kevin/deepseek-harness
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

docs: refresh root README for current capabilities

Browse Source
This commit is contained in:
NI0317
2026-07-30 15:52:20 +08:00
parent 69214b4708
commit 544d3f8267
7 changed files with 147 additions and 105 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.agents/notes/implemented/process/2026-07-22-product-first-root-readme.i18n.yaml
+6
View File
@@ -0,0 +1,6 @@
# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write .agents/notes/implemented/process/2026-07-22-product-first-root-readme.md
2026-07-22-product-first-root-readme.md: bed7d3b49c58a59fa0f6d6637c9852b9fb615b63
2026-07-22-product-first-root-readme.zh.md: fbfb726b2c1a821b323de2fbf419dc97920a6878
.agents/notes/implemented/process/2026-07-22-product-first-root-readme.md
+35
View File
@@ -0,0 +1,35 @@
# Agent Note: Product-first root README
Status: implemented
English | [中文](2026-07-22-product-first-root-readme.zh.md)
## Problem
The root README is the repository's product front door, but a product-only coding-agent description hides the SDK and current runtime breadth, while an SDK-first package inventory delays the shortest path to a working agent. Commands, capability claims, and entry-point descriptions also drift when the README is treated as general marketing instead of a maintained product contract.
## Decision
The root README defines DeepSeek Harness as a plugin-native coding-agent runtime that ships both a composable SDK and the assembled `dsh` agent. It separates mission from shipped facts and leads with the supported one-line installer.
A note before the installer thanks early users, states plainly that the internal preview remains unfinished, has a low overall level of completion, and falls below the experience the team wants to deliver. It invites direct reports of failures, confusion, and friction, assigning those shortcomings to the product rather than the user, while the adjacent pre-release warning keeps the compatibility boundary explicit.
The README names the TUI, Web, headless, ACP, and Python/JSON-RPC entry points with commands or owning links. It summarizes capabilities by coding, orchestration, and operational families, while stating that each composition selects its own plugins. Exhaustive package and service inventories stay in the generated graphs and package-group documentation.
Plugin-native is the organizing principle rather than a slogan: the README ties replaceable services and typed events to composition through `cordis.yml`, and ties model-visible behavior, persistence, replay, queries, telemetry, and UI projections to the authoritative session log. Detailed contracts remain with the architecture, CLI, examples, cookbook, and generated catalogs.
The English and Chinese README sides share the same technical structure. Their community sections point to the primary channel for each language audience. The documentation website keeps its separate user-guide landing page; the repository README is not added to that projection.
## Alternatives considered
**Present only the assembled coding agent.** This gives the shortest product pitch but makes the SDK, alternate front doors, and replaceable runtime seams look incidental even though they are shipped repository surfaces.
**Present the repository as an SDK and package catalog.** This exposes implementation breadth immediately but makes a new reader reconstruct the product from package names. The package map and generated capability graph remain the authoritative inventories.
**Use a long marketing page with screenshots, badges, and duplicated tutorials.** Rich media can demonstrate a stable product journey, but it ages separately from commands and source contracts. The root stays compact and links to runnable examples and owned guides.
**Project the root README as the documentation website home page.** A single landing page avoids two narratives, but the website's user guide and the repository's developer/product front door have different navigation and maintenance needs. They remain separate sources linked to the same architecture and guides.
## Consequences
A new reader can install or choose a runtime surface before learning the package topology, while an SDK reader can see the extension model without a generated catalog being copied into prose. The README must change with any affected command, entry point, pre-release boundary, or high-level capability family, and each claim remains reviewable against source or an owning document.
