dsh-commandcode-provider

English | 简体中文

A lightweight plugin — almost no extra dependencies. It is plain JavaScript (no build step, no framework), ships only two small files plus a one-file patch, and its sole runtime dependency is @deepseek-ai/schemastery, which any dsh plugin already has. Everything else it uses comes from the dsh host itself (ctx.web, the settings seam, the credentials seam, the loopback bridge).

One-click sync of CommandCode subscription-tier models into DeepSeek Harness' llm-pi-ai provider configuration, with a "Create / Update" button in the settings page. Existing target providers are only refreshed (model list updated) — your configured keys and base URLs are preserved. On top of that, it can replace dsh's built-in web search (no separate search API key needed) and show your CommandCode account usage right in the settings card.

Why this plugin

• dsh's llm-pi-ai provider catalog is a static snapshot that never refreshes itself. CommandCode's Provider API currently lists 61 models and grows continuously (each with its own context/pricing/capabilities). Hand-copying them into the provider configuration is impractical and goes stale.
• The official /provider/v1/models endpoint returns only id / name / context_length — no reasoning or vision capability info. This plugin additionally parses the complete per-model catalog embedded in the official GOAT plan page (reasoning / vision / caps / four pricing fields / min plan) and maps capabilities correctly into dsh config.
• Subscription tiers: CommandCode has multiple tiers (Go / GOAT / Pro / Max); models accrue by minPlanName. The plugin offers a "Subscription" dropdown; each tier maps to its own independent model provider, so tiers never overwrite each other.
• Mixed routing: CommandCode serves both an OpenAI-compatible (/chat/completions) and an Anthropic-compatible (/messages) endpoint; sending Claude models to the wrong endpoint returns 400. The plugin splits models automatically: Claude models go into commandcode-<plan>-anthropic (api: anthropic-messages), everything else into commandcode-<plan>-autosync (api: openai-completions).
• Web search without a second API key: stock dsh needs a separate DEEPSEEK_API_KEY before the model's web_search tool works at all. This plugin serves the same tool from your CommandCode account key, so one subscription covers chat and search — no dsh search key to configure or pay for.
• Account usage at a glance: the settings card shows live CommandCode usage (five-hour/weekly window limits with progress bars, monthly credits, request/cost/token totals), so you can see a rate-limit or out-of-credits wall coming before you hit it.

Tiers and providers

Subscription (dropdown)	Total models	Includes	Generated providers
goat (default)	43	Go + GOAT open models (no Claude)	commandcode-goat-autosync
pro	56	everything in goat + Pro tier (Claude Sonnet/Haiku, GPT, Gemini, …)	commandcode-pro-autosync + commandcode-pro-anthropic
max	61	all models (incl. Claude Opus/Fable, Fugu Ultra)	commandcode-max-autosync + commandcode-max-anthropic

Tiers are cumulative (defined by the minPlanName field, matching the official docs: the Pro page states "includes everything in the GOAT plan"). The Go tier offers no API access (403) so it is not an option.

Mapping rules (upstream → llm-pi-ai YAML)

Upstream (official catalog)	llm-pi-ai model field
contextWindow / context_length	contextWindow
vision: true	input: ["text", "image"] (else ["text"])
reasoning: false	reasoningEfforts: false (disable thinking parameters)
reasoning: true	reasoningEfforts: {low: low, medium: medium, high: high, xhigh: xhigh, max: max}
minPlanName	filters models by the selected tier
vendor: Anthropic (or id starting with claude-)	routed into the api: anthropic-messages provider
everything else	routed into the api: openai-completions provider

The catalog marks only reasoning: true/false, no per-model levels, so the map is the identity over the levels the endpoint itself accepts: a reasoning_effort outside low|medium|high|xhigh|max is refused with HTTP 400 naming that set. The map is what tells dsh the model reasons at all — without it dsh treats the model as non-reasoning (no levels in the picker, no reasoning parameter on the wire), while the provider-level compat: {thinkingFormat: "openai", supportsReasoningEffort: true} only decides how the chosen level is sent (overridable via targetCompat).

