面向 DeepSeek Harness 的分级模型路由:think/build 模型分级、视觉委派(视觉模型只负责描述图片,工作模型负责执行操作)、图片生成,以及高影响操作守卫。
安装
# GitHub 源码(首次需按提示配置 allowBuilds 构建授权后重试)
dsh plugin --profile web add github:llmpolska/oh-my-dsh
装任何插件都等于在你的机器上跑第三方代码,权限和你本人一样大——能读你的文件、用你的凭据、访问网络,工具审批管不到它。GitHub 来源的插件还会在安装时执行构建脚本——pnpm 默认拦截,所以安装可能停在 ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED 或 ERR_PNPM_IGNORED_BUILDS;dsh 会打印出需要添加的确切键名,把它加进该 profile 的 pnpm-workspace.yaml 的 allowBuilds 下,重跑一次即可装上。放行构建本身就是一次信任判断:请只安装可信来源,并尽量锁定 commit(github:owner/repo#sha)。
README
该插件的 README 只有英文版本。
Tiered model routing plugin for DeepSeek Harness — split every session into think and build model tiers, delegate image understanding to a vision model (it only describes — your coding model does the work), and generate images from an ordinary chat model. Configured from a settings tab, no YAML editing.
Works on every DSH surface: DSH Desktop, the standalone web profile (dsh web),
and the terminal (dsh --profile tui).
Why oh-my-dsh?
Running every step of a coding session on a frontier model is expensive and slow. oh-my-dsh gives the agent tiers: a strong (think) model for planning, architecture and hard debugging, a cheaper (build) model for day-to-day implementation, an automatic vision delegation layer for images (the vision model describes, the working model acts), and an image-generation path — all without switching sessions manually and without YAML configuration.
Features
- Think / Build tiers — plan mode runs on the strong model, execution on the
cheap one (
automode; per-session override with/omd strong|cheap). - Vision delegation (omp-style) — paste a screenshot of a website and say "build this frontend": the vision model describes it in the background and the working tier answers from that description. The image stays in the chat, the session model never switches, and later turns never fail with unsupported content.
- Image generation —
omd_imagetool and/omd image <prompt>generate images with a configured chat model (e.g.gpt-5.6-lunaon opencode-go,gpt-image-2via Codex), saved under./oh-my-dsh-images/. - High-impact guard —
rm -rf,sudo, force pushes and credential/secret file edits are denied while the build tier executes until the session escalates to think. - Failure auto-escalation — repeated model-step errors temporarily escalate to the think tier.
- Advisor / Reviewer —
omd_advisorandomd_reviewconsult the think tier on demand, before risky decisions and high-risk merges. - Subagent tiering —
omd_workerdispatches bounded task packets to a chosen tier. - System prompt section — every session gets a prompt section describing the
tiers, the
/omdcommand mapping and the routing rules. - Settings tab — providers, models, reasoning efforts, guard and escalation tuning from the "oh my dsh" tab in DSH Desktop settings.
Installation (easy)
Requirements: DeepSeek Harness Desktop (or the dsh CLI) and an internet connection —
the dsh plugin command manages pnpm itself, nothing else to install.
Quickstart — install straight from the GitHub repo (no local download)
Install into every surface you use (each profile is independent):
# DSH Desktop app (the profile the desktop app boots — the settings tab lives here)
dsh plugin --profile desktop add github:llmpolska/oh-my-dsh#v0.1.0
# Standalone web app (dsh web / --profile web)
dsh plugin --profile web add github:llmpolska/oh-my-dsh#v0.1.0
# Terminal (TUI)
dsh plugin --profile tui add github:llmpolska/oh-my-dsh#v0.1.0
The #v0.1.0 suffix pins the exact release tag. The package has no build step
(no prepare script), so pnpm installs it directly — no allowBuilds configuration
is needed. The command also activates the plugin as a profile layer automatically.
Restart
Fully restart DSH Desktop (Cmd+Q → relaunch) or restart the dsh process. The
plugin's host half loads at profile boot — a window refresh is not enough.
On boot the plugin:
- registers the
oh-my-dshsettings namespace and the/omdconfiguration endpoint, - installs the
oh-my-dshagent preset into$DSH_HOME/.agent-presetsand adopts it as the default preset (only when the default is stillstandard), - installs the vision-describe wrapper on the api-proxy (the image gate is passed via a
brief vision header, the turn stays on the working tier, and a background vision
description is queued) and the history scrub on
llm.streamWithRegistration(image parts are replaced by the description for text-only models), - adopts first-boot defaults: think/build take your session default model, vision takes the first image-capable model of that provider. Everything stays editable in the tab.
New sessions get the preset (existing sessions keep theirs — create a new session to get
/omdand theomd_*tools).
Alternative: install from a local tarball (offline / development)
cd oh-my-dsh
pnpm pack # produces oh-my-dsh-0.1.0.tgz
dsh plugin --profile desktop add ./oh-my-dsh-0.1.0.tgz
dsh plugin --profile web add ./oh-my-dsh-0.1.0.tgz
dsh plugin --profile tui add ./oh-my-dsh-0.1.0.tgz
A local path install (dsh plugin add /path/to/oh-my-dsh) links the checkout as-is —
after changing lib/, just restart the app.
