Releases: lessweb/deepcode-cli
Release list
v0.1.34
What's Changed
- refactor(ui): 修改退出后还展示状态提示信息 by @hqwlkj in #201
- style(cli): 优化 AskUserQuestionPrompt 组件选项描述的布局间距 by @hqwlkj in #209
- 新增并优化 Plan Mode 的命令、交互界面及权限限制,并补充 Plan Mode 使用说明 (#232)
- 增加 LLM 错误处理、脱敏与详细日志能力 (#239)
- 优化
read工具,移除pages参数且不再以 Base64 输出 PDF 内容 (#236)
Plan Mode 专题:先规划,再实现
本版本带来并完善了 Plan Mode(规划模式)。面对跨文件重构、新功能开发、架构调整等复杂任务,Deep Code 会先帮助你探索代码库、澄清需求并形成完整方案;在你确认方案前,不会执行修改代码的操作。
如何使用
- 在输入框中按
Shift+Tab,可在 Plan Mode 与 Default Mode 间切换。 - 输入
/plan,或在/命令菜单中选择/plan。 - AI 完成规划后会输出实施方案,并提供三个选择:直接实施、继续完善方案,或切回默认模式。
规划过程更可控
Plan Mode 将协作过程拆为三个阶段:
- 环境摸底:阅读文件、搜索代码和检查配置,建立对项目的理解。
- 意图对齐:明确目标、范围、约束和关键取舍。
- 方案推敲:确定实现路径、边界条件与测试策略,形成可执行的方案。
在规划模式中,文件编辑、写入、删除、代码生成与 Git 历史修改等操作会被限制;即使本地权限配置为自动放行,这些操作也会强制要求确认,确保规划阶
段不会意外改动代码。
实验结果
在使用 deepseek-v4-pro 完成同一算法实现任务的对比实验中,Plan Mode 的测试通过率达到 75%(3/4),而未使用 Plan Mode 的对照组为 0%(0/4)。
Plan Mode 会投入更多时间与推理成本:墙钟时间增加约 32%,费用增加约 26%(约 ¥0.39)。但它通过前置规划、分阶段验证与更聚焦的上下文,在复杂任务中带来了显著更高的可用性。
Plan Mode 推荐用于高难度算法实现、架构设计、多模块协同开发和需要端到端验证的任务;对于简单脚本、低风险配置调整或修正文案等任务,可直接使用默认模式。
更多说明:Plan Mode 文档 | 对比实验报告
Full Changelog: v0.1.33...v0.1.34
v0.1.31
What's Changed
- chore(deps-dev): bump esbuild and tsx by @dependabot[bot] in #174
- chore: 优化deepcode-self-refer skill
- chore: 优化skill-digester skill
Full Changelog: v0.1.30...v0.1.31
v0.1.30
What's Changed
- fix(mcp): fix Windows MCP spawn double-quoting that breaks all MCP servers by @dengmik-commits in #164
- feat: enhance Windows MCP command quoting and add tests for cmd metacharacters
- feat: add plan mode skill and enhance shell init command tests
- feat: add support for implicit invocation control in skills
- feat: update default skill loading so enabledSkills can skip the built-in skills
Full Changelog: v0.1.29...v0.1.30
v0.1.29
What's Changed
- 修复提示输入的换行、光标定位与 busy 状态显示 by @qorzj in #171
- feat: implement bundled built-in skills
- feat: add docs/session-persistence.md
- feat: implement
enabledSkillssupport in settings.json - feat: add raw mode shortcut
ctrl+r
New Contributors
Full Changelog: v0.1.28...v0.1.29
v0.1.28
What's Changed
- refactor: extract OpenAI message converter from SessionManager by @Lellansin in #140
- feat: session name can be edited now by @Feiry-zZ in #159
- feat(ui): 优化 PromptInput 组件的光标显示与布局 by @hqwlkj in #161
- 升级bash tool,支持后台运行(
run_in_background),可解决使用playwright测试场景下,启动server导致流程阻塞的问题 - 修复bash tool多行参数的渲染问题,现在一定会显示LLM提供的description信息
- Enhance cursor handling in PromptInput component to fix IME composition anchoring
- 更新karpathy-guidelines.md,增加提示:
Apply these guidelines silently. Do not cite this document, its title, or guideline names in user-facing responses. - Improve the Markdown underscore rendering
- Add MCP tool name handling with API-safe names
- Implement temperature support for settings.json
