Cardwright

中文 · English

Cardwright 是面向 Windows 本地项目的 Agent 工作台：连接你自己的模型网关，让 Agent 完成任务，并在同一个界面里审阅修改、运行检查、管理小队。0.8 起另有独立的制卡工坊：跟着规划 AI 一步一步，把资料做成 SillyTavern 角色卡。

界面保持安静，确定性的规则由程序执行，技能和补充上下文按需提供给模型。所有模型请求都发往你自己配置的网关；密钥存在本机，由 Windows 凭据加密保护。

• Windows 10 或 11（x64）。
• 一个你自己的模型网关：OpenAI Chat Completions、OpenAI Responses 或 Anthropic Messages 协议都可以。
• 可选：Brave Search 或 SearXNG 的联网搜索凭据（与模型凭据分开保存）。

从 Releases 下载 Cardwright-Setup-1.3.2.exe 并安装。安装包没有代码签名，Windows 首次运行会提示「未知发布者」，可以选择「更多信息 → 仍要运行」。卸载时保留你的资料目录。

也可以自己构建：

npm install
npm run build
node scripts/package.mjs      # 便携版输出到 release/<版本>/Cardwright-win32-x64

1. 在「工作室设置 → Agent 与模型」添加网关，填写协议、Base URL 和密钥；从自动读取的列表里选择模型，或手动添加 ID。
2. 添加项目目录，在输入框里选网关、模型、上下文窗口与权限，然后发送任务。
3. 默认模式下，审阅实际路径和命令后再批准操作。受限命令默认不能联网，必要的本机命令需要另外批准。
4. 打开任务标题旁的侧栏面板，查看文件、改动、检查、终端、浏览器、计划和待办；面板可以拆成上下两块。
5. 在「工作台与验证」里配置项目检查命令。执行结束后，交付卡会分别显示文件修改和验证结果。

• 左栏是会话列表：按项目分组，可以置顶、改名、归档、搜索；每条有状态点（运行中、待批准、已完成、失败、输出被截断）。
• 中间是对话，保持阅读宽度。Agent 切换工作阶段时会标一条章节线，页边有圆点目录。回复或工具行里的 路径:行号 可以点开，右栏直接定位到那一行。
• 右栏是可拆分的面板：文件、改动、检查、终端、浏览器、计划、待办。窗口窄于 960px 时左栏收成图标列，右栏变成覆盖在对话上的抽屉。
• 斜杠命令：/clear、/compact、/context、/cost、/model、/resume、/rewind、/init、/help、/plan、/todos、/memory、/dream，以及你的技能。
• Shift+Tab 在默认审批 → 项目内自动编辑 → 计划 → 完全访问之间循环。
• 子代理：自建的和用户目录里的 .claude/agents/*.md 默认可用；项目文件夹里读到的默认关闭，确认内容可信后在设置里打开。打开的子代理主代理都能派，文件里的 model: 不起作用，在设置里给它选模型。
• 小绘：内置 AI 叫小绘。问它用的是什么模型，它会说出你配置的模型和网关，并说明网关后面实际接的是哪个它核实不了。说话的性格可以在「头像与外观」里关掉。
• 钩子：按 Claude Code 的 settings.json 格式配置 SessionStart、UserPromptSubmit、PreToolUse、PostToolUse、Stop、SubagentStop、Notification，可以从 Claude Code 导入并逐条试跑。
• 对话（工作台和制卡工坊都是）：代码块有语法高亮、语言标签和复制，长代码自动折叠；模型的思考过程边想边显示，想完折成「思考了 N 秒」；连续的同类工具调用合成一行；回复末尾有光标，新字淡入，滚动平滑跟随；表格有斑马纹，太宽时横向滚动。系统开了「减少动画」时这些都不动。
• 新版本提醒：每天查一次 GitHub 上的最新版本号，有新版本时在顶栏提示，点一下打开发布页；不下载、不上传任何东西，可以在「工作室设置」里关掉。
• 上游服务商：网关可以指明地址背后真正处理请求的服务（OpenAI、英伟达 NIM、DeepSeek、OpenRouter、智谱、Moonshot、Together、xAI、Cerebras 等）。接本地中转或轮询时自动识别会失效，指明之后就按那家服务接受的参数发送；认不出的地址只发通用参数。
• 每分钟请求上限：每个网关可选，默认关。任意 60 秒内算一次，这个网关下所有任务、小队成员和重试都算在内，超出的请求在本机排队，状态栏显示本分钟用了多少。服务器回 429 时整个网关按它给的时间一起暂停。重试次数也可以改。
• 中转站不稳时：网关可以设「无响应断开」（填秒数，15–300，默认关）：一次请求超过这个秒数没有收到任何数据（包括回复到一半停住），就断开并重试；模型长时间思考不出字时可能被误判，按你的中转站来设。中转站常见的临时报错（繁忙、超时、上游错误之类）也会自动重试，密钥、额度和请求本身的错误不会。重试次数、每分钟请求上限和上游服务商保存后重启也不会丢。服务器回 429 时，这个网关按它给的时间暂停。排队、冷却、重试都会显示出来：工作台在状态栏，制卡工坊在输入框上方。
• 请求诊断：请求失败时，错误卡片上可以看到实际发出的参数、工具名、各角色的消息条数、HTTP 状态和关键响应头。不含对话正文、工具参数和密钥，可以直接贴给别人看。
• 删除对话：左栏、任务菜单、已归档列表都能删，按住 Ctrl 或 Shift 可以多选。删除会连带子任务、小队成员、会话文件和检查点；项目文件、项目记忆和每天的 Token 统计不受影响。
• 导出为 Markdown：把一段对话连同思考过程和工具调用导出成一个文件。
• 破限：输入框旁的开关，默认关。内置一套创作前提提示词（普通 / 严格两档），也可以从酒馆预设里挑条目导入；导入的只存在本机。开启后随这个任务的每次请求一起发送。
• 思考强度滑块：六档强度做成一条会亮起来的星空滑轨，两端是「更快」和「更聪明」，默认档标「推荐」。打开「减少动画」时不闪。

