DeepCode Files

DeepCode v2.3.0 runs the TUI, Desktop, and a new browser client on one local background service, so work you start in one interface keeps running and can be followed and approved from another. It adds voice dictation, Opper and Amazon Bedrock, correct DeepSeek V4 numbers, and a set of opt-in protections for model traffic that passes through a relay. It also fixes a SQLite locking bug behind occasional "database disk image is malformed" errors and folds in three weeks of community work.

Highlights

• One local service, three interfaces. deepcode, deepcode desktop, and the new deepcode web all connect to the same local background service. Closing a client leaves running work in the background; reconnect from another interface to follow the conversation and answer approvals. Model connections you save work in all three, and you can test streaming and tool calls before relying on a custom server. Manage the service with deepcode service. (#212)
• Dictate your prompt. Add a dictation block to your config and the Desktop and Web composer gains a microphone. Speech is transcribed by a Parakeet server or the local parakeet-mlx command and inserted at the cursor for you to review. (#240)
• More providers, truer numbers. Opper (one EU-hosted, OpenAI-compatible key for many vendors) and Amazon Bedrock join the templates, and DeepSeek V4 models use their own 1M-token window and pricing instead of V3's. (#242, [#208], [#249])
• Desktop polish. Drag the conversation's edge to resize it, import a VS Code color theme, and pick from fonts that are actually installed. (#238, [#207], [#239])
• Opt-in hardening for relayed model traffic. providers.egress limits which hosts receive your prompts; DEEPCODE_COMMAND_SCREEN=strict refuses piped remote scripts, unknown package indexes, and look-alike package names; DEEPCODE_BASH_SCRUB_ENV=1 keeps provider keys out of shell commands; DEEPCODE_TRANSPARENCY_LOG=1 makes the LLM log tamper-evident; DEEPCODE_KEYRING=1 reads keys from the OS keychain. All are off by default. The legacy command executor now applies the destructive-command screen as well, and each session logs its resolved permission posture. (#228–#232, [#246])
• Leaner memory, clearer signals. MEMORY.md can be a short index of topic files that the agent reads on demand, kept deduplicated and within budget; mcp list flags servers worth a second look; the agent gets a one-time reminder when a call keeps returning the same result; a misspelled tool name gets a "did you mean"; python -m core.observability.llmops summarizes cost and latency; and schedules can skip the model call when nothing they watch has changed. (#233, [#241], [#247], [#215], [#221], [#227])
• Steadier everywhere. The state database no longer drops its SQLite locks when several DeepCode processes share it; the memory data boundary now catches every spelling of its closing tag; and Windows fixes landed for GBK consoles, .git/info/exclude, private-storage permissions, and the entry point. (#243, [#210], [#237], [#235], [#236])

Upgrading

pip install -U deepcode-hku. From this release on, deepcode service prepare-upgrade --output <path> stops the background service and snapshots its state before an upgrade; see Upgrading DeepCode.

Thanks

This release includes contributions from @Dev-next-gen, @EduCosta85, @Felixkw12, @lihuiyang1024, and @raymondginger2018-sudo — thank you.

Full changelog: https://github.com/HKUDS/DeepCode/compare/v2.2.0...v2.3.0

DeepCode v2.3.0 让 TUI、Desktop 和新增的浏览器客户端共用一个本地后台服务：在一个界面里开始的任务会在后台继续运行，并能在另一个界面里跟进和审批。本版本新增语音输入、Opper 与 Amazon Bedrock 模板、正确的 DeepSeek V4 参数，以及一组针对经过中转的模型流量、需要主动开启的防护。它还修复了一个 SQLite 锁的问题（此前偶发"database disk image is malformed"错误的原因），并合入了三周来的社区贡献。

亮点

• 一个本地服务，三种界面。 deepcode、deepcode desktop 与新增的 deepcode web 都连接同一个本地后台服务。关闭客户端后，正在运行的工作留在后台；从另一个界面重新连接，即可继续跟进对话、回应审批。保存的模型连接在三种界面通用，使用自定义服务前可以先测试流式输出与工具调用。用 deepcode service 管理后台服务。（#212）
• 用语音输入提示词。 在配置中加入 dictation 后，Desktop 与 Web 的输入框会出现麦克风。语音由 Parakeet 服务或本地 parakeet-mlx 命令转写，插入到光标处，发送前可以检查修改。（#240）
• 更多服务商，更准的参数。 新增 Opper（用一个托管在欧盟的 OpenAI-compatible 密钥调用多家厂商的模型）与 Amazon Bedrock 模板；DeepSeek V4 使用自己的 100 万上下文与价格，不再沿用 V3 的数值。（#242、#208、#249）
• Desktop 细节打磨。 拖动会话区边缘即可调宽，可以导入 VS Code 配色主题，字体选择只列出本机已安装的字体。（#238、#207、#239）
• 可选的中转流量加固。 providers.egress 限定哪些主机能收到你的提示词；DEEPCODE_COMMAND_SCREEN=strict 拒绝下载后直接执行的远程脚本、未知的包索引和仿冒包名；DEEPCODE_BASH_SCRUB_ENV=1 不把服务商密钥交给 shell 命令；DEEPCODE_TRANSPARENCY_LOG=1 让 LLM 日志可以校验是否被篡改；DEEPCODE_KEYRING=1 从系统钥匙串读取密钥。以上默认全部关闭。另外，旧版命令执行器也会先经过危险命令筛查，每个 Session 会在日志里记录最终生效的权限状态。（#228–#232、#246）
• 更轻的记忆，更清楚的提示。 MEMORY.md 可以只是一份主题文件索引，由 Agent 按需读取，并保持去重、不超预算；mcp list 会标出值得再看一眼的 server；同一调用反复返回相同结果时，Agent 会收到一次提醒；工具名写错时会提示"是否想用"；python -m core.observability.llmops 汇总成本与延迟；定时任务在监视的内容没有变化时可以跳过模型调用。（#233、#241、#247、#215、#221、#227）
• 各平台更稳定。 多个 DeepCode 进程共用状态数据库时不再丢失 SQLite 锁；记忆的数据边界能识别结束标签的各种写法；Windows 上修复了 GBK 控制台、.git/info/exclude、私有存储权限与入口文件的问题。（#243、#210、#237、#235、#236）

升级

pip install -U deepcode-hku。从本版本起，升级前可以用 deepcode service prepare-upgrade --output <路径> 停下后台服务并为状态做快照，详见升级指南（英文）。

致谢

本版本包含来自 @Dev-next-gen、@EduCosta85、@Felixkw12、@lihuiyang1024 与 @raymondginger2018-sudo 的贡献——谢谢你们。