Web search (optional)

When enabled, the plugin's Command Code search provider backs dsh's model-facing web_search tool via the Command Code Provider API's /alpha/web-search endpoint — the same API key and account as your chat traffic, so no separate search key or endpoint is configured.

Why this is useful: stock dsh ships web search backed by DeepSeek's own Messages API, which means you need a separate DEEPSEEK_API_KEY (billed on top of whatever you pay for chat) before the model can search at all. With a CommandCode subscription you already have an account key that unlocks search — this plugin reuses that same key for search, so you no longer depend on dsh's own search API key. One subscription, one key: chat and search both covered.

• Served on the dsh web seam (ctx.web) as provider id commandcode, and auto-selected while the toggle is on (restoring the previous search provider when toggled off or when the plugin unloads).
• numResults is clamped to Command Code's range (1–10, default 5); results map to dsh's WebSearchSource shape (url/title/snippet).
• Requires the account key (COMMANDCODE_API_KEY by default — the same credential the chat providers use). Off by default; enable in the settings card (takes effect immediately after Save, no restart).

Usage dashboard (optional)

The settings card also shows account usage — requests, success rate, cost, tokens, credit balances, and the 5-hour/weekly window limits — fetched Host-side from the account endpoints (/alpha/whoami, /alpha/usage/summary, /alpha/billing/credits, /alpha/billing/subscriptions) with the same account key. The key never leaves the host.

What you get at a glance: how many requests you have left in the current five-hour and weekly windows (with progress bars and reset times), how many credits your plan still has this month, and your recent request/success/cost/token totals — so you can see whether a rate-limit or "out of credits" wall is approaching before you hit it, and confirm that a session actually consumed what you expected.

• Each endpoint degrades independently: a transient failure shows a partial-data note instead of blanking the card; when every endpoint fails the same way the card names the cause (invalid key / service unavailable / network).
• The usage endpoints live on the API root (/alpha/*), which is distinct from the chat base /provider/v1. Use the Usage / search API base field in the card if your deployment differs.

Install

The settings page rides dsh's settings API, which changed in the 0.1.7-alpha.1 line, so pick the plugin tag that matches the dsh you run:

dsh	Plugin	Install (web profile)
0.1.7-alpha.1 … <0.2.0	v0.2.2	dsh plugin --profile web add github:CJYLZS/dsh-commandcode-provider#v0.2.2
0.1.7-alpha.1 … <0.2.0	v0.2.0 (Plugins page only)	dsh plugin --profile web add github:CJYLZS/dsh-commandcode-provider#v0.2.0
0.1.5-rc.3 and older	v0.1.3	dsh plugin --profile web add github:CJYLZS/dsh-commandcode-provider#v0.1.3

#<ref> is the Git ref the install resolves. Without one, the repository's default branch is installed, which tracks the newest line.

For a local checkout, link the directory instead of copying it — a later edit applies on the next restart, with no reinstall:

dsh plugin --profile web add link:/absolute/path/to/dsh-commandcode-provider

Restart dsh Web after installing, then open Settings → CommandCode, or the Plugins page's bundle card and Configure on its row — both mount the same page. Choose the subscription (default goat) and click Create / Update.

dsh compatibility

dsh	Plugin	Settings surface
≥ 0.1.7-alpha.1	v0.2.2	A page of its own in Settings (settings.section, nav label CommandCode), beside the Plugins page's row configuration; writes model reasoning maps and runs auto-sync
≥ 0.1.7-alpha.1	v0.2.0	The Plugins page's row configuration (plugins.row.config) only, with no Settings page
≤ 0.1.5-rc.3	v0.1.3	A registered settings section (settings.register / installSection), listed with the model providers

v0.2.2 and v0.2.0 both read and write their configuration as live references of their own Cordis Config; v0.2.2 adds the standalone Settings page and writes the reasoning maps by default, which is what makes the model picker offer thinking levels at all. v0.1.3 uses the removed settings.register / installSection API and the retired settings.plugin.item slot, so it does not load on 0.1.7-alpha.1 or later.

A mismatch shows up as web boot: 1 entry did not activate … pending (waiting for service: settingsScope) in the browser, and as settings.register is not a function for this package in $DSH_HOME/logs/startup-*.log when running v0.1.3 on the newer dsh.