- Agent Skills相关优化:
- 修改Skills的加载路径,使其更符合规范
- 改进skill加载结果,主要是增加
<skill_resources>内容 - 对skills加载路径内的读操作不计入
read-out-cwd的权限限制范围 - identifyMatchingSkillNames的prompt会预加载AGENTS.md
New Contributors
- @dependabot[bot] made their first contribution in #157
- @Feiry-zZ made their first contribution in #159
Full Changelog: v0.1.27...v0.1.28
v0.1.27
What's Changed
- chore(deps): update ink-gradient to 4.0.1 by @fym998 in #135
- chore: 更新API Key not found时的文本显示 by @iamhmx in #137
- 优化edit tool,取消返回findClosestMatch,直接用inferOldStringNotFoundReasonWithLLM生成提示信息
- 改进系统提示词,引入karpathy-guidelines,经验证对于复杂任务可显著提升成功率
New Contributors
Full Changelog: v0.1.26...v0.1.27
v0.1.26
What's Changed
- refactor(ui): 重构代码结构,调整文件路径和导入引用 by @hqwlkj in #122
- fix(session): 修复会话清理内存泄漏并补充回归测试 by @Lellansin in #123
- chore(deps): 更新 ink 依赖到 7.0.4 版本 by @hqwlkj in #125
- fix(prompt-buffer): 修正 getCurrentSlashToken 函数逻辑 by @hqwlkj in #129
- refactor: extract telemetry into separate module with enable/disable toggle by @Lellansin in #130
- Edit 工具增强 — 支持空 old_string 的文件编辑,snippet 处理增强(full-file 支持 + 强制 snippet_id)
- 快照机制增强 - 修复
/undo恢复快照场景的已知问题 - Agent+手动混合修改场景优化 - 基于快照机制检测到手动修改时,自动增加system prompt,可有效防止LLM无脑覆盖手动修改。
Full Changelog: v0.1.25...v0.1.26
v0.1.25
What's Changed
- docs: 更新扩展命令菜单说明和帮助文档 by @hqwlkj in #109
- feat(ui): add bracketed paste with large-paste marker collapsing by @jeoor in #102
- perf: reuse OpenAI client and add undici keep-alive Agent with connection warmup by @Lellansin in #100
- feat(ui): 会话列表支持 Delete 键删除会话 by @xinggitxing in #114
- feat(ui): 增加会话删除及相关UI重置功能 by @hqwlkj in #119
- feat: markdown 表格闭合边框渲染 + CJK/emoji 宽度适配 by @dengmik-commits in #115
- fix(permission): 处理权限拒绝状态与界面更新 by @hqwlkj in #120
- feat: implement checkpoints store only explicit Write/Edit file paths
- 实现权限机制,详见 docs/permission.md
New Contributors
- @jeoor made their first contribution in #102
- @xinggitxing made their first contribution in #114
Full Changelog: v0.1.24...v0.1.25
v0.1.24
What's Changed
- fix(ui): 修正组件路径拼写错误 by @hqwlkj in #93
- fix: resolve CJK composition bug on iOS terminals (backspace packet splitting) by @liante0904 in #94
- feat: MCP 服务器手动重连功能 by @dengmik-commits in #84
- refactor: Extract dropdown components by @hqwlkj in #97
- refactor(ui): 使用 resetPromptInput 简化撤销和回绕处理 by @hqwlkj in #101
- 新增
/undo斜杠命令,同时实现了基于git的快照回滚机制 - 适配七牛/火山引擎等deepseek模型时可能出现的tool call字段相关问题
New Contributors
- @liante0904 made their first contribution in #94
Full Changelog: v0.1.23...v0.1.24
v0.1.23
What's Changed
- feat: 新增
/raw命令交互与终端消息直出 by @hqwlkj in #89 - feat(notify): pass STATUS, FAIL_REASON, BODY, TITLE as env vars to notify hook by @Lellansin in #90
- 修复Windows系统下Bash tool⽆法杀死进程组的问题
- 实现Bash tool的超时机制,默认10分钟,可在ctrl+o界面调整
Full Changelog: v0.1.22...v0.1.23