回复和记忆可能出错。原创代码采用 Apache-2.0；星芒采用项目内非商业许可，其他第三方素材和模型遵循各自许可，详见随包许可文件。

This is an experimental early prototype. Model responses and memory can be unreliable. Original code is Apache-2.0; Xingmang has a project-scoped non-commercial license, and other assets/models retain their own licenses. See the bundled license files.

反馈 / Feedback: GitHub Issues or 952650555@qq.com.

EmbodiedNPC v0.1.0-preview — Early Prototype

EmbodiedNPC 的第一版 Early Prototype，供研究、体验与交流。底层架构和交互体验仍在持续优化，功能和设计可能调整。

Windows x64：下载 EmbodiedNPC-Windows-x64-v0.1.0-preview.zip，完整解压后运行 Start.cmd。包含新的 Unity Player、Python Backend 和 EmbeddingGemma-300M Q4 ONNX；正常情况下无需额外配置 Embedding。主模型 Main LLM 仍需自行准备和配置。 包内提供中英文说明。

首次启动链路、标题页/配置入口与内置 Embedding 加载已检查；后续咖啡馆和玩法验收由项目所有者进行，本次不宣称完整游戏测试或干净机器认证。

代码采用 Apache-2.0；第三方素材遵循各自许可，星芒采用项目内非商业许可；内置模型遵循 Gemma 条款及使用限制。详见包内 LICENSES.md 与各项许可。

---

The first Early Prototype of EmbodiedNPC, shared for experimentation, research and feedback. Architecture and interaction improvements are ongoing; features and designs may change.

Windows x64: download the ZIP, extract it completely, and run Start.cmd. It includes a newly built Unity Player, Python Backend and EmbeddingGemma-300M Q4 ONNX. No separate Embedding setup is normally required. You must still provide and configure a Main LLM service. Chinese and English quick-start guides are included.

The startup chain, title/settings entry point and bundled Embedding loading were checked. Further cafe/gameplay acceptance is being handled by the project owner; this is not a full-gameplay or clean-machine certification.

Original code is Apache-2.0. Third-party assets retain their licenses; Xingmang is project-scoped and non-commercial. The bundled model is subject to the Gemma terms and use restrictions. See LICENSES.md and the included licenses.

Feedback: GitHub Issues / Pull Requests, or 952650555@qq.com.