dsh-nanogpt

dsh-nanogpt

NanoGPT PRO subscription models for DeepSeek Harness: a curated model picker, ×2 rate badges, reasoning levels and the remaining weekly quota — without spending the subscription allowance.

NanoGPT exposes 302 subscription-included text models through one OpenAI-compatible API. Pointing DeepSeek Harness at it directly works, but the model picker then shows all 302, the ×1/×2 billing multiplier is invisible, and the remaining weekly allowance lives on a website. This plugin adds the missing pieces:

• Only the models you mark. The provider card on the Models page lists the subscription catalog with search and a “tool calling only” filter; ticking models writes them into llm-pi-ai.providers.nanogpt.models, and the model picker shows exactly those.
• The ×1/×2 rate. NanoGPT does not return the multiplier from any endpoint — it is server-side accounting. The plugin carries the documented ×2 set (12 model ids) and shows it as a column in the picker and as a ×2 suffix in the model name.
• Reasoning levels. Each model's reasoning_efforts becomes the harness's Effort menu (none → off, the rest keep their names), so “Off” really turns thinking off.
• The remaining quota. A row inside the usage-stats footer block — the same design as its Go/DeepSeek rows — with the text-model percentage and the image percentage, plus a quota bar in the provider card.

The plugin only issues GET /api/subscription/v1/models and GET /api/subscription/v1/usage. It never calls /chat/completions, so it does not spend the allowance it reports.

Requirements

What	Why
DeepSeek Harness ≥ 0.2.2-rc.1	The plugin uses the 0.2.x client contract (settings.models.provider-card, ctx.configForms, ctx.locale) and the llm-pi-ai provider route.
Node.js ≥ 22	The harness runtime.
A NanoGPT PRO subscription	The catalog endpoint returns subscription-included models only.
A NanoGPT API key	Stored as NANOGPT_API_KEY in ~/.dsh/.credentials.yaml (or entered on the Models page).
The llm-pi-ai bundle	The provider route lives there; it ships with the harness.
@xfqz86/dsh-usage-stats (optional)	The quota row is inserted into its footer block. Without it the card still shows the quota.

Install

# 1. Add the bundle
dsh plugin --profile web add dsh-nanogpt

# 2. Make sure the bundle is enabled in the profile: dsh.profile.bundles must list "dsh-nanogpt"
#    (the plugin manager / market toggle does the same thing)

# 3. No restart needed for a new plugin: the loader picks up the new row and the
#    client bundle live. A restart is only needed after editing the host half of
#    an already-loaded plugin, or after moving the plugin to another path.

The plugin adds one row to the profile patch layer:

- id: nanogpt
  name: dsh-nanogpt

The row id nanogpt is the plugin's settings namespace — do not rename it, or the settings form and the patch layer will stop finding each other.

Activation. A new plugin is picked up live: the loader registers the new row and the client bundle without a restart. A restart is needed only after editing the host half (index.js, lib/*.js) of an already-loaded plugin, after changing the plugin's path, or after changing the client package meta (package.json dsh.client/exports). The client half (client.js) and volatile config fields are picked up live.

Then add the provider route to llm-pi-ai in ~/.dsh/profiles/web/cordis.patch.yml (the plugin does not create it for you, because a route needs at least one model):

- id: llm-pi-ai
  config:
    providers:
      nanogpt:
        displayName: NanoGPT
        apiKeyEnv: NANOGPT_API_KEY
        api: openai-completions
        baseURL: https://api.nano-gpt.com/api/v1legacy
        models:
          - id: deepseek/deepseek-v4.1-flash
            name: DeepSeek V4.1 Flash
            contextWindow: 1000000
            maxTokens: 384000
            input: [text, image]
            reasoningEfforts: { off: none, low: low, high: high, max: max }

v1legacy is the base URL NanoGPT's own DeepSeek Harness guide recommends: it returns reasoning in reasoning_content, which the harness keeps across tool calls more often than the reasoning field of /api/v1.

Use

1. Open Settings → Models and find the NanoGPT row.
2. The card shows the remaining weekly allowance and the image quota.
3. Click Choose models, search, tick what you need, and Apply. The model picker now shows exactly those models; ×2 models carry the ×2 suffix.
4. Pick a model and an Effort level in the composer.

If a model was added through the built-in editor instead (its “Fetch available models” reads {baseURL}/models without ?detailed=true, so it gets ids only), the card shows a warning and a Fill from catalog button: it supplies the missing context window, output cap, input types and reasoning levels without touching values you already set.

Settings

Plugins page → the nanogpt row:

Field	Default	Meaning
apiKeyEnv	NANOGPT_API_KEY	Credential reference holding the key.
baseURL	https://api.nano-gpt.com/api	NanoGPT API base.
modelsCacheTtlMs	600000	Catalog cache lifetime.
usageCacheTtlMs	60000	Quota cache lifetime.
extraRate2	''	Extra ×2 model ids, comma-separated.

Verify without spending the allowance

# Live API, GET-only: catalog size, ×2 set, weekly quota, image quota
node tools/probe.mjs

# Offline: route wiring, normalization, no-key/401/429 shapes
node tools/smoke-host.mjs

# Offline: client bundle, provider card, quota row placement
node tools/smoke-client.mjs

How it works

llm-pi-ai (shipped bundle)
  providers.nanogpt ──────────────► the harness model picker
        ▲                                   ▲
        │ configForms.mutate                 │ (only the marked models)
        │                                   │
dsh-nanogpt                                │
  host: GET /dsh-nanogpt/models ───────────┘  (catalog + rate + reasoning levels)
        GET /dsh-nanogpt/usage                 (quota)
  client: settings.models.provider-card        NanoGPT card: quota + picker
          usage-stats footer row               text/image percentages

• The host resolves the key through ctx.credentials on every request, caches both responses, and serves the last good payload when NanoGPT fails.
• The client writes the chosen models through ctx.configForms.get('llm-pi-ai').mutate(...) — the same path the shipped editor uses. llm-pi-ai.providers is volatile, so the picker updates without a harness restart.
• The quota row is inserted into the usage-stats footer block's DOM (data-usage-stats-footer, before the unifiedMain row) and re-inserted if the block re-renders. The usage-stats bundle has no extension point for a fourth source, and patching a third-party npm package that updates itself was not an option.

Limitations

• The ×2 set is a local map (NanoGPT does not expose it). If NanoGPT changes it, add the ids to extraRate2 — no code change needed.
• maxTokens is taken from the model's max_output_tokens (null → 65536). If a route rejects that cap, lower it in the built-in model editor.
• The quota row depends on the usage-stats DOM contract. If that plugin changes data-usage-stats-footer or unifiedMain, the row falls back to a separate footer row.
• The plugin does not create the nanogpt provider route: a route needs at least one model, and the catalog is the user's choice.

License

MIT — see LICENSE.