右栏的「浏览器」里有地址栏和多个标签页，登录数据与主程序分开保存。Agent 也能用它打开网页、读文字和元素、点击、输入、截图、查看控制台与网络请求，你可以随时【接手】，接手期间 Agent 在这个标签页上停手。

本机地址直接打开；其他网站第一次要你允许（允许一次 / 始终允许此网站 / 拒绝），被允许的站点里跳到别的站点同样要问。密码和支付输入一律不代填，提交表单要你确认，下载一律拒绝。

受限命令和终端使用原生 Windows AppContainer，没有网络能力，包括 localhost TCP。模型 API、联网搜索和 HTTP MCP 由独立的主机网络通道处理。隔离失败不会静默转成本机执行。

从工作台左下角进入。卡项目是本地文件夹，只出现在卡库里。规划 AI 按轮提问、给推荐答案，写出设计书后派单；世界书条目、正则、脚本、开场白各是一个组件文件，由应用确定性地拼装成卡。导出整卡 JSON、PNG 卡、世界书和单件，每次导出附一份替换说明。本地预览在隔离的 iframe 里按酒馆的处理顺序渲染，不联网。

• 变量一次做对：「变量结构」分区写一张变量表（路径、类型、默认值、范围、由谁维护），Zod 结构、初始变量 [initvar]、变量列表、变量输出格式和规则里的路径清单都由应用照表生成。拼装检查会核对状态栏和创角页读写的路径、示例补丁能不能打在初始变量上、渲染正则能不能匹配输出格式，还有条数上限和 [initvar] 的设置。导入的卡会推导出一份变量表，只用来核对，不改你的 Zod。
• 前端由应用编译：状态栏、正文美化、开局创角页可以写成「装配单」——选风格、分页、区块和要显示的变量，应用用自带的前端骨架编译成完整的页面。属性表、进度条、环形量表、标签、折叠面板、列表、关系卡、时间线、危机档位、变动徽章都是现成的区块，放不下的内容写成自定义区块。状态栏有三种形态：每楼一个占位符（只画最新一楼）、正文顶部的状态头（新卡默认）、悬浮球加可拖动的状态面板。创角页可以带一个「自定义开局」步骤。
• 八套皮肤：战术档案、鎏金典狱、工业终端、复古电影、粉樱、和纸、霓虹夜，外加按题材自定的一套，规划在设计书里选定。手写 HTML 的前端照旧支持，拼装时应用自动包上 ```html 围栏，酒馆助手才会把它渲染成能交互的页面。
• 导出前先看：预览里有一个「模拟酒馆」，替酒馆助手和 MVU 提供变量、事件和聊天世界书，所以状态栏和创角页也能预览。样例变量可以改了再推一次更新；创角页要写的东西只列成「将写入」，不会真的写；宽度可以切 375、768 和桌面三档。
• 质量检查：手机宽度下横向滚动、没有任何交互反馈、正文对比度不足、直接引用 Google 字体、塞了大图 base64、会卡死的正则会禁止导出；设计令牌太少、没有 @media、循环动画不照顾「减少动画」、强调色用得太多、字体镜像与外链素材、在安卓上表现不同的正则写法会提醒。
• 一处提改动：点卡项目主页的「提改动」，或在任意分区输入 /改动：一句话说要改什么，或者贴一段酒馆里的报错，AI 读整张卡列出影响清单；删掉不要的条目，点「照单开做」，应用按分区的先后顺序一口气改完，再跑拼装检查。只动一个文件时直接改好，可以撤销。没有设计书的卡也能用。
• 导入更顺：导入时先看条目名里的标记（[initvar]、人物总览、地点总览等）再看编号来分区；未分类的条目可以多选后一起移到分区，也可以点「AI 归类建议」，确认之后才搬。
