DeepSeek Harness vs Claude Code：你在选框架，还是在选成品

先给结论：DeepSeek Harness 和 Claude Code 根本不是同一个物种，拿它们对比就像拿「发动机」对比「整车」。

DeepSeek Harness（dsh）是 Agent 运行框架——它负责调度文件读写、bash 命令、工具调用、会话管理，但不自带代码推理能力。你挂 DeepSeek V4 Flash，它就是快但糙的助手；挂 Claude Sonnet，它就是接近成品的能力。框架只是外壳，软件能力的天花板由你接入的模型决定。

Claude Code 是完整成品——Agent 循环、模型推理、项目语义、权限系统、子代理分工，全部出厂调好。你不需要理解它的架构，打开就能干活，软件能力 = 框架 + Claude 模型的组合拳。

一个是给你零件的，一个是给你整机的。搞清楚这个区别，下面的对比才有意义。

一、本质区别：Model + Harness = 智能体

dsh 的公式是 Model + Harness = 完整智能体。模型是大脑，harness 是手脚：读写文件、执行命令、调用工具、跑自动化流程。它的架构基于 Cordis 插件系统，所有能力都是插件——文件编辑是插件、命令执行是插件、模型适配是插件，加载哪些由你定。这就是 README 里那句「Everything is a Plugin」。

Claude Code 的公式是一套出厂调好的系统。2026 年的它已经远不止「终端里的 AI 聊天」：内置 42 种工具（文件、Shell、Web、LSP、子代理、工作树），权限系统是 9400 行 TypeScript 的七层模式，上下文管理有自动压缩，子代理开箱即用。它把「探索代码库 → 制定计划 → 改文件 → 跑测试 → 修复」的循环整个打包了。

一句话：dsh 是给你一套乐高，Claude Code 是给你一台组装机。组装机的性能上限未必比乐高高——但乐高需要你自己会拼。

二、十一个维度逐个拆

1. 从零新建项目

dsh：可完成，质量完全取决于后端模型。挂 Flash 适合小脚本和原型；挂 Pro 能撑中小型全栈；但长任务容易跑偏，需要人盯着纠正。RC 版本稳定性一般，适合你愿意陪跑的场合。

Claude Code：成熟稳定。根据需求生成目录结构、依赖、配置、初始业务代码，Python/TypeScript/Go/Rust 表现都不错。中小型全栈 MVP 是它的舒适区，项目规划能力强——这是模型推理 + 框架协调的合力，不是单靠框架能实现的。

2. 跨文件重构、老项目迭代

dsh：框架层面缺少原生项目图谱。模型需要自己扫描文件理解结构，长会话容易遗忘上下文，复杂重构容易漏改关联文件——你得手动约束读取范围，告诉它先看哪个目录、别碰哪个文件。

Claude Code：强项。内置项目语义图谱，三级上下文自动压缩，能理解模块依赖关系，改一处同步更新多处引用。老项目迭代、大规模迁移是它的核心优势场景。这种「改一个接口，所有调用点跟着变」的能力，靠纯 prompt 约束很难稳定复现。

3. Bug 定位、调试排错

dsh：能读报错日志、能跑命令，但复杂跨模块 bug 的稳定性受后端模型制约。长任务容易状态漂移，需要人工干预回滚任务状态。

Claude Code：第一梯队。读报错、堆栈、日志，自动复现，定位根因，修复之后主动跑测试验证。SWE-bench Verified 榜单上，Claude Sonnet 5 得 85.2%，Opus 5 得 96%（2026-08-19 数据）——注意这是模型分数，但 Claude Code 是这些模型的官方载体，疑难 bug 排查确实是它的看家本领。

4. 单元测试、集成测试、文档生成

dsh：工具支持到位，效果看后端模型。Flash 适合简单测试，Pro 能写复杂测试。

Claude Code：原生擅长。批量补单测、e2e 测试、README、接口文档一气呵成；支持 Git 提交、生成 PR 描述。测试补全 + 文档产出是它日常最省心的场景之一。

