ChatSpeed v2.0.14

[2.0.14]

🚀 New Features

• Model catalog:
  • Added a local Models.dev-backed model catalog with capability adapters, official limits, documentation sources, and usage pricing.
  • Added preset text AI providers with logos and endpoints; the Anthropic preset uses the Claude protocol, and model settings can be prefilled from catalog profiles.
  • Enriched the catalog with new model families and profiles (including qwen3, hunyuan-hy3, glm-5.2/5.3, kimi-k3, and stepfun-flash-multimodal), plus GPT-5.6 reasoning summaries and gpt-oss reasoning support.
  • Added a metadata sync workflow to keep provider catalog data up to date, with heuristic fallback when catalog data is unavailable.
• Codex-compatible Responses API:
  • Responses API requests now mirror the official Codex CLI client surface, including Codex-style headers, client_metadata, and stable per-conversation UUID cache identity for prompt-cache affinity.
  • Added prompt_cache_key support with workflow, conversation, and fallback priority tiers.
  • Models can opt out of the Responses API via a chatCompletions preference in model settings.
• MCP tool management:
  • MCP tool visibility is now driven by the runtime ToolManager, allowing any discovered MCP tool to be selected directly.
  • Isolated MCP tool configuration with workflow sync, simplified tool call names through an alias registry, and showed MCP tool display and server names in the input area.
• Sub-agent approvals:
  • Approval requests and resolutions are now bridged between parent and child workflows, with pending approvals restored on reload.
  • Added a session message pane to view sub-agent activity with a live status badge.
• Workflow:
  • Authorized directories can now be managed from slash commands.
  • The final-review card is hydrated from the structured reviewer result, showing reviewer RESULT and COST directly.
  • Improved the sandbox mode picker and new task button.

🪄 Improvements

• Settings and styling:
  • Compacted the agent, model, sandbox, and skill settings UI and copy.
  • Workflow styles now use design tokens; corrected the automation editor icon and input font size.

🐞 Bug Fixes

• Sub-agent reliability:
  • Fixed sub-agent approval handling: approvals arriving while another tool runs are deferred to the wait state, tool messages are reconciled by tool_call_id, pending approvals are cleared on terminal states, and bulk approval counts are derived from the canonical pending set.
  • Centralized sub-agent runtime cleanup so gateway channels and registry entries are released on every terminal path, and refined sub-agent card status display.
• Workflow reliability:
  • Folded MCP tools now execute after loading; duplicated arguments are stripped from MCP approval previews; home paths are expanded in directory commands.
  • Kept the reading position stable in the message list window, refined ask_user tool-call matching in tail rewind logic, and fixed invisible workflow messages and legacy assistant tool-payload responses.
  • Improved completion report guidance and argument handling, enforced the todo initialization threshold, persisted failed workflow state on resume/start errors, and skipped forced tool choice in thinking mode (DeepSeek compatibility).
  • Refined shell output reduction to keep only shorter reductions, with on-demand temp-file read hints and resumable truncated previews.
• Model handling:
  • Added provider-prefix and short-id fallbacks for model resolution and improved provider model import and catalog-backed model settings.
• Proxy and protocol:
  • Improved CCProxy token accounting and stream error handling, normalized outbound temperature to two decimals, and added indexes to ccproxy_stats in the v4 migration.
• Maintenance and security:
  • Bumped nanoid to 3.3.18 to address CVE-2026-67213, isolated development logs in dev_data, avoided logging long API key suffixes, and removed Windows build warnings.

---

🚀 新功能

• 模型目录：
  • 新增基于 Models.dev 的本地模型目录，包含能力适配器、官方限制、文档来源和用量定价。
  • 新增预置文本 AI 提供商（含图标和接口地址）；Anthropic 预设使用 Claude 协议，模型设置可从目录配置预填充。
  • 目录新增多个模型家族与配置（包括 qwen3、hunyuan-hy3、glm-5.2/5.3、kimi-k3、stepfun-flash-multimodal），并支持 GPT-5.6 推理摘要和 gpt-oss 推理。
  • 新增元数据同步流程以保持提供商目录数据更新；目录数据不可用时自动回退到启发式判断。
• Codex 兼容的 Responses API：
  • Responses API 请求现在对齐官方 Codex CLI 客户端，包括 Codex 风格请求头、client_metadata，以及按会话稳定的 UUID 缓存标识以提升提示缓存命中。
  • 新增 prompt_cache_key 支持，按 Workflow、对话和回退三级优先级生成。
  • 模型可在设置中通过 chatCompletions 偏好选择不使用 Responses API。
• MCP 工具管理：
  • MCP 工具可见性改由运行时 ToolManager 驱动，任何已发现的 MCP 工具都可直接选用。
  • 隔离 MCP 工具配置并与 Workflow 同步；通过别名注册表简化工具调用名称；输入区域显示 MCP 工具显示名和服务器名。
• 子 Agent 审批：
  • 审批请求与结果现在会在父/子 Workflow 之间桥接，重新加载后可恢复待审批项。
  • 新增子会话消息面板查看子 Agent 活动，并显示实时状态徽标。
• Workflow：
  • 支持通过斜杠命令管理授权目录。
  • 最终审查卡片改为从结构化审查结果填充，直接显示审查 RESULT 和 COST。
  • 改进沙箱模式选择器和新任务按钮。

🪄 改进

• 设置与样式：
  • 精简 Agent、模型、沙箱和技能设置界面的布局与文案。
  • Workflow 样式改用设计令牌；修正自动化编辑器图标和输入框字号。

🐞 修复

• 子 Agent 可靠性：
  • 修复子 Agent 审批处理：其他工具运行期间到达的审批会延迟到等待状态处理，工具消息按 tool_call_id 收敛生命周期，终态时清除待审批项，批量审批数量改为基于规范待审批集合计算。
  • 统一子 Agent 运行时清理，确保网关通道和注册表条目在所有终态路径上被释放；优化子 Agent 卡片状态显示。
• Workflow 可靠性：
  • 折叠的 MCP 工具加载后可正常执行；MCP 审批预览去除重复参数；目录命令中展开主目录路径。
  • 消息列表窗口保持阅读位置稳定；优化尾部回退逻辑中的 ask_user 工具调用匹配；修复消息不可见和旧版助手工具载荷响应的问题。
  • 改进完成报告的指引和参数处理，强制执行 todo 初始化阈值，恢复/启动失败时持久化错误状态，思考模式下跳过强制工具选择（DeepSeek 兼容）。
  • 优化 Shell 输出缩减：仅保留更短的缩减结果，提供按需读取临时文件的提示，截断预览可从覆盖行偏移继续。
• 模型处理：
  • 模型解析新增提供商前缀和短 ID 回退；改进提供商模型导入和基于目录的模型设置。
• 代理与协议：
  • 改进 CCProxy 的 Token 统计和流式错误处理；出站 temperature 规范化为两位小数；v4 迁移为 ccproxy_stats 添加索引。
• 维护与安全：
  • 升级 nanoid 至 3.3.18 以修复 CVE-2026-67213；开发日志隔离到 dev_data；避免记录过长 API Key 后缀；消除 Windows 构建警告。