ArtlexYoung/dsh-super-code

一个更快、更节省、更聪明的 DeepSeek Harness 编码插件。简单任务直接做，复杂任务按需组织专业 Agent；任务状态和有来源的经验可以留到后续对话，不必反复交代。

A faster, more efficient coding plugin for DeepSeek Harness. It handles straightforward tasks directly, brings in specialist Agents when useful, and keeps sourced knowledge available for later conversations.

在同一组 SWE-bench Pro hard-100 的校正结果中，0.2.0 分层记忆候选相比官方 minimal，平均回答耗时减少 41%、每题总 token 减少 20%，通过率从 44% 提高到 52%。完整结果与限制

On the same SWE-bench Pro hard-100 set, the corrected results for the 0.2.0 memory-hierarchy candidate show 41% less response time, 20% fewer total tokens, and a pass rate of 52% versus 44% for official minimal. Full results and limitations

0.2.0 候选 · What's New

• 分层记忆： 区分本次任务、当前项目和全局偏好；只自动带入少量相关摘要，需要时再读详情。 Layered memory: Separates task state, project knowledge, and global preferences. Only a few relevant summaries enter context; details are read when needed.
• 记忆库： 在右侧新标签查看和搜索记忆；Agent 详情中能看到带入上下文、读取、保存和删除记录。 Memory library: Browse and search memories in a sidebar tab, and see each Agent's recorded context, reads, saves, and removals in its details.
• 宿主兼容： 保留旧版目录预设，并适配新版声明式预设及消息格式。具体版本和验收范围见兼容矩阵。 Host compatibility: Keeps directory presets on older hosts and supports newer declarative presets and message formats. See the verified versions and scope.

0.2.0 当前为候选，尚未发布到 npm；发布前请使用候选安装包。效果数据来自已冻结的分层记忆实验，不是新增 UI 和兼容改动后的又一轮 100 题评测。

Version 0.2.0 is a candidate, not yet published to npm; use the candidate package before release. Its quality results belong to the frozen memory-hierarchy experiment, not a new 100-task run of the integrated UI and compatibility changes.

开始使用 · Get Started

需要 Node.js 22 或更高版本，以及已验证的 Harness 版本。

Use Node.js 22 or later and a verified Harness version.

1. 安装： 在插件市场搜索 dsh-super-code，或运行以下命令安装 npm 已发布版。

   Install: Search for dsh-super-code in the plugin marketplace, or install the published npm version:

   dsh plugin --profile web add dsh-super-code
   

   候选包可用 dsh plugin --profile web add ./dsh-super-code-0.2.0.tgz 安装。 For the candidate, use dsh plugin --profile web add ./dsh-super-code-0.2.0.tgz.

2. 选中预设： 打开“设置”→“Agent 预设”，选择 super-code。

   Select the preset: Open “Settings” → “Agent Presets” and choose super-code.

3. 开始对话： 新建会话，描述你要完成的编码任务。模型、推理等级、权限和工具沿用 Harness 的设置；子 Agent 默认继承父 Agent 的模型，不用另配模型池。

   Start a conversation: Open a new session and describe the task. Models, reasoning effort, permissions, and tools come from Harness; child Agents inherit the parent's model by default.

没看到预设？在“设置”→ Super Code 查看状态；旧插件版本的入口在“插件配置”中。安装与排查

Can't find the preset? Open “Settings” → Super Code. Older plugin versions place this under “Plugin configuration”. Setup and troubleshooting

右侧执行树显示各 Agent 的状态、用量和缓存率。拖动平移、滚轮缩放；点节点查看详情，不切换主对话。右侧新标签页中的“记忆库”可以查看当前项目和全局偏好。界面与记忆说明

The Agent tree shows status, usage, and cache hit rate. Drag to pan, scroll to zoom, and click a node for details without leaving the main conversation. Open “Memory library” from the sidebar's new-tab page to view project knowledge and global preferences. Interface and memory guide

效果 · Results

同一组 100 题，均完成官方评分；评分环境异常已使用原补丁复核。时间不含官方评分，token 包含缓存输入，并不等于费用。

All four versions received official scores for the same 100 tasks. Grading anomalies were rechecked with the original patches. Time excludes grading; tokens include cache reads and are not a monetary cost estimate.

指标 / Metric	官方 minimal	0.0.9	0.1.2	0.2.0 分层记忆
通过 / Passed	44/100	48/100	52/100	52/100
平均耗时 / Minutes per task	25.69	16.46	16.01	15.15
平均总 token / Tokens per task	11,578,047	9,623,217	9,938,176	9,299,094
模型调用/题 / Model calls	124.85	92.78	99.78	93.83

0.2.0 与 0.1.2 通过数相同，但有 6 题改善、6 题回退，不代表每道题都更好。这是已用于诊断的历史题集，各轮宿主版本也不同，不能把全部差异归因于插件或记忆功能。多维指标、逐题证据与统计口径

Version 0.2.0 ties 0.1.2 overall, with six gains and six regressions—not an improvement on every task. This is a historically exposed diagnostic set, and host revisions differ between runs; the results do not isolate the effect of the plugin or memory alone. Detailed metrics, per-task evidence, and methodology

使用说明 · Usage · 兼容性 · Compatibility · 版本记录 · Changelog