API key setup

After you click Create / Update, the plugin auto-creates a model provider for your tier — commandcode-goat-autosync by default (with commandcode-pro-… / commandcode-max-… for the other tiers). That provider is where the API key goes.

Recommended — via the dsh Models page:

1. In dsh, go to Settings → Models (the page lists every provider, including the one this plugin created).
2. Find commandcode-goat-autosync (or your tier's provider) and click Edit.
3. Paste your CommandCode API key into the API key field and Save.

That's it — the key is stored in dsh's credential store and is used for chat, web search, and the usage card alike.

Alternative — environment variable: export COMMANDCODE_API_KEY=cmd_xxx before starting dsh. (The plugin's config option targetApiKeyEnv renames the variable it reads.)

Where does the key come from? Create one on the commandcode.ai Studio API keys page (all tiers except Go; GOAT/Pro/Max bill against subscription credits, Provider plans are pay-as-you-go).

Configuration

Field	Default	Description
sourceURL	https://api.commandcode.ai/provider/v1/models	Official live model list
catalogURL	https://commandcode.ai/docs/plans/goat	Official GOAT plan page (capability catalog); falls back to the plain list when it fails
plan	goat	Subscription tier: goat / pro / max
targetApiKeyEnv	COMMANDCODE_API_KEY	Credential environment variable name
targetBaseURL	https://api.commandcode.ai/provider/v1	API base URL
targetCompat	{thinkingFormat: "openai", supportsReasoningEffort: true}	compat override for the openai route
extraIds	[]	Extra private model ids to write (outside the catalog, into the openai-route provider)
includeReasoningEfforts	true	Write the low…max reasoning map for catalog-reasoning models; turning it off makes dsh treat them as non-reasoning
autoSync	false	Periodic auto-sync (off by default; enable in the card)
autoSyncIntervalMs	6h	Auto-sync interval (min 60s)
webSearch	false	Serve dsh's web_search with Command Code (/alpha/web-search)
usageBaseURL	https://api.commandcode.ai	API root for usage/search endpoints (/alpha/*)

Target provider names are derived from plan (commandcode-<plan>-autosync / commandcode-<plan>-anthropic); no separate configuration needed.

FAQ

What happens when I switch tiers? Pick the new tier in the dropdown and Save, then click Create / Update: the new tier's provider is created; the old tier's provider is left untouched (tiers never overwrite each other). Remove a stale provider manually.

Why is there no commandcode-goat-anthropic? The GOAT tier contains only open models — no Claude — so no Anthropic-route provider is generated.

How do Claude models work? The official API requires Claude on /messages (Anthropic format). The pro / max tiers automatically create commandcode-<plan>-anthropic (api: anthropic-messages); calling a Claude id through an OpenAI-route provider returns 400.

Why do some models have no reasoning levels? Only the ones the catalog marks reasoning: false: they get reasoningEfforts: false, so dsh sends no thinking parameter and the picker offers no levels. Every reasoning: true model gets the low…max map — that map is what puts the levels in the picker at all. The model list and these maps are rewritten on sync, so after changing an option click Create / Update (or let auto-sync run) before the provider config carries it.

What if I call a model above my tier? The catalog filters strictly by minPlanName, so written models are all inside the selected tier; when upstream adds a model the catalog has not yet indexed, the plugin writes the full plain list (degraded mode, with a warning in the result).

---

简体中文

English | 简体中文

轻量插件——几乎没有额外依赖。 纯 JavaScript 实现（无构建步骤、无框架），只附带两个小文件加一个单文件 patch；唯一的运行时依赖是 @deepseek-ai/schemastery，而这是任何 dsh 插件本来就会装的。其余能力全部来自 dsh 宿主本身（ctx.web、settings 能力缝、凭据能力缝、loopback bridge）。

一键把 CommandCode 所选订阅档位的模型同步到 DeepSeek Harness 的 llm-pi-ai 供应商配置中，并在设置页提供「一键创建/更新」按钮。已存在目标供应商时只刷新模型列表，用户配置的密钥与地址保持不变。除此之外，它还能替代 dsh 自带的 web 搜索（无需单独的搜索 API key），并在设置卡片里直接展示你的 CommandCode 账户用量。