5. 终端环境、编译、构建、Docker

dsh：bash 插件完全可用，能装依赖、编译、打包、操作 Docker。权限策略完全自定义——这意味着所有安全规则都要你自己配，默认放行还是默认拦截，全看你的配置。

Claude Code：开箱即用。内置权限弹窗，高危操作逐次确认；直接执行构建、迁移、容器操作，安全策略出厂固化（七种权限模式：默认逐次确认、只读 plan、自动编辑、跳过确认等），不用自己写插件逻辑。

6. 子 Agent 分工（多 AI 协同开发）

dsh：架构原生支持子 Agent，可以配置多模型分工——小模型做扫描、大模型做编码。但子 Agent 的编排逻辑要自己写配置，没有现成最佳实践，学习成本在这里。

Claude Code：官方成熟的 dispatch 机制。内置 Explore（只读搜索）、Plan、general-purpose 三类子代理，支持自定义子代理（YAML frontmatter 定义角色、工具权限、模型），还能并行派发编码、测试、审查子代理。开箱即用，不需要二次开发。

7. 项目规范约束

dsh：需要自己写 prompt、自己写插件约束。没有原生的 CLAUDE.md 机制。

Claude Code：CLAUDE.md 是灵魂功能。把项目架构、禁止修改的文件、编码规范写进去，每次会话自动加载——相当于给 AI 做入职培训。2026 年还进化出了 .claude/rules/ 路径规则（只在编辑相关文件时加载对应规则）、/init 自动生成、@path 导入。老项目接手时，一份 CLAUDE.md 能让 AI 少犯一半错。

8. 长任务稳定性

dsh：RC 预览版短板。长任务会出现状态丢失、遗忘需求，需要外部持久化任务状态辅助。不适合完全无人值守跑几小时的任务——这是社区实测反馈，也是目前最痛的短板。

Claude Code：生产级。自动 compaction 压缩上下文，长会话记忆稳定，适合长时间持续迭代，可以半自主运行。它是为「挂机干活」设计的，不是为「聊一轮」设计的。

9. 多模态 UI 开发

dsh：本身无视觉能力，需要额外挂载视觉模型做插件扩展。原生不支持看图调试 UI。

Claude Code：支持 Computer Use 屏幕视觉（Mac 上可用），能看页面截图调试 UI 视觉问题；普通模式也支持解析 Figma 的 MCP 协议。做前端、调界面的时候差距明显。

10. 私有化内网开发

dsh：完全开源 MIT，可内网部署；后端可以接本地 GGUF 量化模型，数据不出内网。适合企业做 Agent 平台的二次开发——这是它最大的不可替代性。

Claude Code：原生闭源，官方通道只能走 Anthropic 云端（或 Bedrock/Vertex 企业通道）。社区有魔改版本，但多模态、子代理等特性大量失效，稳定性差，只适合折腾。

11. 成本控制

dsh：可以混合路由——简单任务用低价 Flash，复杂任务上 Pro；也可以本地量化模型。API 成本极低，框架本身免费。这是它性价比最强的维度。

Claude Code：全部消耗 Anthropic token，单价高，任务量大时开销明显。贵，但买的是稳定性。

三、对比总表

四、两个流传很广的误解，今天拆掉

误解一：「Claude Code 只能调用 Anthropic 云端 API」

不对。官方支持两条企业通道：Amazon Bedrock、Google Vertex AI。更关键的是——DeepSeek 官方提供了 Anthropic 兼容端点（https://api.deepseek.com/anthropic），环境变量 ANTHROPIC_BASE_URL 一行切换就能让 Claude Code 跑 DeepSeek 模型。这是官方文档支持的用法，不是魔改。

误解二：「把 Claude Code 后端切 DeepSeek 是 hack，会损坏功能」

不用 hack，但确实有个真坑：Claude Code 2.1.166+ 的子代理功能会和 DeepSeek 兼容层冲突（GitHub issue #1397——子代理发送 thinking:disabled 与 DeepSeek 推荐的 reasoning_effort 参数互斥，API 返回 400）。2.1.165 及之前版本一切正常。结论：官方兼容端点 + 锁版本，功能基本可用；追新版本会踩 400 的坑。这是版本 bug，不是 hack 的锅。

