1.3 Quill
在 Unity Editor 内使用大模型编写 Shader、分析 C# 错误，并通过人工审阅、文件快照和回滚应用修改。
- Unity 2022 或更新版本，推荐 Unity 2022.3 LTS。
- 可访问模型服务的网络，以及对应服务商的 API Key；使用 Ollama 时需自行启动本地服务。
- 插件仅在 Editor 运行，不进入游戏运行时。
| 功能 | 使用方式 |
|---|---|
| Shader / C# 对话 | 描述需求，附加相关文件，获取完整代码 |
| Unity 上下文 | 可选发送当前场景、渲染管线、选中材质和近期错误 |
| 文件审阅 | 将最新回复的第一个代码块送入独立审阅窗口 |
| 文件写入 | 确认目标路径与内容后写入 Assets 内文件 |
| 持久快照 | 写入前保存原文件及其 .meta |
| 回滚 | 恢复已有文件；撤销新建文件及 .meta |
| 自定义接口 | 配置 Chat Completions 兼容地址与模型 ID |
- 下载仓库 ZIP 或克隆仓库。
- 将 Assets/Editor/AIAssistantWindow.cs 和 Assets/Editor/QuillRollback.cs 一起复制到工程 Assets/Editor。
- 等待 Unity 编译，打开 Tools > Shader-CodeAgent4Unity > Open。
- 升级旧版本时覆盖对应文件，避免在其他目录保留同名类的副本。
当前采用源码安装方式，不是 Unity Package Manager 包。
打开 Settings，选择 Provider，填写 Endpoint、Model 和 API Key。Model 字段始终可以手动编辑，应填写服务商实际接受的 API ID。
| 服务 | 配置 |
|---|---|
| DeepSeek | 预设 deepseek-V41-flash、deepseek-v4-pro；地址 https://api.deepseek.com/chat/completions |
| GLM | 预设 glm-5.3；地址 https://open.bigmodel.cn/api/paas/v4/chat/completions |
| OpenAI | 使用支持 Chat Completions 的模型 ID；地址 https://api.openai.com/v1/chat/completions |
| Claude | 原生 Messages 接口；预设 claude-fable-5-1、claude-opus-5-5、claude-sonnet-5-5 |
| Ollama | 默认 http://localhost:11434/v1/chat/completions，填写已安装的模型 ID |
| Custom | 自行填写完整 Chat Completions endpoint、模型 ID 和 Key |
Claude 使用 Messages；其他预设使用 Chat Completions，不包含 Responses/Codex 专用协议适配。供应商是否授权某个模型，以账号和接口实际响应为准。Gemini 和 Qwen 不提供内置预设。
Detect endpoint compatibility 会发送一次小型付费测试请求，并在对话中显示结果。API Key 保存在本机 Unity EditorPrefs；可通过 Forget saved API key 删除插件保存的值。
- 打开插件，配置模型。
- 可用 Attach file 附加现有 Shader；需要时启用渲染管线与场景上下文。
- 示例需求：为 Unity 2022.3 Built-in 管线编写一个青色 Unlit Shader，返回单个完整 ShaderLab 代码块。
- 点击 Send，检查回复，然后点击 Review latest generated code。
- 在 Absolute Assets path 填写完整路径，例如 D:/MyProject/Assets/Shaders/Cyan.shader。
- 审阅全部代码，点击 Apply with snapshot 并确认。
- 等待 Unity 编译，在 Console 检查错误。创建材质并绑定 Shader 后，在自己的场景验证显示效果。
URP/HDRP 请求应明确 Unity、渲染管线版本及目标平台。Shader 能导入不等于所有管线和 GPU 上渲染正确。
使用 Attach file 附加相关脚本，点击 Grab errors 收集错误，描述复现步骤。要求模型返回完整单文件代码，再在审阅窗口指定原脚本路径。当前只提取最新回复中的第一个代码块；多个文件应分别审阅和应用。
打开 Tools > Shader-CodeAgent4Unity > File review and rollback，点击 Rollback latest 并确认。
- 每次写文件前保存快照，包括已有文件和 .meta 的原始字节。
- 恢复最新尚未撤销的快照；新建文件及其 .meta 会被删除。
- 快照保存在当前工程 Library/QuillSnapshots，编辑器重启后文件仍在；删除 Library 会丢失快照。
- 回滚会覆盖目标文件后续的人工修改；它不是三方合并或完整工程备份。
- 新建目录可能保留；场景和外部文件操作不在这套回滚范围内。
在独立 Unity 2022.3.17f1c1 工程执行：
- DeepSeek Flash 真实 API 生成 Shader，Unity 导入检查通过。
