dsh-llm-nous
面向 DeepSeek Harness 的 Nous Portal OpenAI 兼容 LLM 适配器包
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:patrickluvsoj/dsh-llm-nous#044a27b9dc5ba1ae4acf73045b2da128824f733bLlm Nous 以 dsh-llm-nous 发布，当前版本为 0.1.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Give DeepSeek Harness a direct line to the Nous Portal model buffet — without forking Harness or touching its core.
dsh-llm-nous is a standalone Harness bundle. It registers a nous provider route, speaks Nous's OpenAI-compatible API, and keeps the provider-specific weirdness where it belongs: inside the plugin.
noushttps://inference-api.nousresearch.com/v1NOUS_API_KEY[DONE]delta.reasoningnpm is optional. Pick your flavor:
# GitHub — easiest for sharing
dsh plugin --profile web add github:patrickluvsoj/dsh-llm-nous
# Local checkout
dsh plugin --profile web add /absolute/path/to/dsh-llm-nous
# Tarball
dsh plugin --profile web add ./dsh-llm-nous-0.1.0.tgz
# npm, if published there later
dsh plugin --profile web add dsh-llm-nous
For a Harness source checkout, prefix the first two commands with pnpm and add --ignore-workspace-root-check to the install command:
pnpm dsh plugin --profile web add --ignore-workspace-root-check github:patrickluvsoj/dsh-llm-nous
export NOUS_API_KEY='your-nous-portal-key'
For a source checkout, a gitignored .env works too:
NOUS_API_KEY=your-nous-portal-key
The key belongs in the environment or Harness's credential store — not in Cordis YAML, Git, screenshots, or package files. Tiny security basics; civilization survives another day.
dsh --profile web
Or from the Harness source checkout:
pnpm dsh --profile web
The bundle defaults new sessions to:
deepseek/deepseek-v4-flash-0731
That model is paid and requires available Nous Portal credits. The bundle also advertises this free model from the usage-ranked catalog:
poolside/laguna-s-2.1:free
Harness exposes configured providers in Settings → Models. After dsh-llm-nous is loaded, the Nous provider and its advertised models can appear in the model picker.
This is the upstream Harness model-settings UI; the Nous card/model entries appear after the Nous bundle is installed.
Select a model in the picker and it becomes the default for new sessions. A session that has already sent a request keeps the provider/model recorded in its own log. Model changes apply on the next request; no server restart is needed for an ordinary selection change.
With the current Harness composer picker, the live Nous catalog is no longer squeezed into one heroic dropdown. The picker keeps the first 12 models from each provider compact, then lets you:
Search all Nous Portal models and pin favorites
Live discovery is the default (catalogMode: live). The plugin makes an authenticated GET /models request to the configured Nous endpoint and merges compatible live routes into the picker. These twelve curated routes stay pinned first, preserving their OpenRouter trailing-week usage rank:
deepseek/deepseek-v4-flash-0731
xiaomi/mimo-v2.5
tencent/hy3
deepseek/deepseek-v4-flash
openai/gpt-5.6-luna
z-ai/glm-5.2
google/gemini-3.7-flash
deepseek/deepseek-v4-pro
minimax/minimax-m3
poolside/laguna-s-2.1:free
anthropic/claude-opus-5
openai/gpt-5.6-sol
The ranking measures adoption by prompt-plus-completion token volume, not model quality. OpenRouter-only routes and variants absent from Nous are skipped. Source and attribution: OpenRouter rankings, as of 2026-08-25. OpenRouter is not queried at runtime; runtime discovery calls only the configured Nous /models endpoint. The request carries the resolved Nous bearer credential and Harness's standard, non-secret User-Agent attribution header.
The merge is deterministic:
models remain first, in configuration order, and their name, description, context window, and output cap win over conflicting live metadata;:batch routes, tilde aliases such as ~latest, expired routes, non-text-output routes, routes without text input, and entries whose explicit supported_parameters array lacks tools are omitted;listModels() and exact-model resolution share the same successful live snapshot for one hour. Each refresh is limited to five seconds and 4 MiB of response bytes. Before the first successful refresh—or when credentials are missing—the configured/curated list remains available. After a success, an expired snapshot is returned immediately while one background refresh runs. If a refresh fails, the last-good snapshot remains available and credential/catalog retries pause for one minute by default. API keys and raw catalog responses are never cached or written to disk.