.agents/notes/implemented/process/2026-07-22-product-first-root-readme.zh.md
+35
View File
@@ -0,0 +1,35 @@
# Agent Note: 产品优先的根 README
Status: implemented
[English](2026-07-22-product-first-root-readme.md) | 中文
## 问题
根 README 是仓库的产品入口，但仅以产品视角描述 coding agent（编程智能体）会掩盖 SDK 与当前运行时的广度，而以 SDK 为先的包（package）清单则会把启动可运行 agent 的最短路径放到后面。如果把 README 当作通用营销文案，而非持续维护的产品契约，命令、能力声明和入口描述也会逐渐漂移。
## 决策
根 README 将 DeepSeek Harness 定义为以插件为原生构成单元的 coding agent 运行时，同时交付可组合的 SDK 与组装完成的 `dsh` agent。它将使命定位与已交付事实分开，并首先给出受支持的单行安装命令。
安装命令之前的一则说明感谢早期用户，坦率说明内测版本仍未完成、整体完成度还很低，距离团队希望交付的体验还有差距，并邀请用户直接反馈故障、困惑和所有不顺手之处。它明确这些不足是产品的问题，而非用户的问题；紧接其后的预发布提醒则清楚说明兼容性边界。
README 列出 TUI、Web、Headless、ACP（Agent Client Protocol）以及 Python/JSON-RPC 入口，并为每个入口提供命令或归属文档链接。它按编码、编排和运维三个类别概述能力，同时说明每种组合都自行选择插件。包与服务的完整清单仍由生成图和包分组文档维护。
以插件为原生构成单元是 README 的组织原则，而非一句口号：README 通过 `cordis.yml` 将可替换服务、类型化事件与组合方式关联起来，并明确面向模型的行为、持久化、回放、查询、遥测和 UI 投影都以权威会话日志为基础。详细契约仍由架构文档、CLI（命令行界面）、示例、实操手册（cookbook）和生成目录各自维护。
中英文 README 采用相同的技术结构。两侧的社区章节分别指向各自语言受众的主要交流渠道。文档网站继续使用独立的用户指南首页；仓库 README 不加入该投影。
## 考虑过的替代方案
**只展示组装完成的 coding agent。** 这样能给出最简短的产品介绍，但会让 SDK、其他入口和可替换的运行时 seam 显得无足轻重，尽管它们都是仓库中已经交付的组成部分。
**将仓库呈现为 SDK 和包清单。** 这样能立即展现实现广度，却会迫使新读者从包名反推出产品。包索引与生成的能力图仍是权威清单。
**使用包含截图、徽章和重复教程的长篇营销页面。** 富媒体能够展示稳定的产品使用路径，但其内容会独立于命令和源码契约而逐渐陈旧。根 README 保持紧凑，并链接到可运行示例和各自维护的指南。
**将根 README 投影为文档网站首页。** 使用同一个首页可以避免两套叙事，但文档网站的用户指南与仓库面向开发者和产品的入口在导航和维护需求上并不相同。两者继续作为独立来源，并链接到相同的架构文档和指南。
## 结果
新读者可以在了解包拓扑之前完成安装或选择运行时入口，SDK 读者也能理解扩展模型，而无需把生成目录复制进正文。任何受影响的命令、入口、预发布边界或高层能力类别发生变化时，README 都必须同步更新；每项声明都可以依据源码或归属文档进行评审核验。
README.i18n.yaml
+2 -2
View File
@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write README.md
README.md: f9f7294b42e29132d5cd46c0ab6a5f5265a1d8f3
README.zh.md: 88cbf8522d8f1a183a48dc7e80858d1a0ced8f0f
README.md: 4f87a95ecd921748e5ac8b93884ffe1254cb3ddf
README.zh.md: fee6e4ad1c08ac7a9b99e6d7479454e00cf431cf
README.md
+34 -49
View File
@@ -2,81 +2,66 @@
English | [中文](README.zh.md)
DeepSeek Harness (`dsh`) is an open-source coding agent built on the DeepSeek Harness SDK.
DeepSeek Harness is an open-source, plugin-native runtime for coding agents. This repository ships both the composable SDK and `dsh`, a working agent assembled from the same packages.
It uses an architecture where **everything is a plugin**.
**Mission.** Build capable agent products without hard-wiring product choices into one loop. Models, tools, policy, storage, context, interfaces, and even the loop are [Cordis plugins](docs/architecture.md); the session log is the authoritative record from which model history, persistence, replay, queries, telemetry, and UIs derive.