• 小队：「工作室设置 → 制卡」里的「子代理」可以设成关、只读或可写（默认关）。打开后规划和分区可以派成员分头做：只读时派「查资料」，可写时再加「写组件」。写组件只能改分给它的组件，人物总览、出处索引、设计书、变量表只由主 AI 改；成员写的文件算进本轮写入，可以一起撤销。「自行组队」决定由 AI 自己判断什么时候派，还是只在你要求时派。任何分区都能选 Ultra：选了就照样派小队，规划派「查资料」，其他分区两种都派，不受这个开关限制。
• 分批写：分区 AI 一次只写一个组件文件，规划分批登记派单；一键制作遇到还没写完的派单会接着做完（一条派单最多自动接 20 轮，之后停下来等你点【继续】）。
• 说错了话可以撤回：AI 还在写时，撤回会停下这一轮、把这条消息收回，文字放回输入框，它发出的派单回到「未派」；已经写完的消息可以编辑后重新生成，产生新的对话版本。写进卡项目的文件不跟着回滚，要退文件用「本轮写入」里的【撤销本轮】。
• 工坊的标题和正文用内置的开源宋体（Noto Serif SC 的子集），每台电脑上看起来都一样。

• 「工作室设置 → 头像与外观」里可以选主题：跟随系统、深色、浅色、红粉白。主题只换颜色，布局和控件不变；制卡工坊保持自己的样子，只有金色和板块灯光会向主题色偏一点。
• 自己的主题包放在资料目录的 themes/<id>/theme.json 里，可以改颜色、背景图、登场动画和桌宠；读不了的主题包会列出原因。
• 桌宠默认关。打开后它在自己的小窗口里，浮在所有窗口最前面，Cardwright 最小化或收到托盘时也还在。它只报本机的状态：当前任务、审批、结果、一键制作的进度和今天的 Token 用量，不调用模型。单击它回到 Cardwright 并打开它正在报的任务；拖到屏幕上任何位置，下次还在那里；右键有菜单，点 × 收起；系统开了「减少动画」时它不动。宠物包用 Codex 桌宠包格式（pet.json 加 spritesheet.webp），可以从 ZIP 或文件夹安装。
• 内置的绘梨衣桌宠由顾清寒创作，是非官方同人作品，按 CC BY-NC 4.0（署名、非商业）授权；角色权利归原权利人，权利人要求即移除。详见 assets/pets/erii/NOTICE.md。

添加网关时可以选两个预设，都是你自己在本机运行的服务，Cardwright 不附带、不安装、也不管理它们：

• 本地模型：Ollama（http://127.0.0.1:11434/v1）、llama.cpp（http://127.0.0.1:8080/v1）、LM Studio（http://127.0.0.1:1234/v1）。本机地址可以不填密钥，应用会发送占位值 local。
• 本地代理网关：你自己运行的 OpenAI 兼容代理，例如 CLIProxyAPI（http://127.0.0.1:8317/v1），密钥填你在它的配置里给客户端设的那一个。账号能不能这样用，请自行确认服务条款。

【一键自检】先读模型列表，再给默认模型发一次只生成一个 Token 的补全，两步分别报告结果。

资料默认存在 %APPDATA%\Cardwright：设置、会话、检查点、附件和凭据。凭据用 Windows 的 safeStorage 加密。升级时如果数据结构变化，会先在 backups/ 留一份原始备份。

