| Name | Modified | Size | Downloads / Week |
|---|---|---|---|
| Parent folder | |||
| README.md | 2026-09-06 | 4.4 kB | |
| v2.2.0 source code.tar.gz | 2026-09-06 | 10.4 MB | |
| v2.2.0 source code.zip | 2026-09-06 | 10.9 MB | |
| Totals: 3 Items | 21.4 MB | 1 | |
DeepCode v2.2.0 makes long sessions dependable and puts the model under your control: sessions survive resume, compaction, and a second window; settings, models, and reasoning are declared rather than guessed; sub-agents and MCP servers start when needed and stop cleanly; and the Desktop speaks Simplified Chinese. It also folds in a month of community fixes and security hardening.
Highlights
- Sessions that hold up. Resuming a session restores what the agent did, not only what it said; compaction keeps the recent tail and can be triggered on demand with
/compact; context pressure is measured from the provider's own usage; and two windows on one session no longer crash each other — one live writer per session, enforced. - Your models, declared. A dsh-style settings dialog with per-model declarations (context window, capacities, reasoning ladder), live model discovery per connection, a Session-level context-window cap (
/context 64k), and keys that reach exactly the requests they were resolved for. - A terminal that reads like an app. Every bare command opens a selector, the TUI has a status line and tool cards, and a new turn no longer discards the prompt cache.
- Sub-agents and MCP, on demand. Delegate to Codex or Claude Code as sub-agents, hold follow-up conversations with native ones, activate MCP servers lazily, and let hooks see the whole lifecycle (
SessionEnd,PreCompact). - Memory with boundaries. Compaction leaves a note in workspace memory, and injected memory content sits inside an escaped
<untrusted-data>boundary, so a poisoned note cannot forge instructions. - Desktop in Simplified Chinese, eight themes including a high-contrast one, an appearance scale that actually resizes the app, and a project switch that no longer strands the MCP catalog.
- Hardening. Command-executor bypasses closed, native file operations routed through the sandbox, Windows private files protected by a fail-safe ACL, Plugin manifests warned about plaintext credentials, and dependency advisories (pip, pypdf, browserslist) cleared.
Thanks
This release includes contributions from @atirna, @lihuiyang1024, @rifkir23, @MarkHe1222, @imshunsuke, and @raymondginger2018-sudo — thank you.
Full changelog: https://github.com/HKUDS/DeepCode/compare/v2.1.0...v2.2.0
DeepCode v2.2.0 让长会话变得可靠,并把模型真正交到你手里:Session 经得住恢复、压缩和第二个窗口;设置、模型与推理档位改为显式声明而非猜测;子代理与 MCP 服务器按需启动、干净退出;Desktop 说简体中文。同时合入了一个月的社区修复与安全加固。
亮点
- 站得住的 Session。 恢复会话时找回的是 Agent 做过什么而不只是说过什么;压缩保留最近的尾部,也可用
/compact按需触发;上下文压力按服务商报告的真实用量测量;同一 Session 开两个窗口不再互相崩溃——每个 Session 只有一个活跃写入者,由系统锁保证。 - 你的模型,显式声明。 dsh 风格的设置对话框,逐模型声明(上下文窗口、容量、推理档位),每个连接可实时发现模型,Session 级上下文上限(
/context 64k),密钥只到达为它解析的请求。 - 像应用而不是日志的终端。 每个裸命令都打开选择器,TUI 有状态行与工具卡片,新的 Turn 不再丢掉提示缓存。
- 按需的子代理与 MCP。 可把任务委派给 Codex 或 Claude Code 子代理,与原生子代理进行后续对话,MCP 服务器懒启动,Hooks 看得到完整生命周期(
SessionEnd、PreCompact)。 - 有边界的记忆。 压缩会在工作区记忆里留下一条笔记;注入的记忆内容包在转义过的
<untrusted-data>边界里,被投毒的笔记无法伪造指令。 - 简体中文的 Desktop,八套主题(含高对比度),真正缩放整个应用的字号偏好,以及切换项目不再让 MCP 目录卡住。
- 安全加固。 堵上命令执行器的绕过,原生文件操作走沙箱,Windows 私有文件获得 fail-safe 的 ACL,插件清单对明文凭据发出警告,并清除了 pip、pypdf、browserslist 的依赖公告。
致谢
本版本包含来自 @atirna、@lihuiyang1024、@rifkir23、@MarkHe1222、@imshunsuke 与 @raymondginger2018-sudo 的贡献——谢谢你们。