为什么要用这个插件

• DSH 的 llm-pi-ai 供应商目录是静态快照，不会自己刷新。CommandCode 的 Provider API 现有 61 个模型且持续上新（每款有各自的上下文/价格/能力），手抄进供应商配置既不现实也容易过期。
• 官方的 /provider/v1/models 接口只返回 id / name / context_length，没有任何推理（thinking）或视觉能力信息。插件额外解析官方 GOAT 计划页 内嵌的完整目录（每模型含 reasoning / vision / caps / 四项定价 / 最低计划要求），把能力正确映射进 DSH 配置。
• 订阅分档：CommandCode 区分多个档位（Go / GOAT / Pro / Max），模型按 minPlanName 累计归属。插件提供「订阅类型」下拉框，每个档位对应独立的模型供应商，互不覆盖。
• 混合路由问题：CommandCode 提供 OpenAI 兼容（/chat/completions）与 Anthropic 兼容（/messages）两套端点，Claude 系列走错端点会直接 400。插件按模型自动拆分：Claude 进 commandcode-<档位>-anthropic（api: anthropic-messages），其余进 commandcode-<档位>-autosync（api: openai-completions）。
• 搜索无需第二把 API key：原版 dsh 要单独的 DEEPSEEK_API_KEY，模型的 web_search 工具才能用。本插件直接用你的 CommandCode 账户 key 提供同一个搜索工具——一个订阅同时覆盖聊天和搜索，不用再配置、再付费买 dsh 的搜索 key。
• 用量一眼可见：设置卡片直接展示 CommandCode 实时用量（5 小时/周窗口限额带进度条、本月剩余额度、请求/成本/Token 汇总），限流或额度耗尽之前就能提前看到。

订阅档位与供应商

订阅类型（下拉框）	模型总数	包含	生成的供应商
goat（默认）	43	Go + GOAT 开源模型（无 Claude）	commandcode-goat-autosync
pro	56	goat 全部 + Pro 档（Claude Sonnet/Haiku、GPT、Gemini 等）	commandcode-pro-autosync + commandcode-pro-anthropic
max	61	全部模型（含 Claude Opus/Fable、Fugu Ultra）	commandcode-max-autosync + commandcode-max-anthropic

档位是累计包含关系（minPlanName 字段定义，与官方文档一致：Pro 页写明 "includes everything in the GOAT plan"）。Go 档不提供 API 访问（403），因此不在选项中。

映射规则（上游 → llm-pi-ai YAML）

上游（官方目录）	llm-pi-ai 模型字段
contextWindow / context_length	contextWindow
vision: true	input: ["text", "image"]（否则 ["text"]）
reasoning: false	reasoningEfforts: false（禁用思考参数）
reasoning: true	reasoningEfforts: {low: low, medium: medium, high: high, xhigh: xhigh, max: max}
minPlanName	按所选档位过滤模型
vendor: Anthropic（或 id 以 claude- 开头）	归入 api: anthropic-messages 供应商
其余模型	归入 api: openai-completions 供应商

官方目录只标注 reasoning: true/false，没有每个模型的档位列表，所以这份映射取的是端点自己接受的档位字面量：reasoning_effort 只要不是 low|medium|high|xhigh|max，就会收到 HTTP 400 并被告知这一集合。这份映射同时是 dsh 判定「该模型会思考」的依据——没有它，dsh 会把模型当作不支持思考（选择器里没有挡位、请求也不带思考参数）；供应商级 compat: {thinkingFormat: "openai", supportsReasoningEffort: true} 只决定选中的档位怎么发（可通过 targetCompat 覆盖）。

Web 搜索（可选）

开启后，插件的 Command Code 搜索供应商 为 dsh 的模型 web_search 工具提供后端，走 Command Code Provider API 的 /alpha/web-search 端点——与聊天同一个 API Key、同一个账户，无需单独配置搜索 key 或端点。