界面某一页出错时，那一页换成错误卡片，可以一键复制诊断信息；同一条记录也写进 logs/renderer.log，不含卡和对话的内容。「工作室设置 → 数据与导出」里可以打开日志文件夹，反馈问题时附上它。

MIT，见 LICENSE。assets/ 下的素材不在 MIT 之内，各自的来源与条件见对应文件夹里的 NOTICE.md；内置桌宠只能非商业使用。第三方组件的许可声明见 THIRD_PARTY_NOTICES.md。

• 顾清寒
• Claude（Anthropic，AI 结对）

Cardwright is a local agent workbench for Windows projects: connect your own model gateway, let an agent do the work, and review the changes, run the checks and manage a squad in the same window. Since 0.8 it also has a separate card studio: a planning AI walks you through turning your material into a SillyTavern character card.

The interface stays quiet, deterministic rules are executed by the program, and skills and extra context are handed to the model only when needed. Every model request goes to the gateway you configured; keys stay on your machine, encrypted by Windows credential storage.

• Windows 10 or 11 (x64).
• Your own model gateway: OpenAI Chat Completions, OpenAI Responses or Anthropic Messages.
• Optional: Brave Search or SearXNG credentials for web search, kept apart from the model credentials.

Download Cardwright-Setup-1.3.2.exe from Releases. The installer is not code-signed, so Windows shows an "unknown publisher" warning the first time; choose "More info → Run anyway". Uninstalling keeps your data directory.

Or build it yourself:

npm install
npm run build
node scripts/package.mjs      # portable output in release/<version>/Cardwright-win32-x64

1. Add a gateway under Studio settings → Agents & models: protocol, base URL and key. Pick models from the list it reads, or add IDs by hand.
2. Add a project folder, choose gateway, model, context window and permission in the composer, then send a task.
3. In the default mode you approve each action after reading the actual paths and commands. Restricted commands have no network; host commands are approved separately.
4. Open the side panel beside the task title for files, changes, checks, terminal, browser, plan and todos. The panel splits into two.
5. Configure project check commands under Workbench & verification. When a run ends, the delivery card keeps file changes and verification apart.

• Left: conversations grouped by project, with pin, rename, archive and search, and a status dot each (running, waiting for approval, finished, failed, output truncated).
• Middle: the conversation at reading width. The agent marks a chapter when the work changes phase; a rail of dots in the margin jumps between them. A path:line in a reply or a tool row opens that file in the side panel at that line.
• Right: a splittable panel — files, changes, checks, terminal, browser, plan, todos. Under 960px the sidebar becomes an icon column and the panel slides over the conversation.
• Slash commands: /clear, /compact, /context, /cost, /model, /resume, /rewind, /init, /help, /plan, /todos, /memory, /dream, and your skills.
• Shift+Tab cycles ask → auto-edit in project → plan → full access.
• Subagents: your own, and the ones in .claude/agents/*.md in your home folder, are on from the start; ones found in a project folder start off, so turn one on in settings once you trust its file. The lead can send any subagent that is on. A file's model: line has no effect; pick the model for it in settings.
• 小绘: the AI built into Cardwright is called 小绘. Ask which model it runs on and it names the model and the gateway you configured, and says it cannot verify which model actually sits behind the gateway. Its way of talking can be switched off under Profile & appearance.
• Hooks: SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, SubagentStop and Notification, configured in Claude Code's settings.json shape, importable from Claude Code and testable one by one.
• Conversation (in the workbench and the card studio alike): code blocks are syntax-highlighted with a language label and a copy button, and long ones fold; the model's thinking shows as it happens and folds to "Thought for N s" when done; consecutive calls of the same tool fold into one row; a reply ends in a cursor while it writes, new text fades in and scrolling follows smoothly; tables are striped and scroll sideways when wide. Under reduced motion none of it moves.
• New version reminder: once a day Cardwright reads the latest version number on GitHub; when there is a newer one, the top bar says so and opens the release page on a click. Nothing is downloaded or uploaded, and Studio settings can turn it off.
• Upstream service: a gateway can name the service that actually answers behind its address (OpenAI, NVIDIA NIM, DeepSeek, OpenRouter, Z.ai, Moonshot, Together, xAI, Cerebras and others). A local relay or proxy defeats detection; naming the service sends the fields that service accepts, and an unrecognised address receives only widely accepted ones.
• Requests per minute: optional per gateway, off by default. Counted over any rolling 60 seconds across every task, squad member and retry on that gateway; a request over the limit waits on this computer and the status line shows how full the minute is. A 429 pauses the whole gateway for as long as the service asks. The retry count is configurable too.
• When a relay is unstable: a gateway can set Disconnect when silent (seconds, 15 to 300, off by default): a request that receives no data for longer than that, even halfway through a reply, is cut off and retried. A model that thinks for a long time without streaming any output can be mistaken for a stall, so set it to suit your relay. A relay's usual temporary errors (busy, timed out, upstream error and the like) are retried automatically; a wrong key, an empty balance or a bad request is not. The retry count, the requests-per-minute limit and the upstream service are kept across restarts. When the server answers 429, that gateway pauses for as long as it asks. Waiting in the queue, cooling down and retrying all show on screen: in the status line in the workbench, above the composer in the card studio.
• Request diagnostic: when a request fails, the error card shows the parameters that went out, the tool names, how many messages of each role, the HTTP status and the response headers worth reading. It holds no message text, no tool arguments and no credential, so it can be pasted into a bug report as it stands.
• Delete conversations: from the session list, the task menu or the archived list, with Ctrl or Shift for several at once. Deleting takes sub-tasks, squad members, session files and checkpoints; project files, project memory and daily token totals are untouched.
• Export as Markdown: a conversation, its thinking and its tool calls as one file.
• Framing prompts (破限): a toggle beside the composer, off by default. Two built-in sets, or entries imported from a SillyTavern preset; imported sets stay in your data folder. When on, they lead every request this task makes.
• Reasoning effort slider: the six levels become a night sky that lights up as the level rises, labelled faster and smarter, with your default marked. It stops moving under reduced motion.

