dsh-llm-motomoto

dsh-llm-motomoto

A DeepSeek Harness bundle for the MotoMoto relay: everything a request must do that the provider route cannot say, plus a status card in Settings → MotoMoto relay and the sidebar's Plugins → dsh-llm-motomoto detail page (dsh >= 0.1.7).

What belongs to whom

This bundle deliberately ships no provider route. The route — protocol, baseURL, credential reference, models — is the user settings layer's to own: the llm-pi-ai: section of ~/.dsh/settings.yaml. dsh-settings resolves that section from the user layer alone when a bundle ships no base, so the MotoMoto group in your model picker is exactly what you configured, and this bundle never fights your edits.

What the bundle contributes is the relay's wire contract, which no route field can express:

Everything is scoped to the relay host (motomoto.lol by default): requests to any other host pass through the fence untouched, and unloading the plugin restores the fetch it replaced.

MotoMoto's public homepage describes the API as best-effort, with no availability guarantee. A valid Chat Completions request can still receive an upstream 502; the plugin cannot repair an outage behind the relay. See motomoto.lol.

The settings section

On dsh >= 0.1.7, open Settings → MotoMoto relay (Chinese: 设置 → MotoMoto 中转站) for the route status card, or Plugins → dsh-llm-motomoto from the sidebar. The runtime's live configuration is served under the llm-motomoto entry. On older hosts, the fence's choices live under the llm-motomoto: key of ~/.dsh/settings.yaml:

llm-motomoto:
  host: motomoto.lol              # requests to this host carry the identity below
  userAgent: codex_cli_rs/0.46.0 (Windows 10.0.26200; x86_64) WindowsTerminal
  originator: codex_cli_rs        # empty string omits the header
  headers: {}                     # extra headers, applied last; an empty value removes the header
  stripBodyParams:                # request-body fields the backend rejects
    - max_output_tokens
  announce: true

Every knob except announce is live: a settings change reaches the next request without a restart. announce is read once at activation, so it stays out of the form.

On dsh >= 0.1.7 the Host derives this form from the exported Config schema and serves only the fields marked volatile, so host, userAgent, originator, headers and stripBodyParams carry that mark (liveField in lib/index.js); an unmarked field would be absent from the page, not read-only. The written value is committed into the running configuration reference and read again per request, which is what makes the change apply without a reload. announce is deliberately unmarked and lives in the composed bundle config.

The card itself has a seat per generation: dsh >= 0.1.7 binds the route once through configForms.get('llm-pi-ai') and shares it between settings.section (list id llm-motomoto, with the localized relay title) and plugins.bundle.config, keyed by the package name (dsh-llm-motomoto). Both slots are awaited independently, so either page can be absent or reload without removing the other. Older hosts use the settingsScope binder and settings.plugin.item, keyed by namespace. Both are reached through a dynamic ctx.inject, so the static inject list holds only slots and locale — declaring a service no provider supplies leaves the fiber inactive, which the Host reports as a renderer boot failure for the whole bundle.

The route to configure beside it

The bundle expects the motomoto provider in your user layer, e.g.:

llm-pi-ai:
  providers:
    motomoto:
      displayName: MotoMoto
      apiKeyEnv: MOTOMOTO_API_KEY
      api: openai-completions     # MotoMoto advertises /chat/completions
      baseURL: https://motomoto.lol/v1
      compat:
        supportsDeveloperRole: false
        maxTokensField: max_tokens
      models:
        - id: gpt-5.5
          ...

Route notes: set maxTokensField: max_tokens for Chat Completions relays that expect that field; the adapter may otherwise choose max_completion_tokens for newer OpenAI model IDs. A chatTemplateKwargs: {} left by the settings page is an empty object and is ignored.

Where to look after installing

1. Model picker — the MotoMoto group, from your route.
2. Settings → MotoMoto relay (Chinese: 设置 → MotoMoto 中转站) — a direct navigation entry for the live route status (group, endpoint, credential reference, models) on dsh >= 0.1.7.
3. Sidebar Plugins → dsh-llm-motomoto — the same card on the package detail page. Older hosts still show it under Settings → Plugins.

After editing the hand-written lib/client.js in a linked local install, refresh the current DSH page. Without a build watcher, do not expect automatic client updates.

Security notice

The API key is never stored in this package. Provide it as the MOTOMOTO_API_KEY credential.

Install from this source directory

git clone https://github.com/HandsYe/dsh-llm-motomoto
dsh plugin --profile desktop add "file:<克隆下来的仓库路径>"

Then fully restart the desktop app. Verify the load:

Select-String -Path "$env:APPDATA\DSH Desktop\logs\dsh-*.log" -Pattern 'llm-motomoto' | Select-Object -Last 5

Test

npm install
npm test

28 offline checks: the patch's shape (it loads the runtime half and declares no route), the manifest and secret scan, the client bundle's registration and rendering (including that no static inject names a service this Host no longer provides), the live-field marks the Plugins page derives its form from, and the fence's identity rewrite, header extras, body strip, stream close at each terminal event, error-body pass-through, and unload behavior — plus one check that a live edit takes effect with no installSection and no re-apply, which is the dsh >= 0.1.7 posture.

test/live-probe.mjs (not part of the suite) sends one request shaped exactly like a real agent turn, directly and through the fence:

node test/live-probe.mjs fenced gpt-5.5 high
node test/live-probe.mjs raw gpt-5.5 high

Uninstall

dsh plugin --profile desktop remove dsh-llm-motomoto