github Darkatse/TauriTavern v2.1.0
v2.1.0 - 试试SubAgent吧！
on GitHub

latest releases: Canary, v2.1.1, mobile-auto-v2.1.1-dev...
one month ago

SubAgent文档已经更新

Testflight Canary测试邀请链接 (内部测试，更新频繁，但是可能有Bug)

本次更改内容：


Agent系统

  • Agent目前可以独立指定预设与模型，与当前实际选择状态彻底解耦
    • 请注意，模型需要先在API连接管理中先保存过仅有模型的配置
  • 新增SubAgent支持，主Agent可以派出子Agent完成次级任务，随后根据子Agent的汇报结果继续推进工作
    • 子Agent可以设置独立的预设与模型
  • 当前生效的Agent配置会在面板中着重显示，并可用下拉框进行切换。编辑其余配置不会影响实际生效的配置。
  • workspace write file工具允许在已有文件后追加新的内容，而不用全部重写（append mode）

  • 新增掷骰子工具dice roll，采用dnd基本掷骰规则，跑团玩家狂喜。此工具默认禁用
  • Agent前端运行面板现在可以看到从头到尾所有的历史轮次调用了，不会被剪裁
  • 修复了Agent 读取skill时强制限制最高80k tokens内容的问题，现在遵循Agent配置中的数值了
  • 修复了Agent Result未正确处理，造成Agent性能下降的问题
  • 不再强制要求Agent System Prompt为System Role，与DS缓存优化脚本等兼容
    • 注意，Agent System Prompt不为System可能会造成效力下降，请自行权衡利弊
  • Agent模式下直接输出内容不再立即失败，而是存放到output/direct_output.md文档内，以便之后重新利用。这次失败请求的返回消息中会引导Agent使用工具而非直接输出。
  • 修复了Agent模式下Swipe可以看到之前的被Swipe掉的最后一层楼的问题。
  • 对请求成功但解析失败的Agent过程，允许进行重试。感谢@yhny1001
  • 当Agent误将文件夹当作文件读取时不再立刻失败，允许进行一次重试。感谢@yhny1001
  • 若Agent在提交过一次内容后发生指令漂移（例如输出的内容不合法），允许进行一次重试，而不是直接失败。感谢@yhny1001
  • 当Agent运行失败后自动回滚消息，感谢@yhny1001
  • 优化了内置的默认Agent系统提示词，感谢@makerlinck
  • 增加了一些Agent犯蠢时候的兜底

Skill管理

  • 将skill管理变为单独的扩展面板
  • 引入 全局/预设/Agent配置/角色卡 四级层级划分，多预设skill不再互相干扰
  • 可以在前端直接进行skill内容修改


  • 新增了SKILL复制导入以及从网络下载导入的选项
  • 将SKILL默认导出格式更改为.zip

更新上游1.18.0


Bug修复

  • 修复了正则的Rust后端加速会执行已经禁用的正则的问题

  • 修复了iOS上无法删除背景的问题
  • 修复Android端偶发的移动文件不稳定问题
  • 修复更新1.18.0前端后部分界面输入法抬高适配失效的问题
  • 恢复支持部分旧群聊文件，读取时不再报错

杂项

  • 允许添加自定义模型
名称
  • 允许只保存模型连接配置，将预设与模型解耦
  • 可以直接导入SillyTavern网页备份的数据压缩包
  • 新增Claude Opus 4.8模型
  • 新增xhigh与max推理强度，自动适用于OpenAI或Claude的支持模型
  • 接入AWS Bedrock渠道
    
- 重命名OpenRouter应用归属请求标头
  • 允许对所有渠道添加附加参数，感谢@exynos967
  • 可以使用 /trigger 激活agent模式，感谢@calm-prussian
  • 可以使用 /llmlog 命令快速打开LLM API Log面板，感谢@Moeblack
  • 数据迁移扩展允许导入.tar.gz压缩文件



下个版本会有什么

  • 多Agent接力
  • 同步功能小修，加入Agent文件同步
  • 历史Agent运行回顾与管理

SubAgent Documentation Has Been Updated

Testflight Canary Test invitation link (internal testing, frequently updated, but may contain bugs)

Changes in this release:

Agent System

  • Agents can now specify their own presets and models independently. They are no longer tied to the currently selected preset or model in the main chat.
    • Note: the model must first be saved as a model-only configuration in API Connection Management.
  • SubAgent support has been added. A main Agent can delegate secondary tasks to child Agents, then continue working based on the SubAgent reports.
    • SubAgents can use their own presets and models.
  • The currently active Agent Profile is now highlighted in the panel and can be switched with a dropdown. Editing other Profiles does not change the Profile currently used for runs.
  • The workspace.write_file tool can now append content to an existing file instead of rewriting the whole file.
  • Added the dice.roll tool. It uses basic D&D-style dice notation and may be useful for tabletop and roleplay scenarios. This tool is disabled by default.
  • The Agent frontend run panel can now show the full run history from start to finish without tail-only truncation.
  • Fixed the hard 80k-token cap when Agents read Skills. Skill reading now follows the values configured in the Agent Profile.
  • Fixed incorrect Agent Result handling that could reduce Agent performance.
  • Agent System Prompt is no longer forced to use the System role. This improves compatibility with cache optimization scripts and similar setups.
    • Note: using a non-System role for Agent System Prompt may reduce its effect. Please choose based on your own setup.
  • Direct text output in Agent Mode no longer fails immediately. It is saved to output/direct_output.md so it can be reused later. The failure response now guides the Agent to use tools instead of replying directly.
  • Fixed an issue where Agent Mode swipe generation could see the last discarded swipe layer.
  • Agent runs can now retry when the request succeeds but response parsing fails. Thanks to @yhny1001.
  • When an Agent mistakenly reads a folder as a file, the run no longer fails immediately and can retry once. Thanks to @yhny1001.
  • If an Agent drifts after committing once, for example by producing invalid output, it can retry instead of failing immediately. Thanks to @yhny1001.
  • Agent failure now rolls back the message automatically. Thanks to @yhny1001.
  • Improved the built-in default Agent system prompt. Thanks to @makerlinck.
  • Added more fallback handling for common Agent mistakes.

Skill Management

  • Skill management has been moved into its own extension panel.
  • Added four Skill scopes: global, preset, Agent Profile, and character card. Skills from different presets no longer interfere with each other.
  • Skill content can now be edited directly from the frontend.
  • Added options to import Skills by copying them in, and to import Skills from a network download.
  • The default Skill export format is now .zip.

Upstream 1.18.0 Update

Bug Fixes

  • Fixed an issue where Rust backend regex acceleration could execute disabled regex rules.
  • Fixed background deletion on iOS.
  • Fixed occasional unstable file moves on Android.
  • Fixed input method lift adaptation issues in some UI areas after updating the frontend to 1.18.0.
  • Restored support for some older group chat files. They no longer fail during loading.

Miscellaneous

  • Custom model names can now be added.
  • Model connection configurations can now be saved without presets, allowing presets and models to stay decoupled.
  • SillyTavern web backup archives can now be imported directly.
  • Added the Claude Opus 4.8 model.
  • Added xhigh and max reasoning effort levels. They can be applied to supported OpenAI or Claude models.
  • Added AWS Bedrock channel support.
  • Renamed the OpenRouter application attribution request header.
  • Additional parameters can now be added to all channels. Thanks to @exynos967.
  • /trigger can now activate Agent Mode. Thanks to @calm-prussian.
  • Added the /llmlog command to quickly open the LLM API Log panel. Thanks to @Moeblack.
  • The data migration extension can now import .tar.gz archives.

What Is Coming Next

  • Multi-Agent handoff.
  • Sync improvements, including Agent file sync.
  • Agent run history review and management.

What's Changed

  • example: add recommend agent instruction of deepseek-v4-series_ZH and deepseek-v4-series_EN by @makerlinck in #43
  • 【功能】【低优先级】新增 /llmlog 斜杠命令，快捷打开 LLM API 日志窗口 by @Moeblack in #50
  • feat(openai): 允许非自定义提供商添加附加参数 by @exynos967 in #64
  • feat(diagnostics): log upstream body preview when chat completion JSON parse fails + reclassify as transient by @yhny1001 in #61
  • fix(agent): translate directory access to recoverable error (#54) + emit run_rollback_targets on instruction drift (#55) [backend] by @yhny1001 in #62
  • fix(agent): wire #55 frontend UI to userRetryable + run_rollback_targets by @yhny1001 in #63
  • feat(agent): soft drift recovery for model.tool_call_required + strengthen workspace.commit hint by @yhny1001 in #68
  • feat: 允许 /trigger 斜杠命令触发 Agent 模式 by @calm-prussian in #58
  • Sync/upstream 1.18.0 by @Darkatse in #71
  • fix(chat): custom_* 附加参数字段容忍 JSON null by @yhny1001 in #77
  • feat(chat): 新增 AWS Bedrock 通道（7 大 provider + 自定义入口） by @yhny1001 in #76
  • version 2.1.0 by @Darkatse in #78

New Contributors

Full Changelog: v2.0.0...v2.1.0

Check out latest releases or
releases around Darkatse/TauriTavern v2.1.0

Don't miss a new TauriTavern release

NewReleases is sending notifications on new releases.

Get notifications