五、选型建议

普通业务开发者，以写软件、出产品为主：优先 Claude Code。开箱即用，软件开发综合能力第一梯队，稳定性是拿时间换不来的。你不需要理解 Agent 循环，它已经把十年工程化经验打包好了。

Agent 平台研发、私有化部署、想深度改造 Agent 逻辑：选 DeepSeek Harness。MIT 开源、内网部署、本地模型、插件可定制——这些是闭源产品给不了的自由。同时后端挂强推理模型（Claude Sonnet 或 DeepSeek V4 Pro），把外壳的调度能力和模型的推理能力都吃满。

两头都想要：DSH 框架 + Claude API，或者 Claude Code + DeepSeek 兼容端点。前者拿可控可审计，后者省钱省心——注意锁 Claude Code 版本 2.1.165。

最后记住一件事：不要把 Harness 本身当成会写代码的 AI。它是执行外壳，软件能力上限由你接入的大模型决定。选框架还是选成品，本质是问自己：你愿意花时间调教工具，还是愿意花钱买时间。

版权声明：本文首发于 cn-res.vip，作者 Grout。转载需保留本声明与全文。

Let's get the premise straight first: DeepSeek Harness and Claude Code are not the same species. Comparing them is like comparing an engine to a car.

DeepSeek Harness (dsh) is an agent runtime framework. It orchestrates file I/O, bash commands, tool calls, and session management — but it carries no code reasoning ability of its own. Plug in DeepSeek V4 Flash and you get a fast but rough assistant; plug in Claude Sonnet and you approach out-of-the-box capability. The framework is just the shell. The ceiling is set by the model you attach.

Claude Code is a finished product — agent loop, model reasoning, project semantics, permission system, subagent orchestration, all tuned at the factory. You don't need to understand its architecture to use it. Software capability = framework + Claude model, as one package.

One gives you parts. The other gives you the whole machine. Get that distinction right and the rest of the comparison falls into place.

Part One: The Essential Difference

dsh's formula is Model + Harness = complete agent. The model is the brain; the harness is the hands — reading files, running commands, calling tools, executing automation. Its architecture is built on the Cordis plugin system: everything is a plugin. File editing, command execution, model adaptation — you choose what to load. That's what "Everything is a Plugin" actually means.

Claude Code's formula is a factory-tuned system. In 2026 it's far beyond "AI chat in the terminal": 42 built-in tools (files, shell, web, LSP, subagents, worktrees), a permission system of roughly 9,400 lines of TypeScript with seven modes, automatic context compaction, and out-of-the-box subagents. The whole "explore repo → plan → edit → test → fix" loop is packaged.

One sentence: dsh gives you Lego; Claude Code gives you an assembled machine. The assembled machine isn't necessarily more capable — but it doesn't require you to know how to build.

Part Two: Eleven Dimensions, One by One

Greenfield projects. dsh: workable, quality depends entirely on the backend model. Flash for small scripts, Pro for small-to-mid full-stack — but long tasks drift, so you have to watch it. RC stability is middling. Claude Code: mature. Generates directory structure, dependencies, config, initial business code; strong in Python/TypeScript/Go/Rust. Small-to-mid full-stack MVP is its comfort zone.

Cross-file refactoring and legacy projects. dsh: no native project graph. The model scans files itself; long sessions forget context; complex refactors miss related files — you manually constrain its read scope. Claude Code: a core strength. Built-in project semantic graph, three-tier context compaction, understands module dependencies, updates all references when one changes. That "change one interface, all call sites follow" behavior is hard to reproduce with prompt discipline alone.

Bug hunting and debugging. dsh: can read logs and run commands, but cross-module debugging stability is model-bound; long tasks drift. Claude Code: first tier. Reads errors, stacks, logs; reproduces automatically, finds root cause, then runs tests to verify the fix. On the SWE-bench Verified leaderboard (as of August 19, 2026): Claude Sonnet 5 at 85.2%, Opus 5 at 96%. Those are model scores, but Claude Code is the official vehicle for those models — and hard-bug hunting is its signature skill.