## Install
## Before you begin, thank you
Install `dsh` with one command:
Thank you for taking the time to try DeepSeek Harness. It is still in internal testing, and it is far from complete. It is nowhere near the product we want to ship. Some features are unfinished, and some parts are rough to use. Problems that show up in real use may lead us to rethink designs we have today.
We will keep working to get these parts right, and we want to hear what using it is actually like. Please tell us plainly where it fails. We also want to know what is confusing or gets in your way. If it does not help you, or makes your work harder, we have not done our job. The specific problems you run into and any suggestions you have will help us decide what to fix first. Thank you for spending time with it before it is ready, and for helping us make it better one step at a time.
> **Pre-release notice:** Package APIs, configuration, and persisted formats may change without compatibility shims until the first tagged release.
## Start in one command
```sh
curl -fsSL https://raw.githubusercontent.com/deepseek-harness/deepseek-harness/master/scripts/install.sh | sh
```
The installer requires `git` and Node `^22.19 || >=24`, offers to install `pnpm` when it is missing, and prompts for a DeepSeek API key.
The installer requires `git` and Node `^22.19 || >=24`, offers to install `pnpm`, prompts for a DeepSeek API key, and launches the TUI in the current directory. It keeps managed checkouts under `~/.dsh/source`; run the same command again to update. [`scripts/install.sh`](scripts/install.sh) documents alternate locations and non-interactive options.
The installer keeps every checkout under `~/.dsh/source`: the master clone at `~/.dsh/source/master` and each install's staging checkout as a git worktree `~/.dsh/source/staging-<timestamp>`. The stable symlink `~/.dsh/source/current` points at the active staging worktree, and `dsh` in `~/.local/bin` links to `current/bin/dsh`, so an upgrade repoints one symlink and the `dsh` on PATH never moves. Re-running the command adds a fresh staging worktree from an updated master and repoints `current` at it. See [`scripts/install.sh`](scripts/install.sh) for alternate install locations and other options.
## Choose a surface
## Use DeepSeek Harness
| Surface | Entry point |
|---|---|
| Full-screen TUI | `dsh` |
| Browser UI | `pnpm run demo:web` from a source checkout, or `dsh web` from a built checkout |
| One-shot headless task | `pnpm run demo:headless "summarize this workspace"`, or `dsh -p "summarize this workspace"` from a built checkout |
| ACP automation server | `pnpm run demo:acp` from a source checkout |
| Python / JSON-RPC SDK | [`python/`](python/README.md) with its bundled runtime |
### Web UI
The one-line installer launches the source-running TUI without a build. The `dsh web` and `dsh -p` entries additionally need the frontend and client bundles from `pnpm run build`; `pnpm run demo:web` performs that build itself. The TUI, Web, and headless entries use the invoking directory as the workspace. See the [`dsh` CLI contract](apps/cli/README.md) for configuration, resume, provider, and workspace details; the [examples](examples/README.md) show the thinner ACP, JSON-RPC, Code Mode, and self-referential compositions.
For the recommended local interface, build the frontend after installation and after each update, then start the Web UI. Resolve the running checkout from the `dsh` launcher so the command holds regardless of which staging worktree is current (the launcher resolves through the stable `current` symlink):
## What ships
```sh
dsh_bin=$(cd "$(dirname "$(command -v dsh)")" && pwd -P)/$(basename "$(command -v dsh)")
while [ -L "$dsh_bin" ]; do
link=$(readlink "$dsh_bin")
case $link in /*) dsh_bin=$link ;; *) dsh_bin=$(cd "$(dirname "$dsh_bin")" && cd "$(dirname "$link")" && pwd -P)/$(basename "$link") ;; esac
done
dsh_dir=$(cd "$(dirname "$dsh_bin")/.." && pwd -P)
pnpm --dir "$dsh_dir" run build && pnpm --dir "$dsh_dir" run build:web
dsh web
```
Capabilities are selected by composition. The repository's shipped plugins cover:
The Web UI is served at `http://127.0.0.1:3080` by default.
- **Coding:** filesystem read/write/edit and search, shell and persistent PTY execution, LSP navigation, web search/fetch, reusable skills, and model-written Code Mode programs.
