ChatSpeed v2.0.16

[2.0.16]

🚀 New Features

• Zhugeliang advisor sub-agent:
  • Added a read-only Zhugeliang child agent that diagnoses difficult coding problems and compares solution strategies, wired into the Coding agent's planning and execution guidance.
• MCP tool executor:
  • Added an MCP tool executor dialog: a new run_mcp_tool command invokes MCP tools from the settings page, with parameters rendered dynamically from the tool input schema; the execute button is shown on running servers only.
• Workflow:
  • Clicking the primary path in the titlebar now expands the sidebar (if collapsed) and switches to the Files tab, providing direct access to authorized paths.
  • Added a dedicated lite model role for title generation and one-shot user-input language detection, selectable in agent/workflow model configuration.
  • Runtime-generated observations delivered to the model are now wrapped in a <RUNTIME_OBSERVATION> envelope so the model can distinguish them from genuine user input; the system prompt now also clarifies that English runtime content is not a language signal.

🪄 Improvements

• Performance and maintenance:
  • The Models.dev catalog now loads asynchronously so the multi-MB catalog snapshot no longer blocks application startup; removed an unused built-in agent version-sync key.
  • Renamed the tray model label to "AI Provider"; added a dev Docker Compose setup, added the xlrd dependency to the python-slim image, and removed the chatspeed-docs submodule.

🐞 Bug Fixes

• Model handling:
  • Vendor-prefixed models (e.g. ModelScope's Qwen/Qwen3.8-Flash-Next) now resolve through the short-id fallback, and a new ModelScope qwen transport rule resolves the qwen thinking adapter, preventing the reasoning_effort + thinking_budget conflict.
  • Model overrides are now applied when the selection changes: proxy aliases resolve to the first real backend model and populate contextSize, maxTokens, thinking, and temperature, including on initialization.
• Workflow:
  • Centralized message list scroll handling: window anchors are preserved and stale scroll targets cleared, and user scrolling immediately cancels pending auto-follow scrolls, keeping the reading position stable during message trimming, streaming, resizing, and scrollbar dragging.
  • Language detection now parses the chat JSON envelope for visible content with a strengthened system prompt; usage after clear-context is attributed to the current segment; the MCP executor dialog sorts required parameters first; overlay height and border-radius are fixed; i18n keys are re-sorted.
• Dependencies:
  • Bumped @xmldom/xmldom to 0.8.15.

---

🚀 新功能

• Zhugeliang 顾问子 Agent：
  • 新增只读的 Zhugeliang 子 Agent，用于诊断疑难编码问题并对比解决方案策略，已接入 Coding 智能体的规划与执行指引。
• MCP 工具执行器：
  • 新增 MCP 工具执行对话框：通过新的 run_mcp_tool 命令在设置页调用 MCP 工具，参数按工具输入 Schema 动态渲染；仅对运行中的服务器显示执行按钮。
• Workflow：
  • 点击标题栏主路径现在会展开侧边栏（若已折叠）并切换到 Files 标签页，方便直接访问授权路径。
  • 新增专用的 lite 模型角色，用于标题生成和一次性的用户输入语言检测，可在 Agent/Workflow 模型配置中选择。
  • 送入模型的运行时观察事件现在包裹在 <RUNTIME_OBSERVATION> 信封中，模型可据此区分真实用户输入与运行时内容；系统提示词同时明确"英文运行时内容不代表语言信号"。

🪄 改进

• 性能与维护：
  • Models.dev 模型目录改为异步加载，多 MB 的目录快照不再阻塞应用启动；移除未使用的内置 Agent 版本同步键。
  • 托盘模型标签更名为 "AI Provider"；新增开发用 Docker Compose 配置、为 python-slim 镜像补充 xlrd 依赖，并移除 chatspeed-docs 子模块。

🐞 修复

• 模型处理：
  • 带供应商前缀的模型（如 ModelScope 的 Qwen/Qwen3.8-Flash-Next）现在可通过短 ID 回退匹配；新增 ModelScope qwen 传输规则以解析 qwen 思考适配器，避免 reasoning_effort 与 thinking_budget 冲突。
  • 模型选择变化时正确应用模型覆盖：代理别名解析到第一个真实后端模型并填充 contextSize、maxTokens、thinking 和 temperature，初始化时同样生效。
• Workflow：
  • 消息列表滚动处理集中化：保留窗口锚点、清除过期滚动目标，用户滚动时立即取消待执行的自动跟随滚动，在消息裁剪、流式输出、窗口缩放与拖拽滚动条时保持阅读位置稳定。
  • 语言检测现在会解析聊天 JSON 信封提取可见内容，并加强系统提示词；clear-context 后的检测用量归属到当前 segment；MCP 执行对话框参数按必填优先排序；修复 overlay 高度与圆角；i18n 键重新排序。
• 依赖：
  • @xmldom/xmldom 升级至 0.8.15。