Which surface does what
| Surface | Settings tab | /omd + omd_* tools |
Vision describe | Image generation |
|---|---|---|---|---|
| DSH Desktop | ✅ | ✅ | ✅ | ✅ |
dsh web (web profile) |
✅ | ✅ | ✅ | ✅ |
| TUI / headless | — (no settings UI) | ✅ (agent plane activates process-wide) | — (no image prompts) | ✅ |
The plugin is one codebase: the client half (settings tab) is web-platform, the host half (endpoint, vision wrapper) mounts where a web server exists, and the agent plane (roles, tools, guard, scrub) activates through the preset on web surfaces or process-wide in the TUI, where no preset roster exists.
Usage
Slash commands
/omd status
/omd strong | cheap | auto | off # per-session tier (think/build aliases work too)
/omd plan # auto + plan mode + think header
/omd models # list registered providers and models
/omd set <think|build|vision> <provider> <model> [off|high|max]
/omd subagent <inherit|cheap|strong>
/omd advisor <question> # one think-tier consultation
/omd review <focus> # think-tier review
/omd image <prompt> # generate an image (saved under ./oh-my-dsh-images)
The automatic vision description (visionAuto) is toggled in the settings tab; the
describing model is set with /omd set vision <provider> <model>.
Tools
omd_status, omd_route, omd_configure, omd_advisor, omd_review, omd_worker,
omd_image.
How the vision delegation works
- You send an image → the message appears immediately in the chat, image intact.
- The vision model describes the image in the background (one vision call; the description is cached keyed by the image's sha256).
- The turn runs on the working tier (think/build): its first request waits for the description, then answers from it — the vision model never runs a turn.
- Later turns: the image part in history is replaced by the description for text-only models, so nothing ever fails with unsupported content.
The settings tab
Open DSH Desktop → ⚙ Settings → oh my dsh:
- Routing: mode (auto/strong/cheap/off), subagent policy, escalation tuning, guard toggle.
- Think tier: provider + model + reasoning effort.
- Build tier: provider + model + reasoning effort.
- Vision tier: picker limited to image-capable models, auto-describe toggle.
- Image generation: provider picker (registered providers — the plugin derives the
base URL and reuses the provider's API key) + model id (e.g.
gpt-5.6-luna), size, output directory, and the "image output modalities" toggle. The "custom endpoint" option is the manual fallback (base URL + API key env var/inline key).
For developers
lib/index.js— host half:/omdendpoint, vision describe wrapper, preset install.lib/agent.js— agent plane: tier routing, guard, escalation,omd_*tools, prompt section,llm.streamWithRegistrationhistory scrub.lib/client.js— the "oh my dsh" settings tab (plain JS, served by the shell).lib/config.js— settings schema and shared state (incl. the vision description store).
Tests
node --test "tests/**/*.test.mjs" # 25 unit tests for the guard and tier decisions
node --check lib/index.js && node --check lib/agent.js && node --check lib/client.js
Uninstall
dsh plugin --profile desktop remove oh-my-dsh
dsh plugin --profile web remove oh-my-dsh
dsh plugin --profile tui remove oh-my-dsh
# optionally: delete $DSH_HOME/.agent-presets/oh-my-dsh and the `oh-my-dsh` section
# in $DSH_HOME/settings.yaml, then set the agent-presets default back to standard.
Keywords
deepseek-harness · dsh-plugin · dsh · llm · model-routing · tiered-routing · coding-agent · vision-model · image-generation · open-source · plugin
License
MIT
Built and maintained by LLM Polska — plugins, tools and automations for DeepSeek Harness. https://llmpolska.pl
链接
同类插件
Ebony-Vinyl/dsh-our-free-model★ 77498
为 dsh 接入免费模型车道(含 DeepSeek V4.1 Flash、Kimi K3 等),无需登录、注册或填写 API Key,并提供本地 OpenAI 兼容网关、降级路由与用量看板。
dingminhua/dsh-connect-workbuddy★ 1695
将本机登录的 WorkBuddy 模型(国内版与国际版)接入 DSH,每次刷新按上游目录读取每个模型的积分倍率,支持逐模型图片输入开关、带自动换号的账号池、手动每日签到,以及卡片与输入框里的只读积分胶囊读数。
Alice-Marx/dsh-model-router★ 970
为 DeepSeek Harness 规划有依赖的工作包,将其分配给已配置的模型 API 或支持的厂商 CLI,并提供费用预览。
V1ki/dsh-plugin-subscriptions★ 619
把 ChatGPT(Codex)、Claude、Grok 订阅当作 DeepSeek Harness 的 LLM 提供方:设置页登录、模型目录、用量展示,以及 image_generate、video_generate 与 x_search 工具。
XDTrees/dsh-workbuddy-xdpool★ 547
把本机 WorkBuddy 桌面 App 登录过的所有账号自动并入 DSH 模型池:多账号轮换自动容错、实时积分,模型目录按模型标注积分倍率与限时免费 / 夜间折扣。
XDTrees/dsh-xdpool★ 547
把本机 WorkBuddy 桌面 App 登录过的所有账号自动并入 DSH 模型池:多账号轮换自动容错、429 冷却、实时积分,模型目录按模型标注积分倍率与限时免费 / 夜间折扣,扫码登录 13 家厂商,聚合免费额度渠道。账号重启不丢。
社区评论
评论公开保存在 GitHub Discussions。加载评论会连接 GitHub 和 Giscus;发表内容需要 GitHub 账号。