- **Orchestration:** subagents, background tasks, worker-thread workflows, same-session goals, plan state, todos, and user questions.
- **Operations:** workspace sandboxing and approvals, session persistence/resume/fork/query, compaction and spill, projections, titles, and OpenTelemetry export.
### TUI
Anything visible to the model must be reconstructable from the session log. That makes alternate UIs, persistence backends, replay, and operational tooling consumers of one event stream instead of parallel sources of truth.
Start the full-screen terminal interface:
## Extend the harness
```sh
dsh
```
A swappable capability normally separates its interface, implementation, and consumer. Add or replace a provider behind a service such as `ctx.llm`, `ctx.fs`, `ctx.pty`, `ctx.web`, or `ctx.subagents`; register model-facing behavior through `ctx.tools`; attach policy and request shaping through typed events; compose the result in `cordis.yml` without forking the agent loop.
### Headless
Start with the [first-plugin guide](docs/user/develop/basic/index.md) and [extension cookbook](docs/cookbook/extension-cookbook.md). Use the [architecture](docs/architecture.md) for the system map, the generated [capability graph](docs/capability-seams.md) for current service relationships, and the [package map](packages/README.md) when you need ownership details.
Run one task, print the final answer, and exit:
```sh
dsh -p "summarize this workspace"
```
## Why DeepSeek Harness
Built-in capabilities cover file reading, editing, and search; shell execution; reusable skills; task tracking; subagents and workflows; persistent sessions; and context compaction. The TUI also includes Plan Mode.
- **Everything is a plugin.** Models, tools, policies, storage, context management, and interfaces are composable [Cordis plugins](docs/user/develop/basic/index.md), so deployments can extend or replace behavior without forking the agent loop. See the [architecture](docs/architecture.md) for the underlying design.
- **Code Mode (opt-in).** It exposes a `run_code` tool and a generated TypeScript SDK; only program output re-enters model context. See [Code Mode](packages/core/tools/README.md#code-mode).
- **Self-referential Cordis tools are opt-in.** They let the agent inspect its live runtime and mount or unmount plugins while it runs. See the [Cordis tools](packages/cordis/tool-cordis/README.md).
## Community
Follow <a href="https://x.com/Deepseekharness">DeepSeek Harness on Twitter</a> for project updates.
## Development
## Develop
```sh
pnpm install
pnpm run test:coverage
pnpm run demo:tui
```
Start with the [development guide](docs/development.md) and read the [architecture](docs/architecture.md) before changing packages.
Set `DEEPSEEK_API_KEY` in the environment or root `.env`. The [development guide](docs/development.md) owns setup and validation; read the [architecture](docs/architecture.md) before changing `packages/`, and follow [AGENTS.md](AGENTS.md) when working in this repository.
For agents, follow [AGENTS.md](AGENTS.md).
## Community
DeepSeek Harness is currently pre-release.
Follow <a href="https://x.com/Deepseekharness">DeepSeek Harness on X</a> for project updates.