为什么值得开： 原版 dsh 的 web 搜索由 DeepSeek 自己的 Messages API 提供，意味着你需要再配一个单独的 DEEPSEEK_API_KEY（在聊天费用之外另行计费），模型才能搜索。而只要你有 CommandCode 订阅，账户 key 本身就解锁搜索能力——本插件直接复用这把同一个 key 做搜索，从此不再依赖 dsh 自己的搜索 API key。一个订阅、一把 key，聊天和搜索都搞定。

• 注册在 dsh web 能力缝（ctx.web）上，provider id 为 commandcode；开关开启期间自动被选中（关闭或插件卸载时恢复之前的搜索供应商）。
• numResults 会被钳制在 Command Code 的范围内（1–10，默认 5）；结果映射为 dsh 的 WebSearchSource 结构（url/title/snippet）。
• 需要账户 key（默认 COMMANDCODE_API_KEY——与聊天供应商同一个凭据）。默认关闭；在设置卡片里开启（保存后立即生效，无需重启）。

用量统计（可选）

设置卡片同时展示账户用量——请求数、成功率、成本、Token、额度余额以及 5 小时/周窗口限额——数据在宿主侧用同一个账户 key 从账户端点（/alpha/whoami、/alpha/usage/summary、/alpha/billing/credits、/alpha/billing/subscriptions）抓取。key 不会离开宿主。

一眼看清： 当前 5 小时/周窗口还剩多少请求（带进度条与重置时间）、本月套餐还剩多少额度，以及最近的请求数/成功率/成本/Token 汇总——在撞上「限流」或「额度耗尽」之前就能提前发现，也能确认某次会话实际消耗是否符合预期。

• 每个端点独立降级：某个端点临时失败时显示局部数据提示而不会清空整卡；当所有端点以同一方式失败时，卡片会点明原因（key 无效 / 服务不可用 / 网络错误）。
• 用量端点位于 API 根路径（/alpha/*），与聊天的 /provider/v1 基址不同。如果部署环境不同，请使用卡片中的「用量/搜索 API 地址」字段。

安装

设置页依赖 dsh 的 settings API，而它在 0.1.7-alpha.1 这一代有变更，所以请按你正在运行的 dsh 选择插件 tag：

dsh	插件版本	安装命令（web profile）
0.1.7-alpha.1 … <0.2.0	v0.2.2	dsh plugin --profile web add github:CJYLZS/dsh-commandcode-provider#v0.2.2
0.1.7-alpha.1 … <0.2.0	v0.2.0（仅插件页）	dsh plugin --profile web add github:CJYLZS/dsh-commandcode-provider#v0.2.0
0.1.5-rc.3 及更早	v0.1.3	dsh plugin --profile web add github:CJYLZS/dsh-commandcode-provider#v0.1.3

#<ref> 就是安装时解析的 Git ref。不写时安装仓库默认分支，也就是最新的一代。

本地检出则用链接方式安装，避免拷贝——之后每次改动在重启后生效，不需要重新安装：

dsh plugin --profile web add link:/absolute/path/to/dsh-commandcode-provider

安装后重启 DSH Web，打开 设置 → CommandCode，或从 插件（Plugins） 页进入本 bundle 卡片并在该行点 配置（Configure）——两处打开的是同一个页面。选择订阅类型（默认 goat），点击 一键创建/更新。

dsh 兼容性

dsh	插件版本	设置界面
≥ 0.1.7-alpha.1	v0.2.2	设置页里独立的 CommandCode 页（settings.section），与插件页的行配置并存；会写入模型思考档位映射并真正执行自动同步
≥ 0.1.7-alpha.1	v0.2.0	仅插件页的行配置（plugins.row.config），没有设置页
≤ 0.1.5-rc.3	v0.1.3	注册式设置节（settings.register / installSection），与模型供应商并列显示

v0.2.2 与 v0.2.0 都把配置读写为自身 Cordis Config 的实时引用；v0.2.2 在此之上增加了设置页里的独立页面，并默认写入思考档位映射——正是它让模型选择器里出现思考挡位；v0.1.3 用的是已被删除的 settings.register / installSection 与已退役的 settings.plugin.item，因此在 0.1.7-alpha.1 及之后无法加载。

