Skip to content

ChatSpeed v2.0.15

Choose a tag to compare

View all tags
@github-actions github-actions released this 04 Sep 14:48
v2.0.15
cba1831

[2.0.15]

🚀 New Features

  • Copy in the Workflow UI:
    • Added a copy button to AI message outputs and streaming content that copies the rendered raw Markdown.
    • Approved submit_plan tool titles now show a "Copy Raw Plan" button on hover in the message list and the plan approval dialog; it copies the plan Markdown exactly as rendered when the tool entry is expanded.
  • AMD Radeon support:
    • Added an AMD thinking adapter for developer.amd.com.cn endpoints: unsupported thinking, enable_thinking, and thinking_budget fields are stripped before forwarding, and reasoning_effort is normalized to the tiers the provider accepts (DeepSeek-V4-Flash: low/high; Qwen3.8-Flash-Next: low/medium).
    • Registered the adapter in the model catalog and transport catalog by endpoint host, with documented sources.

🪄 Improvements

  • Context compression:
    • Compression history now uses a structured <messages> XML format with separate <reasoning>, <content>, and <tool_use> nodes, preserving assistant tool-call intent and tool result/error identity alongside existing projection and filtering semantics.
    • The compressor model now prefers the utility model only when its confirmed context capacity covers the current phase action model's required context bound; unknown or insufficient capacity safely falls back to the action model.
  • Usage summary pricing:
    • Terminal usage summaries now report the actual pricing rates, cache rate, and multiplier captured at request time, backfilled from the latest ccproxy_stats pricing snapshot per provider and model instead of empty pricing fields.

🐞 Bug Fixes

  • Localization: Added the missing settings.agent.selectedSkills key used by the skill selection form label in all supported locales.
  • Config transfer: Agent export and import now validate mcp_tool_exposure as an McpToolConfig object, while legacy array values remain accepted for backward compatibility.
  • Message list scrolling: Scrolling up now pauses auto-scroll immediately so streaming output no longer fights the user's reading position; auto-scroll resumes only after scrolling back to the bottom.

🚀 新功能

  • Workflow 界面拷贝：
    • AI 消息输出和流式内容新增拷贝按钮，可一键复制渲染后的原始 Markdown。
    • 消息列表和计划审批对话框中，已批准的 submit_plan 计划标题悬停时显示"复制原始计划"按钮；复制的计划 Markdown 与展开渲染的内容完全一致。
  • AMD Radeon 支持：
    • 为 developer.amd.com.cn 端点新增 AMD 思考适配器：转发前剥离不支持的 thinking、enable_thinking 和 thinking_budget 字段，并将 reasoning_effort 归一化为厂商接受的档位（DeepSeek-V4-Flash 支持 low/high，Qwen3.8-Flash-Next 支持 low/medium）。
    • 适配器已按端点主机注册到模型目录和传输目录，并附有文档来源。

🪄 改进

  • 上下文压缩：
    • 压缩历史改用结构化的 <messages> XML 格式，为每条消息区分 <reasoning>、<content> 和 <tool_use> 节点，在保留既有投影和过滤语义的同时，完整保留助手工具调用意图以及工具结果/错误身份。
    • 压缩模型选择增加上下文容量门控：仅当工具模型的确认上下文容量覆盖当前阶段执行模型所需上下文边界时才优先使用；容量未知或不足时安全回退到执行模型。
  • 用量摘要定价：
    • 终态用量摘要现在会报告请求发生时的实际单价、缓存单价与倍率（按提供商和模型回填 ccproxy_stats 中最近一次定价快照），不再显示空的定价字段。

🐞 修复

  • 本地化：为技能选择表单标签补充所有支持语言中缺失的 settings.agent.selectedSkills 翻译键。
  • 配置传输：Agent 导出/导入时 mcp_tool_exposure 现按 McpToolConfig 对象校验，同时继续兼容遗留的数组格式。
  • 消息列表滚动：向上滚动会立即暂停自动跟随，流式输出不再与用户的阅读位置"抢滚"；只有滚动回到底部后才恢复自动跟随。
Assets 22
Loading