## License
README.zh.md
+33 -52
View File
@@ -2,85 +2,66 @@
[English](README.md) | 中文
DeepSeek Harness（`dsh`）是一款基于 DeepSeek Harness SDK 构建的开源 coding agent（编程智能体）。
DeepSeek Harness 是一个面向 coding agent（智能体）的开源、插件原生运行时。本仓库同时提供可组合的 SDK，以及由同一组包（package）组装而成、可直接运行的 agent `dsh`。
它采用了**一切皆插件**的架构。
**使命。** 构建能力强大的 agent 产品，而不把产品选择硬编码到单一循环中。模型、工具、策略、存储、上下文、接口，乃至循环本身，都是 [Cordis 插件](docs/architecture.md)；会话日志是权威记录，模型历史、持久化、回放、查询、遥测和 UI 均从中派生。
## 安装
## 使用前，想先说声谢谢
使用一条命令安装 `dsh`：
感谢你愿意花时间试用 DeepSeek Harness。它还在内测，整体完成度不高，也远没有达到我们想交付的样子。有些功能还没做完，有些地方用起来会很粗糙。真实使用中暴露出来的问题，也可能让我们推翻现在的设计。
我们会继续认真把这些地方做好，也希望你能把真实感受直接告诉我们。哪里失败了，哪里让你困惑或不好用，都请直说。如果它没有帮到你，反而给工作添了麻烦，那就是我们没有做好。你遇到的具体问题和任何建议，都会帮助我们判断接下来先改什么。谢谢你愿意在它还不成熟的时候花时间试用，也谢谢你愿意和我们一起把它一点点做好。
> **预发布说明：** 在首个带标签的版本发布之前，包 API、配置和持久化格式可能直接变更，不提供兼容层。
## 一条命令开始
```sh
curl -fsSL https://raw.githubusercontent.com/deepseek-harness/deepseek-harness/master/scripts/install.sh | sh
```
安装器要求系统已安装 `git` 和 Node `^22.19 || >=24`，缺少 `pnpm` 时可代为安装，并会提示输入 DeepSeek API 密钥。
安装器要求系统已安装 `git` 和 Node `^22.19 || >=24`，可代为安装 `pnpm`，会提示输入 DeepSeek API 密钥，并在当前目录启动 TUI。它把受管检出放在 `~/.dsh/source` 下；再次运行同一命令即可更新。其他安装位置和非交互选项见 [`scripts/install.sh`](scripts/install.sh)。
安装器会把所有检出都放在 `~/.dsh/source` 下：master 克隆位于 `~/.dsh/source/master`，每次安装的 staging 检出是一个 git worktree `~/.dsh/source/staging-<时间戳>`。稳定符号链接 `~/.dsh/source/current` 指向当前生效的 staging worktree，`~/.local/bin` 中的 `dsh` 链接到 `current/bin/dsh`，因此升级只需重指一个符号链接，PATH 上的 `dsh` 从不移动。再次运行该命令会基于更新后的 master 新增一个 staging worktree，并把 `current` 重指到它。其他安装位置和选项见 [`scripts/install.sh`](scripts/install.sh)。
## 选择使用方式
## 使用 DeepSeek Harness
| 使用方式 | 入口 |
|---|---|
| 全屏 TUI | `dsh` |
| 浏览器 UI | 在源码检出中运行 `pnpm run demo:web`，或在已构建的检出中运行 `dsh web` |
| 一次性无头任务 | 运行 `pnpm run demo:headless "summarize this workspace"`，或在已构建的检出中运行 `dsh -p "summarize this workspace"` |
| ACP（Agent Client Protocol）自动化服务器 | 在源码检出中运行 `pnpm run demo:acp` |
| Python / JSON-RPC SDK | [`python/`](python/README.md) 及其自带的运行时 |
### Web UI
一行安装命令可直接启动从源码运行的 TUI，无需构建。`dsh web` 和 `dsh -p` 入口还需要先通过 `pnpm run build` 生成前端与客户端构建产物；`pnpm run demo:web` 会自行执行该构建。TUI、Web 和无头入口都把调用命令时的目录用作工作区。配置、会话恢复、提供方及工作区细节见 [`dsh` CLI（命令行界面）契约](apps/cli/README.md)；[示例](examples/README.md)展示了更精简的 ACP、JSON-RPC、Code Mode 和自指组合。
推荐在本地使用 Web UI。安装完成后以及每次更新后，请先构建前端，再启动 Web UI。通过 `dsh` 启动器解析当前运行的检出，这样无论当前是哪个 staging worktree，命令都成立（启动器会经由稳定的 `current` 符号链接解析）：
## 当前提供的能力
```sh
dsh_bin=$(cd "$(dirname "$(command -v dsh)")" && pwd -P)/$(basename "$(command -v dsh)")
while [ -L "$dsh_bin" ]; do
link=$(readlink "$dsh_bin")