- 修改后的文件及 .meta 回滚，逐字节比对通过。
- 新建文件及 .meta 回滚删除通过。
- 插件 C# 编译检查通过。
尚无窗口录屏 GIF；未完成所有渲染管线、当前用户场景及所有 Provider 的端到端测试。
- 当前是人工审阅后写单文件的助手，尚未实现自主多步执行、Shell、场景或系统环境修改（以后会弄？）。
- 仅允许 Assets 内目标文件，不支持链接目录/文件；生成代码仍需人工审阅（咱就是说也别太懒）。
- 发送的上下文及附加文件会传给所选服务商，API 调用按供应商规则计费。
- EditorPrefs 不是加密密钥保险库。不要提交 Key、个人配置、请求日志或私有工程内容。
- 建议使用 Git 备份工程，并先在测试工程试用。
本项目采用 MIT License
- 工具菜单和窗口名称统一为 Shader-CodeAgent4Unity。
- 新增生成代码审阅入口、Assets 文件写入、持久快照及回滚。
- 更新 DeepSeek 和 GLM 预设，允许手动填写模型 ID。
- 修复 Windows 路径格式检查、对话上下文位置和接口检测流程。
- 增加本机已保存 Key 的删除按钮。
-
更新模型：OpenAI 的 gpt-6-astra、gpt-6.1-sol、gpt-6-luna；Claude 的 claude-fable-5-1、claude-opus-5-5、claude-sonnet-5-5。
-
恢复可选择的模型预设，修复下拉框首项选择问题；支持从供应商刷新账号可用模型列表。
-
增加 Claude 原生 Messages 请求、鉴权和文本响应解析。
-
Unity 2022.3 编译验证通过。OpenAI/Claude 未使用付费 Key 完成端到端调用验证；截图确认的是模型 ID，不代表所有模型支持当前 Chat Completions 接口。Responses 适配仍未实现。
1.3 Quill · Unity 2022+ · MIT License
Write Shaders, analyze C# errors, and apply changes with LLMs inside the Unity Editor — with human review, file snapshots, and rollback.
- Unity 2022 or newer; Unity 2022.3 LTS recommended.
- Network access to your model provider, plus the corresponding API key. For Ollama, start the local service yourself.
- The plugin runs in the Editor only; it is not included in game runtime.
| Feature | How to use |
|---|---|
| Shader / C# chat | Describe what you need, attach relevant files, get complete code |
| Unity context | Optionally send the current scene, render pipeline, selected materials, and recent errors |
| File review | Send the first code block of the latest reply to a dedicated review window |
| File write | Write to files under Assets after confirming the target path and content |
| Persistent snapshots | Save the original file and its .meta before writing |
| Rollback | Restore existing files; undo newly created files and their .meta |
| Custom endpoint | Configure a Chat Completions–compatible URL and model ID |
- Download the repository ZIP or clone the repository.
- Copy both
Assets/Editor/AIAssistantWindow.csandAssets/Editor/QuillRollback.csinto your project'sAssets/Editor. - Wait for Unity to compile, then open Tools > Shader-CodeAgent4Unity > Open.
- When upgrading from an older version, overwrite the corresponding files and avoid keeping copies of the same class in other directories.
This is currently a source-based install, not a Unity Package Manager package.
