@fevex/deepseek

Official DeepSeek ModelGateway adapter for Fevex. See the compatibility and release policy.

npm install @fevex/core @fevex/deepseek

import { createDeepSeek } from '@fevex/deepseek';

const model = createDeepSeek({
  apiKey: process.env.DEEPSEEK_API_KEY!,
  schemaPolicy: 'best-effort',
})('deepseek-v4-flash');

The adapter uses native fetch and implements only the Fevex model contract. Runtime behavior such as tool execution, validation, events and retries belongs to @fevex/core. Returned gateways include metadata: { provider: "deepseek", model: modelId } for traces, eval cost calculators and OpenTelemetry export.

Text, tools and continuation

Without outputSchema, DeepSeek content is returned as the exact provider string, even when it looks like JSON. With outputSchema, the adapter requires valid JSON before Fevex performs local schema validation. Truncated, filtered or malformed responses fail explicitly.

Thinking-mode tool calls require reasoning_content to be sent back on the next request. The adapter keeps it in opaque providerState, and Fevex carries that state automatically between model steps. Direct ModelGateway consumers must pass it unchanged; it is run-local and must not be inspected. The adapter's stateCodec lets Fevex persist it only inside a private durable checkpoint.

reasoning: "none" disables thinking. low, high and max map directly to DeepSeek's native reasoning_effort; neutral xhigh maps to high for deepseek-v4-flash and max for deepseek-v4-pro. minimal and medium are rejected before HTTP with PROVIDER_REASONING_UNSUPPORTED. Provider-only low, high and max remain available through modelOptions when core reasoning is provider-default.

The adapter always uses native Chat Completions SSE with usage enabled. Visible content is streamed immediately, while reasoning and partial tool calls are accumulated privately until the required [DONE] marker.

Tool names must match [A-Za-z0-9_-]{1,64}. Fevex-owned fields such as model, messages, tools, tool choice, schema format, parallel calls and output caps override conflicting modelOptions. Provider tools outside Fevex's function contract are not enabled by this gateway.

Schema policy

schemaPolicy defaults to "strict". DeepSeek strict tool schemas require a baseURL ending in /beta; the adapter validates their documented subset and sets strict: true on every function according to the Tool Calls guide. Unsupported keywords throw DeepSeekError with code: "PROVIDER_SCHEMA_UNSUPPORTED" before fetch runs.

DeepSeek JSON Output guarantees valid JSON but does not enforce a supplied JSON Schema, so final outputSchema is rejected in strict mode. Use schemaPolicy: "best-effort" to include the schema as a model instruction and let Fevex validate the response locally. This mode does not guarantee adherence and the caller may need to retry invalid output.

const strictTools = createDeepSeek({
  apiKey: process.env.DEEPSEEK_API_KEY!,
  baseURL: 'https://api.deepseek.com/beta',
})('deepseek-v4-flash');