case $link in /*) dsh_bin=$link ;; *) dsh_bin=$(cd "$(dirname "$dsh_bin")" && cd "$(dirname "$link")" && pwd -P)/$(basename "$link") ;; esac
done
dsh_dir=$(cd "$(dirname "$dsh_bin")/.." && pwd -P)
pnpm --dir "$dsh_dir" run build && pnpm --dir "$dsh_dir" run build:web
dsh web
```
能力由组合决定。本仓库交付的插件涵盖：
Web UI 默认通过 `http://127.0.0.1:3080` 提供服务。
- **编程：** 文件系统读写、编辑与搜索，shell 和持久 PTY 执行，LSP 导航，Web 搜索与抓取，可复用 skill（技能），以及由模型编写的 Code Mode 程序。
- **编排：** subagent、后台任务、工作线程工作流、同一会话内的目标、计划状态、待办事项，以及向用户提问。
- **运维：** 工作区沙箱与审批、会话持久化／恢复／fork／查询、压缩（compaction）与 spill、投影、标题，以及 OpenTelemetry 导出。
### TUI
凡是模型可见的内容，都必须能从会话日志中重建。这样一来，不同 UI、持久化后端、回放和运维工具都成为同一事件流的消费方，而不是彼此并行的真源。
启动全屏终端界面：
## 扩展 harness
```sh
dsh
```
一项可替换能力通常会将接口、实现和消费方彼此分离。你可以为 `ctx.llm`、`ctx.fs`、`ctx.pty`、`ctx.web` 或 `ctx.subagents` 等服务添加或替换提供方；通过 `ctx.tools` 注册面向模型的行为；通过类型化事件挂接策略和请求整形；再在 `cordis.yml` 中组合这些部分，无需 fork agent loop（智能体循环）。
### Headless
运行一项任务，打印最终答案后退出：
```sh
dsh -p "summarize this workspace"
```
## 为什么选择 DeepSeek Harness
内置功能涵盖文件读取、编辑与搜索、shell 执行、可复用 skill（技能）、任务跟踪、subagent 与工作流、持久化会话，以及上下文压缩（context compaction）。TUI 还包含 Plan Mode。
- **一切皆插件。** 模型、工具、策略、存储、上下文管理和界面均可组合为 [Cordis 插件](docs/user/develop/basic/index.md)，部署方无需 fork agent loop（智能体循环）即可扩展或替换行为。底层设计见[架构文档](docs/architecture.md)。
- **Code Mode（需显式启用）。** 它会提供 `run_code` 工具和生成的 TypeScript SDK，只有程序输出会重新进入模型上下文。参见 [Code Mode](packages/core/tools/README.md#code-mode)。
- **自指 Cordis 工具需显式启用。** 这些工具可让 agent 检查自身的实时运行时，并在运行中挂载或卸载插件。参见 [Cordis 工具](packages/cordis/tool-cordis/README.md)。
## 社区
扫描二维码，或打开 <a href="https://wj.qq.com/s2/27234598/03eb/">DeepSeek Harness 微信社区申请页面</a> 申请加入。
<p>
<img src="assets/community-wecom-survey.png" alt="DeepSeek Harness 微信社区二维码" width="240">
</p>
从[第一个插件指南](docs/user/develop/basic/index.md)和[扩展实操手册](docs/cookbook/extension-cookbook.md)开始。需要系统图时查看[架构](docs/architecture.md)，需要当前服务关系时查看生成的[能力图](docs/capability-seams.md)，需要所有权细节时查看[包图](packages/README.md)。
## 开发
```sh
pnpm install
pnpm run test:coverage
pnpm run demo:tui
```
请先阅读[开发指南](docs/development.md)；修改包之前，请阅读[架构文档](docs/architecture.md)。
将 `DEEPSEEK_API_KEY` 设置在环境变量或根目录 `.env` 中。环境搭建和验证由[开发指南](docs/development.md)统一说明；修改 `packages/` 前请阅读[架构](docs/architecture.md)，在本仓库工作时请遵循 [AGENTS.md](AGENTS.md)。
面向 agent：遵循 [AGENTS.md](AGENTS.md)。
## 社区
DeepSeek Harness 目前处于预发布阶段。
前往 <a href="https://wj.qq.com/s2/27234598/03eb/">DeepSeek Harness 微信社区</a>关注项目动态。
## 许可证
scripts/snapshots/translation-prompt-v4/request-response.expected.json
+2 -2
View File
File diff suppressed because one or more lines are too long