Skip to content

Cardwright 1.3.0

Choose a tag to compare

View all tags
@1798547983tt 1798547983tt released this 03 Oct 05:21
· 2 commits to main since this release
v1.3.0
f91e102

中文

Cardwright 1.3.0。

  • 小绘：内置 AI 现在叫小绘。问它用的是什么模型，它会说出你配置的模型和网关名，并说明网关后面实际接的是哪个模型它自己核实不了。它说话有自己的性格：话不多但说清楚，真诚，有一点安静的好奇心；写进文件的内容和应用要解析的固定格式不受影响。不喜欢可以在「头像与外观」里关掉「小绘的性格」，名字和如实说出模型不变。
  • 制卡工坊的小队：「工作室设置 › 制卡 › 小队」里的「子代理」可以设成关、只读或可写，默认关。打开后，规划和分区可以派小队成员分头做：「查资料」只读，交回带出处的要点；「写组件」只改分给它的组件。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」，【撤销本轮】会一起撤销；对话里能看到每位成员在做什么，也可以单独停下。「自行组队」决定由 AI 自己判断什么时候派，还是只在你要求时才派。同时跑几个，看「同时运行的 Agent」；每个主 AI 最多带几个成员，看「成员额度」。
  • 分批写：分区 AI 一次请求只写一个组件文件；规划先写设计书，再一批一批登记派单。中转站中途掐断时，损失的只是一小段。一键制作遇到还没写完的派单（比如逐个写人物），会在同一个对话里接着做，最多自动续 20 轮；上次停在一半的派单，下次一键制作也会接上。
  • 中转站更稳：
    • 网关新增「无响应断开」，默认关：回复中途超过设定的秒数（15–300）一点数据都没有，就断开重试。
    • 「上游服务商」「每分钟请求上限」和「重试次数」保存后，重启不再丢失。
    • 服务器回 429 时，整个网关的冷却真正生效。
    • 中转站常见的临时报错（上游繁忙、超时、Bad Gateway 等）也会自动重试；密钥、余额、内容审核这类错误不会白白重试。
    • 排队、冷却、重试会显示在工作台的状态栏和制卡工坊的输入框上方。
  • 工作台子代理：自己建的子代理和读到的 .claude/agents、.agents/agents 子代理都能派出去，主代理也看得到；自建子代理多了「说明」栏。探索员、规划师始终只读。「成员额度」（原「默认小队人数」）真正生效。派成员的工具和内置角色改用 Cardwright 自己的叫法：执行员、探索员、规划师。
  • 显示修复：变量表编辑框的每一列都能完整显示，表头不再压住第一行；选模型的菜单无论在哪里打开都不会被裁掉。

升级须知

  • 项目文件夹里读到的子代理（.claude/agents、.agents/agents）改为默认关闭；要用的话，到「设置 › 工作区能力 › 子代理」里重新打开。
  • 1.2.0 每次重启都会丢掉网关的「上游服务商」「每分钟请求上限」和「重试次数」。升级后请再填一次，之后会一直保留。
  • 在开发者模式下改过内置提示词的：这次更新过的提示词会标为「默认已更新」。想用上新的小绘身份、分批登记派单和接着做的标记，请先恢复默认再改。
  • 自建角色如果用了 executor、explorer、planner、researcher、writer 这几个编号，会改名为「原编号-custom」，它的任务和设置跟着走。
  • 「成员额度」现在真的会限制：额度设得低时，派不出更多成员。
  • 下载新安装包直接覆盖安装，资料目录保留。

安装

下载下方的 Cardwright-Setup-1.3.0.exe。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（%APPDATA%\Cardwright）。

SHA-256：0afd55d20eb04d0b3a351e65b896e2624a835a1d82ef4fa875a8009e5dd5795e

English

Cardwright 1.3.0.

  • 小绘: the built-in AI is now called 小绘. Ask which model it runs on and it names the model and gateway you configured, and says it cannot verify which model actually sits behind the gateway. It talks with a personality of its own — says little but says it clearly, honest, quietly curious — while the files it writes and the fixed formats the app reads are unaffected. Switch off "小绘's personality" under Profile & appearance if you prefer; the name and the honest model answer stay.
  • Squads in the card studio: Studio settings › Card studio › Squad has a Subagents switch: off, read-only or writable, off by default. When it is on, planning and section AIs can send squad members to work in parallel: 查资料 (research) only reads and reports key points with their sources; 写组件 (write components) changes only the components assigned to it. The character overview, source index, design book and variable table are written by the lead only. Members' writes appear under "Written this turn" and "Undo this turn" reverts them too; you can see what each member is doing in the conversation and stop any one of them. "Self-organised squads" decides whether the AI chooses when to send members or waits for you to ask. How many run at once follows "Concurrent agents"; how many members one lead may hold follows "Member quota".
  • Batched writing: a section AI writes one component file per request; planning writes the design book first and then registers dispatches in batches, so a relay cutting a reply short loses only a little. One-click making continues an unfinished dispatch (such as writing characters one by one) in the same conversation, up to 20 automatic rounds, and the next run picks up a dispatch that was stopped halfway.
  • Steadier behind relays:
    • New per-gateway "Disconnect when silent", off by default: a reply that receives nothing for the set number of seconds (15–300) is cut and retried.
    • The upstream service, requests per minute and retry count no longer reset when the app restarts.
    • The gateway-wide cooldown after a 429 now actually takes effect.
    • Common temporary relay errors (upstream busy, timeouts, Bad Gateway and the like) are retried automatically; key, balance and content-filter errors are not retried in vain.
    • Queueing, cooldown and retries show in the workbench status line and above the card studio composer.
  • Workbench subagents: your own subagents and those read from .claude/agents or .agents/agents can all be dispatched, and the lead sees them; custom subagents get a Description field. 探索员 (explorer) and 规划师 (planner) always only read. The Member quota (formerly the default squad size) is now enforced. The dispatch tools and built-in roles use Cardwright's own names: executor, explorer, planner.
  • Display fixes: every column of the variable table editor is readable and the header no longer covers the first row; the model menu is never clipped, wherever the picker sits.

Upgrading

  • Subagents found in a project folder (.claude/agents, .agents/agents) are now off by default; turn the ones you use back on under Settings › Workspace intelligence › Subagents.
  • 1.2.0 lost a gateway's upstream service, requests per minute and retry count on every restart. Enter them once more after upgrading; they stay from now on.
  • If you edited built-in prompts in developer mode, the prompts this release changes are marked "default updated". Restore the defaults to get the 小绘 identity, batched dispatch registration and the continue marker, then edit again.
  • A custom role saved under executor, explorer, planner, researcher or writer is renamed to "-custom"; its tasks and settings follow it.
  • The Member quota is now enforced: a low value refuses larger squads.
  • Download the new installer and install it over the old version; your data directory is kept.

Install

Download Cardwright-Setup-1.3.0.exe below. 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 (%APPDATA%\Cardwright).

SHA-256: 0afd55d20eb04d0b3a351e65b896e2624a835a1d82ef4fa875a8009e5dd5795e

Assets 3
Loading