@arnilo/prism-provider-neuralwatt

NeuralWatt first-party provider package for Prism (OpenAI-compatible Chat Completions).

import { createExtensionKernel } from "@arnilo/prism";
import { createNeuralWattProviderPackage } from "@arnilo/prism-provider-neuralwatt";

const kernel = createExtensionKernel();
await kernel.load([createNeuralWattProviderPackage({ apiKey: "fake-neuralwatt-key" })]);

Exports

• createNeuralWattProviderPackage() — inert provider package (provider + models + api_key auth).
• createNeuralWattProvider() — raw provider adapter.
• defineNeuralWattModel() — add custom NeuralWatt model configs.
• neuralWattModels — curated registry of NeuralWatt featured aliases (glm-5.2, glm-5.2-fast, glm-5.2-short, glm-5.2-short-fast, kimi-k2.6, kimi-k2.6-fast, kimi-k2.7-code, qwen3.5-397b, qwen3.5-397b-fast, qwen3.6-35b, qwen3.6-35b-fast) plus the legacy kimi-k2 entry.
• listNeuralWattModels() — explicit one-call /v1/models discovery helper that maps returned aliases to ModelConfig/ModelCost.
• getNeuralWattQuota() — explicit one-call /v1/quota helper returning typed account quota (balance, usage, limits, subscription, key). Requires an API key; caller owns throttling (1 rps per customer).
• neuralWattBody() / neuralWattEvents() / toUsage() — serializer, SSE parser, usage mapper.
• neuralWattEventsWithTelemetry() — like neuralWattEvents() but also yields neuralwatt:telemetry events parsed from : energy / : cost SSE comments.
• parseNeuralWattComment() / parseNeuralWattEnergy() / parseNeuralWattCost() / mapNeuralWattTelemetry() — energy/cost telemetry parsers for streaming comments and non-streaming top-level fields.
• classifyNeuralWattError() / neuralWattHttpError() — retry classifier for NeuralWatt status codes, Retry-After, and retry_strategy.

NeuralWatt-specific request fields flow through ProviderRequestOptions.compat/extra: reasoning_effort, thinking_token_budget, chat_template_kwargs, tool_choice. Default base URL: https://api.neuralwatt.com/v1.

The static catalog records documented context windows, text/image input support, tool calling, reasoning vs. fast variants, JSON-mode compat metadata, and implicit caching. It does not guess ModelConfig.cost; exact per-alias input/output/cache-read prices come from NeuralWatt's /v1/models response via listNeuralWattModels().

import { listNeuralWattModels } from "@arnilo/prism-provider-neuralwatt";

const models = await listNeuralWattModels({ apiKey: "fake-neuralwatt-key", fetch });

import { getNeuralWattQuota } from "@arnilo/prism-provider-neuralwatt";

const quota = await getNeuralWattQuota({ apiKey: "fake-neuralwatt-key", fetch });
// quota.usage?.current_month?.energy_kwh, quota.balance?.balance_usd

Discovery is never called during package setup or generation.

Energy and cost telemetry

NeuralWatt streams : energy {...} / : cost {...} as SSE comments and emits top-level energy/cost fields on non-streaming responses. Standard SSE clients ignore comments, so use neuralWattEventsWithTelemetry() (or the parse* / mapNeuralWattTelemetry() helpers) to observe them. generate() uses neuralWattEvents() and stays streaming-only, so telemetry is opt-in. Telemetry carries usage/cost numbers only — never prompts, API keys, or headers.

Security defaults

• No network calls during import, setup, build, or default tests.
• No automatic environment, file, keychain, or shell credential lookup.
• API keys are resolved per request from caller-supplied values or resolvers and redacted from errors via redactSecrets.
• Provider-owned headers (content-type, authorization) are applied last and win over caller-supplied ProviderRequest.options.headers.
• Live tests are opt-in behind NEURALWATT_API_KEY (plus PRISM_LIVE_PROVIDER_TESTS=1); default tests are network-free.

Cache behavior

• cache: { kind: "implicit" } — NeuralWatt prefix caching is automatic; no explicit cache payload (cache_control, cacheKey, prompt_cache, or cacheRetention) is sent regardless of cache options.
• cacheRetention: "none" disables Prism cache-control hints only; it does not disable the implicit backend prefix cache. Hosts relying on cache hits should keep their stable prompt prefix byte-stable.
• prompt_tokens_details.cached_tokens maps to Usage.cacheReadTokens. NeuralWatt does not report a cache-write token today, so Usage.cacheWriteTokens is never fabricated (stays undefined).

See docs/providers/neuralwatt.md for the full provider API page.