DeepSeek Harness 的原生 Rapid-MLX 提供者：dsh 自动发现已部署模型并在会话内管理，上下文压缩贴合 Mac 统一内存上限。
DSH 适配
原生运行时
作者声明
安全审计
未审计
最后核验
2026-08-30
许可证
Apache-2.0
01它能帮你完成什么？
Run local Rapid-MLX models inside DeepSeek Harness without hand-maintaining per-model facts
`dsh` auto-discovers served models from `/v1/models` — context window, reasoning/tool parsers, MoE/hybrid, modalities — with no hand-written settings.yaml
DeepSeek Harness users on Apple Silicon running Rapid-MLX who want accurate, server-driven model facts
See, pull, remove and health-check Rapid-MLX models without leaving the dsh session
Five agent tools — `rapid_mlx_serving`, `rapid_mlx_cached`, `rapid_mlx_pull`, `rapid_mlx_remove`, `rapid_mlx_health` — plus a `/rapid-mlx` command
Agents and developers who want to manage served/downloadable models and check health in-session
Get truthful reasoning controls and machine-fitted context compaction
Reasoning selector only appears for models that actually have a reasoning parser; compaction timed to the server's `max_model_len` (unified-memory ceiling), not a hand-written number
Developers hit by silent failures from stale context windows or dead reasoning selectors when switching Rapid-MLX models
02如何接入 DeepSeek Harness？
前置条件
- Node ≥ 22.15 (dsh imports Node's Zstd stream API without declaring it)
- a running Rapid-MLX server
安装步骤
- 01
$ dsh plugin --profile web add @raullenchai/dsh-provider
- 02
$ dsh plugin --profile web add github:raullenchai/rapid-mlx-dsh-provider
- 03
export RAPID_MLX_BASE_URL=http://localhost:8000/v1 # optional; this is the default
- 04
$ dsh web
- 05
# $DSH_HOME/settings.yaml agent-default-model: provider: rapid-mlx model: qwen3.6-35b-8bit
验证接入成功
- Installs and activates as a profile layer; the entry shows up in `dsh --profile headless --dump-config` with no "declares no dsh.bundle" warning.
- Registers the `rapid-mlx` route with `ctx.llm` and serves real queries.
03DSH 适配与能力边界
Native dsh LLM adapter: installed via `dsh plugin --profile web add`, registers the `rapid-mlx` route and reads model facts from the Rapid-MLX server's `/v1/models`.
Server-driven model discovery
Rapid-MLX `/v1/models` HTTP endpoint→served models with context window, reasoning/tool parsers, MoE/hybrid, modalities — deduped
In-session model management tools
the active dsh agent session→five tools (`rapid_mlx_serving`, `rapid_mlx_cached`, `rapid_mlx_pull`, `rapid_mlx_remove`, `rapid_mlx_health`) plus a `/rapid-mlx` command
`rapid_mlx_pull` and `rapid_mlx_remove` change on-disk cached models (non-interactive, forced `-y`)Memory-fitted context compaction
dsh-compaction-basic compaction request→compacts at `thresholdRatio × capacity` (0.8 default) using server `max_model_len` when available, else `context_window`
Conformant LLM adapter (cookbook contract)
dsh LLM calls — chat, tool calls, streaming, abort→streaming responses meeting the official adapter protocol obligations
04适合谁？何时不该用？
适合
- DeepSeek Harness users on Apple Silicon running Rapid-MLX who want accurate, server-driven model facts
- Agents and developers who want to manage served/downloadable models and check health in-session
- Developers hit by silent failures from stale context windows or dead reasoning selectors when switching Rapid-MLX models
不适合
- The route is registered as `rapid-mlx`. If your settings.yaml also declares a `rapid-mlx` provider under `llm-pi-ai`, the two compete for one route name (`registerAdapter` owns provider exclusivity). Use one or rename ours.
- Several server-reported facts are read but not yet acted on — `recommended_sampling`, `tool_call_parser` (no fast-fail on models that can't emit tool_calls), and `is_hybrid`/`is_moe`/`capabilities`; images are refused with `UNSUPPORTED` rather than carried.