Tests and documentation. dsh: tooling is there, output follows the model. Claude Code: natively good — batch unit tests, e2e, READMEs, API docs, plus Git commits and PR descriptions.

Terminal, builds, Docker. dsh: bash plugin works; permissions are fully custom — meaning you write every safety rule yourself. Claude Code: out of the box. Permission prompts for risky operations, factory-default security policy, no plugin logic required.

Subagent orchestration. dsh: natively supports subagents and multi-model division of labor, but you write the orchestration config yourself; no established best practices yet. Claude Code: mature dispatch. Explore (read-only), Plan, general-purpose built in, plus custom subagents defined in YAML frontmatter, parallel encoding/testing/review agents. Zero second development.

Project conventions. dsh: prompt discipline and custom plugins; no native CLAUDE.md. Claude Code: CLAUDE.md is the soul feature — architecture, protected files, coding standards, auto-loaded every session. It's onboarding for the AI. By 2026 it has path-scoped rules (.claude/rules/), /init auto-generation, and @path imports. On a legacy codebase, a good CLAUDE.md cuts the AI's mistakes in half.

Long-task stability. dsh: RC preview weakness. State loss, forgotten requirements; needs external task-state persistence. Not suited for unattended multi-hour runs — that's community-verified and the sharpest pain point today. Claude Code: production-grade. Automatic compaction, stable long-session memory, semi-autonomous iteration. It's designed to be left running.

Multimodal UI work. dsh: no native vision; needs a vision-model plugin. Claude Code: Computer Use screen vision on Mac, screenshot-based UI debugging, plus Figma MCP parsing. The gap is obvious on frontend work.

Private/offline deployment. dsh: fully open source (MIT), deployable on intranet, GGUF local models — data never leaves the building. This is its irreplaceable advantage. Claude Code: closed source; official paths are Anthropic cloud or Bedrock/Vertex enterprise channels. Community hacks lose subagents, multimodality, and stability.

Cost. dsh: mixed routing — cheap Flash for simple tasks, Pro for hard ones; local quantized models optional; the framework itself is free. Claude Code: all Anthropic tokens, higher unit price, noticeable spend at scale.

Part Three: The Comparison Table

Part Four: Two Widely Repeated Myths, Demolished

Myth one: "Claude Code can only call the Anthropic cloud API." Wrong. Official enterprise channels include Amazon Bedrock and Google Vertex AI. More importantly — DeepSeek ships an official Anthropic-compatible endpoint (https://api.deepseek.com/anthropic). One environment variable (ANTHROPIC_BASE_URL) switches Claude Code onto DeepSeek models. Documented, supported, not a hack.

Myth two: "Switching Claude Code to DeepSeek is a hack that breaks features." No hack required — but there is a real trap: Claude Code 2.1.166+ subagents conflict with the DeepSeek compatibility layer (GitHub issue #1397 — subagents send thinking:disabled, which the DeepSeek endpoint rejects when combined with the officially recommended reasoning_effort; HTTP 400). Version 2.1.165 and earlier work fine. Bottom line: official endpoint + pinned version ≈ fully usable; chasing the latest version hits the 400. It's a version bug, not a hack's fault.

Part Five: How to Choose

Business developer shipping software → Claude Code. Out-of-the-box, first-tier software engineering capability, stability you can't buy with time. You don't need to understand agent loops; ten years of engineering practice is already packaged.

Agent platform developer, private deployment, deep customization → DeepSeek Harness. MIT, intranet, local models, pluggable — freedoms closed source can't give. Pair it with a strong reasoning model (Claude Sonnet or DeepSeek V4 Pro) and you get both the shell's orchestration and the model's reasoning.

Want both → dsh + Claude API, or Claude Code + DeepSeek endpoint. First gives auditability and control; second saves money — just pin Claude Code to 2.1.165.

One last thing: never mistake the harness for an AI that writes code. It's an execution shell. The ceiling is your model. Choosing between framework and product is really asking yourself: do you want to spend time tuning tools, or spend money to buy time?

© 2026 cn-res.vip. Author: Grout. Reproduction requires this notice in full.