The Browser tab in the side panel has an address bar and tabs, and keeps its logins apart from the rest of the application. The agent can open pages, read the text and the elements, click, type, take screenshots and read the console and network list. You can take a page over at any time; while you have it, the agent's tools stop on that tab.

Local addresses open straight away. Any other site asks you first (allow once / always allow this site / refuse), and a link that leaves an allowed site asks again. Password and payment fields are never typed into, submitting a form asks you, and downloads are refused.

Restricted commands and the terminal run in a native Windows AppContainer with no network at all, including localhost TCP. Model APIs, web search and HTTP MCP go through a separate host channel. A failed isolation never silently falls back to running on the host.

Enter from the bottom of the sidebar. A card project is a local folder and appears only in the card library. The planning AI asks one question at a time with a recommended answer, writes a design document and then dispatches the work. World-book entries, regexes, scripts and greetings are each a component file, and the application assembles the card deterministically. Export the whole card as JSON or PNG, the world book, or single pieces; every export comes with a note on how to replace it in SillyTavern. The local preview renders in an isolated iframe in SillyTavern's own order, offline.

• Variables right the first time: the variable structure section writes a variable table (path, type, default, range, who maintains it), and the application generates the Zod schema, the [initvar] initial variables, the variable list, the output format and the paths in the rules from it. The assembly checks verify the paths the status bar and the start page read and write, apply the sample patch to the initial variables, match the renderer regex against the output format, and check list caps and the [initvar] settings. An imported card gets a derived table that is only checked against; your Zod stays as it is.
• Front-ends compiled by the application: a status bar, body renderer or start page can be an assembly sheet — a style, pages, blocks and the variables they show — that the application compiles with its built-in front-end skeleton into a complete page. Stat tables, bars, ring gauges, tags, folds, lists, relationship cards, timelines, crisis tiers and change badges are ready-made blocks; anything else goes in a custom block. The status bar comes in three forms: a placeholder on each message (only the latest one is drawn), a status head at the top of the body (the default for new cards), or a floating orb with a draggable status panel. A start page can include a free-text custom opening step.
• Eight skins: Tactical dossier, Gilded ward, Industrial terminal, Vintage cinema, Sakura, Washi and Neon night, plus one derived from the card's subject, chosen in the design document. Hand-written HTML front-ends still work; on assembly the application wraps them in the ```html fence the SillyTavern helper needs to render them as interactive pages.
• See it before you export: the preview carries a simulated tavern that stands in for the SillyTavern helper and MVU (variables, events, the chat world book), so status bars and start pages can be previewed as well. Edit the sample variables and push an update; what a start page would write is listed, never written; switch between 375, 768 and desktop widths.
• Quality checks: horizontal scrolling at phone width, no interaction feedback at all, body text below 4.5:1, Google's own font links, big base64 images and regexes that can hang block the export; too few design tokens, no @media, looping motion that ignores reduced motion, an overused accent, font mirrors and external media, and regex writings that behave differently on Android are warnings.
• Change it in one place: Ask for a change on the card project home, or type /改动 in any section: say in a sentence what to change, or paste an error from SillyTavern, and the AI reads the whole card and lists what the change affects. Take out what you do not want, press Go ahead, and the application works through the sections in dependency order and runs the assembly checks. A change to a single file is made directly and can be undone. Cards without a design document work too.
