
Apageoflove/DSH-arena
30Last commit Aug 15, 2026
DSH-arena DSH plugin
DSH Arena is a plugin that lets you compare multiple candidate profiles against the same task and repository snapshot. It filters out runs that fail hard gates, performs Pareto ranking, and generates reports with automatic credential sanitization. It also includes a web workbench for interactive experimentation.
How to install the DSH-arena DSH plugin
dsh plugin --profile arena add dsh-arena-0.1.15.tgzCopying does not run this command. Review the repository and version before installing the DSH-arena DSH plugin.
DSH-arena DSH plugin data source
DSH-arena DSH plugin snapshot date: Aug 16, 2026
discovered
What the DSH-arena DSH plugin can do
- Atomic persistent storage with concurrent queue limits, abort, and timeout
- Deterministic hard gates that fail profiles immediately without entering ranking
- Pareto ranking and winner summary across quality, duration, token, and cost metrics
- Web matrix interface with real-time refresh via Shadow DOM components
- Adapter-driven execution without guessing undocumented DSH headless commands
- Export reports (JSON, JSONL, Markdown) with automatic sanitization of authorization, token, key, secret, password, and cookie fields
- Offline report import: load a local Arena JSON report and render it with audit alerts, no upload or telemetry
Where the DSH-arena DSH plugin fits
- Compare multiple AI model profiles on the same task to find the best performance
- Run local experiments with hard gate constraints to filter out poor candidates
- Generate Pareto-optimal recommendations and winner summaries for decision-making
- Audit experiment runs with credential leak detection in exported reports
- Analyze offline reports without network access by importing Arena JSON files
Who the DSH-arena DSH plugin is for
- DSH users who want to evaluate and compare different model configurations
- Developers building custom evaluation pipelines that need local persistence and ranking
- Researchers running head-to-head comparisons of AI models on private tasks
DSH-arena DSH plugin limitations
- Only implements the documented Cordis contract (name, apply, ctx, Config schema); does not use undocumented DSH internal interfaces
- Real DSH Web/Headless E2E tests have not been completed yet, so 100% compatibility is not claimed
- Currently private: true and not published to npm; requires building from source or using a local tarball
- Persistence defaults to .dsh-arena/state.json within dataRoot; paths outside dataRoot are rejected
- Requires Node 24 for building (TypeScript stripping), but runtime may work on other versions
- The executor adapter must be provided by the caller; the plugin does not invent DSH headless commands or initiate model calls
DSH-arena DSH plugin: from the repository README
Quoted from the Apageoflove/DSH-arena README, the upstream source of the DSH-arena DSH plugin. Copyright remains with the original authors.
DSH Arena 是 DeepSeek Harness(DSH)的本地实验与评估工作台插件。它拿同一个任务、同一份仓库快照去对比多个候选 profile:先把没过硬性门槛(hard gate)的跑法拦下来,再做 Pareto 排序;导出报告时自动脱敏常见凭据;还带一个 Web 实验台。   ## 现在能做什么 - 原子化的本地持久化存储,运行队列有并发上限,支持中止和超时 - 确定性 hard gates:不满足直接判失败,不进入排序 - Pareto 排名与 winner 汇总 - Web 矩阵界面,实时刷新 - 执行走 adapter 驱动,不猜、不调 DSH 没文档化的 headless 命令 ## 兼容性边界 - 只按官方文档化的 Cordis 契约写:`name`、`apply(ctx, config)`、Schemastery 的 `Config` schema - 不碰 DSH Client/Host 没文档化的内部接口;能力识别用特性探测(`ctx.arena`),不用私有槽位 - 真实 DSH 上的 Web / Headless E2E 还没跑通,所以本仓库不声称 100% 兼容。等真机验证过了再改口 ## 构建 Node 24 自带 TypeScript 类型剥离,构建不需要额外依赖: ```powershell cd dsh-arena npm run build npm test ``` 产物是 `lib/index.js`(host 入口)和 `lib/client.js`(浏览器入口)。package.json 里声明了 `dsh.bundle.patch` 和 `dsh.client`;目前 `private: true`,还没发 npm。 ## 装进 DSH 先构建: ```powershell npm run build ``` 然后在官方 DeepSeek Harness 检出目录里执行(或把 `pnpm dsh` 换成全局安装的 `dsh` 命令): ```powershell pnpm dsh plugin --profile arena add <本目录路径> pnpm dsh --profile arena --dump-config pnpm dsh --profile arena web ``` 不想走源码目录安装的话,先 `npm pack` 打出本地 tgz,再直接 add 那个文件: ```powershell pnpm dsh plugin --profile a
Read the full READMERepository license: MIT
DSH-arena DSH plugin questions
How do I install DSH Arena?
First build the plugin from source: run `npm run build` in the dsh-arena directory. Then in your DeepSeek Harness checkout, run `pnpm dsh plugin --profile arena add <path-to-dsh-arena>` or use `npm pack` to create a tarball and add it with `pnpm dsh plugin --profile arena add dsh-arena-0.1.15.tgz`. The plugin is not yet published on npm, so local installation is required.
What are hard gates in DSH Arena?
Hard gates are deterministic criteria that a candidate profile must satisfy before being considered for ranking. If a profile fails a hard gate, it is immediately marked as failed and does not proceed to Pareto sorting. This allows you to enforce minimum performance thresholds early in the evaluation process.
Can I use DSH Arena without a DSH host?
No, DSH Arena is a plugin for DeepSeek Harness and requires a DSH environment to run. It extends DSH's functionality by adding experiment workbench capabilities. You cannot use it standalone; you need to install it into a DSH client/host.
How does credential sanitization work in reports?
When exporting reports in JSON, JSONL, or Markdown format, DSH Arena automatically scans field names and values for patterns like authorization, token, key, secret, password, and cookie. It sanitizes those values before writing the output, reducing the risk of leaking sensitive information.
Is DSH Arena compatible with headless DSH?
The plugin has not been fully tested with headless DSH. The README states that Web/Headless E2E tests have not been run yet, so 100% compatibility is not claimed. It is designed to work with the documented Cordis contract and does not rely on undocumented internal interfaces, but actual headless usage may require further validation.