@alaxrpg/dsh-sensenova-provider

Unofficial DeepSeek Harness (DSH) LLM provider plugin for SenseNova (OpenAI-compatible API). Registers the sensenova provider route with a Models-page card, a live model catalog, and multi-account API-key rotation on a single shared base URL.

Reference implementation: @mars-sea/dsh-commandcode-provider (MIT). This plugin keeps only the multi-account connection feature; it drops login flows, usage dashboards, plan windows, and command-line tools.

Features

• Registers provider route sensenova (display name SenseNova).
• Live model catalog via GET {apiBase}/v1/models.
• Streaming over both wire protocols: POST {apiBase}/v1/responses (default, see below) and POST {apiBase}/v1/chat/completions (SSE; automatic fallback).
• Multiple API keys (accounts) on one shared base URL:
  • 429 Too Many Requests → does not cool down or rotate the key (protects the server-side per-key prompt cache); raises RATE_LIMIT with providerRetryAfterMs for the host retry layer (honors Retry-After, capped at 60 s; quota-class 429s carry a backoff floor). With quotaRotation enabled, quota-class 429s stickily switch to the next untried key.
  • 401 Unauthorized → rotate to the next key; the rejected key stays disabled until the stored credential changes.
  • All keys exhausted → surfaces RATE_LIMIT / INVALID_CREDENTIAL.
• Web settings page (Models-page card + dedicated settings section).
• Default wire protocol is Responses (auto mode): used whenever the request can be expressed faithfully, and automatically downgraded to Chat Completions when semantics would be lost (e.g. a stop sequence, or a reasoning effort away from the model's Responses default). Notable Responses differences: max_output_tokens counts reasoning tokens, temperature defaults to 0.6 (Chat uses 1), reasoning is replayed as a summary only, and stop is not supported (auto-downgrades to Chat).
• Image inputs: outbound images in production depend on a host-side attachment bridge (the 0.1.7-rc.2 adapter type surface has no channel for it yet). In this version images are sent as host-generated text placeholders — never silently dropped; the bridge is a follow-up host dependency.

Requirements

• DSH host >=0.1.2-alpha.3 (alpha line).
• Node.js >=22.

Compatibility note: the runtime import surface is a stable subset of @deepseek-ai/dsh-llm present since 0.1.1-rc.2 (LlmAdapter, LlmError, ReasoningEffortId, assertUsableApiKey, attributionHeaders, errorChain, resolveRetryPolicy); the drifting ToolCallId symbol is defined locally (src/brand.ts, identity-backed). Runtime loading on 0.1.1-rc.2 hosts works; the declared peer floor >=0.1.2-alpha.3 is the type-level support start (the ToolCallId type first appears there).

Install

dsh plugin --profile <name> add dsh-sensenova-provider

Or install from a local path:

dsh plugin --profile <name> add ./dsh-sensenova-provider

Then configure SenseNova in the Models page: the default credential environment variable is SENSENOVA_API_KEY, and the default base URL is https://token.sensenova.cn/v1.

Configuration

The plugin installs a llm-sensenova settings section with these fields:

Field	Type	Default	Meaning
apiBase	string	https://token.sensenova.cn/v1	Shared base URL for all accounts.
apiKeyEnv	credential-ref	SENSENOVA_API_KEY	Default account credential reference.
accounts[]	array	[]	Extra accounts: { id, label, apiKeyEnv }.
activeAccount	string	""	Preferred account id; empty means auto / first usable.
wireProtocol	auto/responses/chat-completions	auto	Wire protocol mode: auto prefers Responses and downgrades to Chat when semantics would be lost; forcing responses errors on requests it cannot express; chat-completions is the legacy escape hatch.
reasoningSummary	auto/concise/detailed	auto	Responses-only: verbosity of the reasoning summary (reasoning.summary).

API keys are stored through the DSH credentials service; they are never logged and never sent to the model.

Development

pnpm install
pnpm run typecheck
pnpm run build    # tsdown → lib/
pnpm test         # node --import tsx --test tests/**/*.test.ts

License

MIT