• Smoother imports: entries are sorted by the markers in their names first ([initvar], character and place overviews and the like) and by their order number second; unclassified entries can be moved to a section several at a time, or you can ask for AI suggestions and confirm them before anything moves.
• Squads: the "Sub-agents" setting under Studio settings → Card studio can be Off, Read only or Can write (off by default). When it is on, planning and the section AIs can send members to work in parallel: 查资料 (research) when it is Read only, 写组件 (write a component) as well when it can write. A 写组件 member can change only the components it was given; the character overview, the source index, the design document and the variable table stay with the main AI. What the members write counts in the turn's writes and is undone together with them. "Self-organised squads" decides whether the AI judges when to send members, or sends them only when you ask. Ultra is offered in every section and sends a squad whatever the setting: planning sends 查资料, the other sections both kinds.
• Batched writing: a section AI writes one component file per request, and planning registers its dispatches a board at a time. One-click making carries on with a dispatch that is not finished yet (at most 20 rounds for one dispatch, then it stops and waits for you to press Continue).
• A message can be withdrawn: while the AI is still writing, withdrawing stops the turn, takes the message back, returns its text to the composer and puts a dispatch it sent back to not sent. A finished message can be edited and regenerated into a new conversation version. Files already written stay as they are; use Undo this turn in the turn's writes to take them back.
• The studio's headings and text use a bundled open-source serif (a subset of Noto Serif SC), so it looks the same on every computer.

• Studio settings → Profile & appearance offers Follow system, Dark, Light and Red, pink & white. A theme changes colours only; the card studio keeps its own look, with only its gold and board light leaning towards the theme.
• Your own theme packs go in themes/<id>/theme.json in the data directory and can change colours, a background picture, the entrance and the pet; a pack that cannot be read is listed with the reason.
• The desk pet is off by default. When on, it floats in its own small window above every other window, and stays there while Cardwright is minimized or in the tray. It only reports local state: the task at hand, approvals, results, one-click making's progress and today's tokens; it never calls a model. Click it to bring Cardwright back on the task it reports; drag it anywhere on the screen and it stays there next time; right-click for its menu, or close it with its ×. Under reduced motion it stands still. Pet packs use the Codex pet format (pet.json and spritesheet.webp) and install from a ZIP or a folder.
• The built-in Erii pet is unofficial fan art by 顾清寒, licensed CC BY-NC 4.0 (credit, non-commercial). The character belongs to its rights holders, and the pet is removed at a rights holder's request. See assets/pets/erii/NOTICE.md.

When adding a gateway you can start from two presets. Both are services you run yourself on this computer; Cardwright does not ship, install or manage them:

• Local model: Ollama (http://127.0.0.1:11434/v1), llama.cpp (http://127.0.0.1:8080/v1) or LM Studio (http://127.0.0.1:1234/v1). On a local address the key may be left empty; the placeholder local is sent.
• Local proxy gateway: an OpenAI-compatible proxy you run yourself, such as CLIProxyAPI (http://127.0.0.1:8317/v1); enter the client key you set in its config. Whether your accounts may be used this way is for you to check.

Self-test lists the models, then sends one completion of a single token to the default model, and reports each step.

Everything lives in %APPDATA%\Cardwright: settings, sessions, checkpoints, attachments and credentials. Credentials are encrypted with Windows safeStorage. If the data shape changes across versions, the original file is copied into backups/ first.

When a page of the interface fails, an error card takes its place and copies the diagnostics in one click; the same record goes into logs/renderer.log, without card or conversation content. Studio settings → Data & export opens the log folder; attach it when you report a problem.

MIT, see LICENSE. The files under assets/ are not covered by the MIT License; each folder's NOTICE.md gives its source and terms, and the built-in desk pet is for non-commercial use only. Third-party notices are in THIRD_PARTY_NOTICES.md.

• 顾清寒
• Claude (Anthropic, AI pair)