版本不匹配时的表现：浏览器里 web boot: 1 entry did not activate … pending (waiting for service: settingsScope)；在新版 dsh 上运行 v0.1.3 时，$DSH_HOME/logs/startup-*.log 里本包会有 settings.register is not a function。

API Key 配置

点击「一键创建/更新」后，插件会自动创建对应档位的模型供应商——默认是 commandcode-goat-autosync（其它档位为 commandcode-pro-… / commandcode-max-…）。API key 就配在这个自动创建的供应商上。

推荐方式——在 dsh 的模型页配置：

1. 进入 dsh 的 设置 → 模型（该页列出所有供应商，包括本插件自动创建的）。
2. 找到 commandcode-goat-autosync（或你档位对应的供应商），点击 编辑。
3. 在 API 密钥 输入框粘贴你的 CommandCode API key，点击 保存。

完成。key 会存入 dsh 的凭据库，聊天、web 搜索和用量卡片共用这一把。

备选方式——环境变量： 启动 dsh 前 export COMMANDCODE_API_KEY=cmd_xxx。（插件配置项 targetApiKeyEnv 可修改它读取的变量名。）

Key 从哪来？在 commandcode.ai Studio 的 API keys 页面创建（除 Go 计划外均可用；GOAT/Pro/Max 按套餐额度计费，Provider 计划按量付费）。

配置项

字段	默认值	说明
sourceURL	https://api.commandcode.ai/provider/v1/models	官方实时模型列表
catalogURL	https://commandcode.ai/docs/plans/goat	官方 GOAT 计划页（能力目录）；抓取失败时降级为纯列表
plan	goat	订阅档位：goat / pro / max
targetApiKeyEnv	COMMANDCODE_API_KEY	凭据环境变量名
targetBaseURL	https://api.commandcode.ai/provider/v1	API 基地址
targetCompat	{thinkingFormat: "openai", supportsReasoningEffort: true}	openai 路由的 compat 覆盖
extraIds	[]	额外写入的私有模型 id（目录之外，进 openai 路由供应商）
includeReasoningEfforts	true	给目录里标 reasoning: true 的模型写入 low…max 档位映射；关掉后 dsh 会把它们当作不支持思考
autoSync	false	定时自动同步（默认关闭，需在卡片里手动开启）
autoSyncIntervalMs	6h	自动同步间隔（最小 60s）
webSearch	false	用 CommandCode 提供 dsh 的 web_search（/alpha/web-search）
usageBaseURL	https://api.commandcode.ai	用量/搜索端点所在 API 根路径（/alpha/*）

目标供应商名由 plan 推导（commandcode-<plan>-autosync / commandcode-<plan>-anthropic），不需要单独配置。

FAQ

切换档位会怎样？ 下拉框选择新档位并保存后，点「一键创建/更新」会创建该档位的供应商；旧档位的供应商保留不动（互不覆盖）。想清理旧供应商需手动删除。

为什么 goat 档没有 commandcode-goat-anthropic？ GOAT 档只含开源模型，没有 Claude，因此不会生成 Anthropic 路由供应商。

Claude 模型怎么用？ 官方 API 要求 Claude 走 /messages（Anthropic 格式）。pro / max 档自动创建 commandcode-<plan>-anthropic 供应商（api: anthropic-messages）；若在 OpenAI 路由供应商里调用 Claude id，会得到 400。

为什么有些模型没有思考档位？ 只有目录里标了 reasoning: false 的那些：它们写入 reasoningEfforts: false，dsh 因此不发思考参数，选择器里也不给档位。标 reasoning: true 的模型一律写入 low…max 映射——正是这份映射让挡位出现在选择器里。模型列表和映射都在同步时才重写：改了选项后要点一次「一键创建/更新」（或等自动同步跑一轮）才会落到供应商配置里。

调用超出档位的模型会怎样？ 目录按 minPlanName 严格过滤，写入的模型都在所选档位内；上游列表若新增模型而目录尚未收录，插件会按纯列表写入全部（降级模式，结果中带 warning 提示）。