The Nous catalog is much larger. A model ID omitted from discovery can still be passed through unchanged by setting it explicitly; it then uses the plugin-wide context and output defaults.
The plugin does not add a separate terminal picker command. For terminal or headless runs, select the model with a normal Harness patch:
- id: agent-default-model
config:
provider: nous
model: stepfun/step-3.7-flash:free
Run it:
dsh --profile headless --patch ./nous-model.cordis.yml \
'Reply with exactly: NOUS_OK'
Create nous-model.cordis.yml:
- id: agent-default-model
config:
provider: nous
model: openai/gpt-oss-120b
Then launch:
dsh --profile web --patch ./nous-model.cordis.yml
The adapter passes the model ID through to Nous unchanged. The model does not have to be in the advertised catalog.
The defaults are live discovery, a one-hour successful-response cache, a five-second request timeout, and a one-minute failed-refresh cooldown:
- id: llm-nous
config:
apiKeyEnv: NOUS_API_KEY
baseURL: https://inference-api.nousresearch.com/v1
catalogMode: live
catalogCacheTtlMs: 3600000
catalogTimeoutMs: 5000
catalogRetryCooldownMs: 60000
models:
- id: deepseek/deepseek-v4-flash-0731
name: DeepSeek V4 Flash 0731
contextWindow: 1310720
models is the ordered curated list and the fallback used when live discovery has not succeeded. It is separate from the favorites a user pins in the picker. Supplying models replaces the built-in twelve-entry list. Set catalogMode: curated to disable /models requests completely and expose only this configured list. catalogCacheTtlMs, catalogTimeoutMs, and catalogRetryCooldownMs must be positive integer milliseconds no greater than the platform timer limit. The cooldown applies after failed initial or background refreshes, preventing repeated credential lookup and /models requests until it expires.
A patch replaces the entire targeted config, so keep apiKeyEnv and baseURL when overriding discovery settings.
maxTokens limits generated tokens for one response. That includes visible text, reasoning, and model-generated tool calls. It is not the conversation-memory limit; that is the model's context window.
Raise it for a model with enough context:
- id: llm-nous
config:
apiKeyEnv: NOUS_API_KEY
baseURL: https://inference-api.nousresearch.com/v1
maxTokens: 32768
A patch replaces the entire targeted config, so keep apiKeyEnv and baseURL when overriding llm-nous.
Per-model caps are supported too:
- id: llm-nous
config:
apiKeyEnv: NOUS_API_KEY
baseURL: https://inference-api.nousresearch.com/v1
maxTokens: 32768
models:
- id: stepfun/step-3.7-flash:free
contextWindow: 256000
maxTokens: 8192
- id: deepseek/deepseek-v4-pro-0813
contextWindow: 1048576
maxTokens: 32768
Effective precedence:
explicit request maxTokens
→ model-specific maxTokens
→ plugin-wide maxTokens
→ 8,192 default
Inspect the composed profile:
dsh --profile web --dump-config
Look for the dsh-llm-nous layer and provider: nous.
Run a tiny smoke test:
dsh --profile headless \
--patch ./nous-model.cordis.yml \
'Reply with exactly: NOUS_OK'
The request reached Nous, but the selected model is paid and the Portal account has no available credits. Choose a free model or add credits.
The requested output cap plus the prompt, conversation, and tool definitions exceeds the model's total context window. Lower maxTokens or choose a model with more context.
MISSING_CREDENTIALHarness cannot find NOUS_API_KEY. Export it in the launching environment or store it through the Harness credential UI.
DUPLICATE_ADAPTERTwo plugins are registering the nous route. Keep only one dsh-llm-nous bundle active in the profile.
pnpm install
pnpm run build
pnpm test
The runtime artifact is lib/index.mjs. It bundles its runtime dependencies, so an installed bundle does not need a sibling Harness checkout.
MIT