dsh-commandcode-provider

Languages: English | 中文

A custom LLM provider plugin for DeepSeek Harness (dsh) that connects dsh to Command Code — a faithful port of pi-commandcode-provider onto dsh's LLM seam.

Works with every Command Code plan, including the $1/month Go plan. Even the Go plan — the only one without Provider API access — gives you an API key in the Studio; that key authenticates the CLI/agent login. This plugin uses that same key against Command Code's own /alpha/generate endpoint, not the traditional Provider API protocol, so it works even when a plan has no Provider API access.

Disclaimer: This is an unofficial, community-maintained integration. It is not affiliated with, endorsed by, or supported by Command Code. You need your own Command Code account. Command Code's terms, availability, and pricing apply.

Requires dsh 0.2.0-rc.2 or newer. This release targets the 0.2 LLM and settings seam; on an older dsh the entry will not load. Settings edited on Settings → Models → Command Code → Edit apply live — no dsh restart.

Quick Start

Newbie-friendly: one command to install — the plugin mounts itself, no config files to edit.

1. Install

dsh plugin --profile web add dsh-commandcode-provider@0.2.0

The package declares dsh.bundle, so dsh plugin add installs it AND automatically joins it to the profile's bundle layers — the shipped cordis.patch.yml mounts the provider row for you.

Do not also add a manual cordis.patch.yml row for this plugin — the bundle patch already mounts it; a second row would double-register the provider.

2. Restart dsh

The provider then appears in the web UI automatically.

3. Get your API key

Even on the Go plan, Command Code gives you an API key used for authentication:

1. Open commandcode.ai and sign in.
2. Go to Studio → API Keys (in the sidebar).
3. Click Generate API key and copy it.

This key is an authentication key for using Command Code — it is not a "Provider API" key. The plugin speaks Command Code's own /alpha/generate protocol with it, so no Provider API access is required, and the Go plan works.

4. Configure the API key (either way)

** Command:** In a dsh chat, run:

/commandcode-setkey

and paste the key when prompted.

That's it — pick a Command Code model from the model picker and start chatting.

How it works

This plugin is not a traditional API-protocol provider:

• It talks to Command Code's own streaming endpoint https://api.commandcode.ai/alpha/generate with the same wire protocol the official CLI uses (x-command-code-version, x-cli-environment, x-project-slug, … headers).
• The API key from Studio is an authentication credential for that endpoint — the same one cmd login writes to ~/.commandcode/auth.json.
• Because it rides the CLI protocol instead of the Provider API, it works on every plan, including Go, and even models the Provider API cannot serve.
• Model discovery still uses the public Provider API catalog endpoint when reachable, with an offline cache at <dsh home>/commandcode/commandcode-models.json; if that endpoint is unreachable (e.g. a Go plan), the last cached catalog keeps serving.

Functionality

Everything the pi plugin does, on the dsh host plane:

• Provider registration — registers the commandcode provider route on ctx.llm with the same /alpha/generate streaming protocol, retry/timeout/abort semantics, and message/tool conversion as the pi plugin (paired tool-call filtering, data-URL image forwarding, legacy schema normalization).

• Model discovery with offline cache — fetches the catalog from https://api.commandcode.ai/provider/v1/models and caches it locally (versioned, atomic writes); offline, the last cached catalog keeps serving.

• Reasoning metadata — models with known Command Code effort support advertise their thinking levels through the LLM seam; a selected level is sent as params.reasoning_effort.

• Image input — models marked image in the catalog accept image blocks (resolved through dsh's durable attachment service); text-only models reject images before any network request.

• Authentication — credentials resolve per request through, in order:

  1. the Harness credential seam (ctx.credentials, what the web Models page writes),
  2. the trusted launch environment (COMMANDCODE_API_KEY),
  3. existing Command Code auth files: ~/.commandcode/auth.json, ~/.pi/agent/auth.json, ~/.omp/agent/auth.json.

  No browser OAuth flow — the key is entered in the web UI (Settings → Models → Command Code → Edit) or via /commandcode-setkey.

• Commands — /commandcode-refresh (re-fetch + re-register the catalog), /commandcode-status (redacted diagnostics), /commandcode-setkey (store an API key).

• Pricing display — the static per-model cost table from the pi plugin (USD per million tokens).

• Error hygiene — context-overflow wording is normalized to CONTEXT_WINDOW_EXCEEDED, and every error/diagnostic text is redacted.

Configuration

The commandcode-provider settings section accepts (all optional):

Field	Default	Meaning
apiKeyEnv	COMMANDCODE_API_KEY	Credential reference resolved per request
displayName	Command Code	Name shown by selectors
baseURL	COMMANDCODE_API_BASE env → https://api.commandcode.ai	API base for /alpha/generate (usually leave at default)
modelsUrl	COMMANDCODE_MODELS_URL env → Provider API	Model discovery endpoint
modelsTimeoutMs	COMMANDCODE_MODELS_TIMEOUT_MS env → 10000	Discovery timeout
modelsCachePath	COMMANDCODE_MODELS_CACHE env → <dsh home>/commandcode/commandcode-models.json	Catalog cache path
models	—	Optional explicit catalog entries; each overrides (or adds to) the discovered model by id
defaultContextWindow	262144	Context capacity for models neither the catalog nor an override sizes
defaultMaxTokens	32768	Output capability for models neither the catalog nor an override sizes
timeoutMs	—	Per-attempt HTTP request timeout
streamIdleTimeoutMs	300000	Max idle time while one stream read is outstanding
retryPolicy	normal defaults	Provider-owned model-request retry policy

# $DSH_HOME/settings.yaml
commandcode-provider:
  apiKeyEnv: COMMANDCODE_API_KEY   # credential reference the Models page writes

Changelog

0.2.0

• Adapted to the dsh 0.2.0-rc.2 seam: ToolCallId, first-class role: 'tool' result messages, JsonObject tool arguments, and the schema-derived settings form that replaced installSettingsSection.
• Settings edits apply live: every config field is a live reference (Volatile), so the plugin re-reads settings per use and refreshes its provider route, catalog, and Models-page directory entry on loader/volatile-update.
• Seam versions pinned to the dsh release being targeted, so a future seam drift fails at install instead of silently not loading.
• Tests and typecheck resolve the published seam packages — no local deepseek-harness checkout required.

0.1.2

• Removed the browser OAuth flow (and the local callback server): no more /commandcode-login, no browser pop-up. The API key is entered directly in Settings → Models → Command Code → Edit (single API key field) or via /commandcode-setkey.
• Models-page card for Command Code shows only the API key field; no API address to configure.

0.1.0 / 0.1.1

• Initial port of pi-commandcode-provider: /alpha/generate streaming, model discovery + cache, reasoning efforts, image input, pricing display.

Development

npm install        # dev + test dependencies
npm run typecheck  # strict tsc against the harness seam packages
npm run build      # emit lib/ (ESM + declarations)
npm test           # vitest suite (wire protocol, discovery, cost, adapter, plugin entry)

Typecheck and vitest resolve the @deepseek-ai/* seam packages from node_modules, pinned in package.json to the dsh release this plugin targets — the same published API surface the host loads, with no local harness checkout.

License

MIT

---

中文文档: README.zh.md