Open Settings, choose a Provider, and fill in Endpoint, Model, and API Key. The Model field is always editable and should contain the actual API ID accepted by the provider.
| Service | Configuration |
|---|---|
| DeepSeek | Presets deepseek-V41-flash, deepseek-v4-pro; endpoint https://api.deepseek.com/chat/completions |
| GLM | Preset glm-5.3; endpoint https://open.bigmodel.cn/api/paas/v4/chat/completions |
| OpenAI | Use a model ID that supports Chat Completions; endpoint https://api.openai.com/v1/chat/completions |
| Claude | Native Messages API; presets claude-fable-5-1, claude-opus-5-5, claude-sonnet-5-5 |
| Ollama | Defaults to http://localhost:11434/v1/chat/completions; enter an installed model ID |
| Custom | Provide the full Chat Completions endpoint, model ID, and key yourself |
Claude uses Messages; the other presets use Chat Completions and do not include Responses/Codex-specific protocol support. Whether a provider authorizes a given model depends on your account and the actual API response. Gemini and Qwen have no built-in presets.
Detect endpoint compatibility sends one small paid test request and shows the result in the chat. The API key is stored in the local Unity EditorPrefs; you can delete the value saved by the plugin via Forget saved API key.
- Open the plugin and configure a model.
- Optionally use Attach file to attach an existing Shader; enable render pipeline and scene context if needed.
- Example request: Write a cyan Unlit Shader for the Unity 2022.3 Built-in pipeline, returning a single complete ShaderLab code block.
- Click Send, review the reply, then click Review latest generated code.
- In Absolute Assets path, enter the full path, e.g.
D:/MyProject/Assets/Shaders/Cyan.shader. - Review all the code, click Apply with snapshot, and confirm.
- Wait for Unity to compile and check the Console for errors. Create a material, assign the Shader, and verify it in your own scene.
URP/HDRP requests should explicitly state the Unity version, render pipeline version, and target platform. A Shader that imports successfully is not guaranteed to render correctly on every pipeline and GPU.
Use Attach file to attach the relevant scripts, click Grab errors to collect errors, and describe the reproduction steps. Ask the model to return complete single-file code, then specify the original script path in the review window. Currently only the first code block of the latest reply is extracted; review and apply multiple files separately.
Open Tools > Shader-CodeAgent4Unity > File review and rollback, click Rollback latest, and confirm.
- A snapshot is saved before every file write, including the original bytes of existing files and their .meta.
- Restores the most recent snapshot that has not yet been undone; newly created files and their .meta are deleted.
- Snapshots are stored in the current project's
Library/QuillSnapshotsand survive an editor restart; deleting Library loses the snapshots. - Rollback overwrites any manual changes made to the target file afterward; it is not a three-way merge or a full project backup.
- Newly created directories may remain; scene and external file operations are outside the scope of this rollback.
Executed in a standalone Unity 2022.3.17f1c1 project:
- Shader generated via the real DeepSeek Flash API; Unity import check passed.
- Rollback of a modified file and its .meta, byte-for-byte comparison passed.
- Rollback and deletion of a newly created file and its .meta passed.
- Plugin C# compile check passed.
No window recording GIF yet; end-to-end tests across all render pipelines, the current user's scene, and all providers are not complete.
- This is currently a human-reviewed, single-file writing assistant. Autonomous multi-step execution, Shell, and scene or system-level modifications are not implemented yet (maybe later?).
- Only target files under Assets are allowed; linked directories/files are not supported. Generated code still requires human review (come on, don't be that lazy).
- Sent context and attached files are transmitted to the selected provider; API calls are billed according to the provider's rules.
- EditorPrefs is not an encrypted key vault. Do not commit keys, personal configuration, request logs, or private project content.
- Back up your project with Git, and try it in a test project first.
This project is licensed under the MIT License.
- Unified the tool menu and window name to Shader-CodeAgent4Unity.
- Added a generated-code review entry point, Assets file writing, persistent snapshots, and rollback.
- Updated the DeepSeek and GLM presets and allowed manual model ID entry.
- Fixed Windows path format checks, chat context position, and the endpoint detection flow.
- Added a button to delete the locally saved key.
- Updated models: OpenAI's
gpt-6-astra,gpt-6.1-sol,gpt-6-luna; Claude'sclaude-fable-5-1,claude-opus-5-5,claude-sonnet-5-5. - Restored selectable model presets and fixed the dropdown first-item selection issue; added support for refreshing the account's available model list from the provider.