05兼容性、维护与安全提示
- The route is registered as `rapid-mlx`. If your settings.yaml also declares a `rapid-mlx` provider under `llm-pi-ai`, the two compete for one route name (`registerAdapter` owns provider exclusivity). Use one or rename ours.
- Several server-reported facts are read but not yet acted on — `recommended_sampling`, `tool_call_parser` (no fast-fail on models that can't emit tool_calls), and `is_hybrid`/`is_moe`/`capabilities`; images are refused with `UNSUPPORTED` rather than carried.
- DSH is still a developer preview that moves fast; dsh 0.1.0-rc.8 is API-compatible with rc.7 but the author treats compatibility as tracking a moving target, not a frozen promise.
Apache-2.0 · published to npm as @raullenchai/dsh-provider (latest release v0.2.0, 2026-08-19)
06常见问题
如何为 DeepSeek Harness 安装 dsh-provider？
运行 `dsh plugin --profile web add @raullenchai/dsh-provider`（或 `dsh plugin --profile web add github:raullenchai/rapid-mlx-dsh-provider` 装源码）。需要 Node ≥ 22.15 和一个正在运行的 Rapid-MLX 服务端，之后执行 `dsh web`。
安装前提是什么？
Node ≥ 22.15（dsh 会直接 import Node 的 Zstd 流 API 而未声明）以及一个正在运行的 Rapid-MLX 服务端。仅当服务端不在默认地址 http://localhost:8000/v1 时才需要设置 `RAPID_MLX_BASE_URL`。
它如何接入 DeepSeek Harness？
它以原生 LLM 适配器身份注册为 `rapid-mlx` 路由；dsh 连接本地 Rapid-MLX 服务端的 OpenAI 兼容 `/v1` 端点，并从 `/v1/models` 读取模型信息。在 settings.yaml 中将 `agent-default-model` 指向 `provider: rapid-mlx` 即可。
与 dsh 自带的 openai-completions 路由有何不同？
通用路由需要你手写每模型的 facts 到 settings.yaml，且只知道你填的内容。本适配器读取 `/v1/models`，切换模型无需重新配置，且推理档位只在实际具备推理解析器的模型上出现。
排障：我的 settings.yaml 里已有 rapid-mlx 提供者怎么办？
两者都会注册 `rapid-mlx` 路由并争抢同一路由名（`registerAdapter` 拥有提供者排他性）。只用其中一个，或重命名本插件的路由。注意 dsh 0.1.0-rc.8 与 rc.7 API 兼容，但 DSH 仍是快速迭代的开发者预览版。
07相关的 DSH 工作流
dsh-routing-suite
作者 yjh051108
DeepSeek Harness（dsh）路由套件：先装运行时注入器，再加载任务感知的推理模式路由预设（已实测 P1-P23）。
brooks-lint
作者 hyhmrright
基于 12 本经典工程书籍的 AI 代码评审工具，提供腐化风险诊断、书籍引用、严重度标签及 6 种分析模式。
dsh-plugin-shop
作者 livxue
DeepSeek Harness 最全面的插件市场，每日更新、全网采集、发布前审核。
dsh-our-free-model
作者 zouyuxuan122
在 dsh 里装上这个插件即可，无需登录、注册或填 API Key，就能使用包括 Muse Spark 1.3、MiMo V2.6 在内的前沿模型——完全免费，不限量。 All you do is install this plugin in dsh: no login, no sign-up, no API key — the frontier models are just there, Muse Spark 1.3 and MiMo V2.6 among them. Completely free, with no usage cap.
08数据与来源
A native [Rapid-MLX](https://github.com/raullenchai/Rapid-MLX) provider for
Registers the `rapid-mlx` route with `ctx.llm` and serves real queries.
页面基于项目公开文档、仓库元数据和 DSH Plugins 的结构化解析生成；最后核验于 2026-08-30。发现错误？提交更正。
最佳 DeepSeek Harness 插件
从全目录挑出的 12 个值得优先安装的插件，覆盖各个分类。