Xiaomi Reference Implementation

The Xiaomi provider integrates the Xiaomi MiMo API (api.xiaomimimo.com/v1), an OpenAI-compatible Chat Completions API with native thinking/reasoning support. It demonstrates boolean reasoning effort mapping, multi-turn reasoning content preservation, and max_tokens to max_completion_tokens request patching.

Module Structure

src/providers/xiaomi/
├── spec.ts         # XIAOMI_PROVIDER_SPEC declaration with capabilities and accessors
├── client.ts       # createXiaomiProviderEdge() factory
├── hooks.ts        # Provider-specific hooks (patchRequest, usage mapping, deltas)
├── protocol/       # Xiaomi-specific type definitions
│   ├── completions.ts  # Request/response DTOs
│   └── index.ts        # Barrel exports
└── index.ts        # Public exports

Spec Declaration

The XIAOMI_PROVIDER_SPEC in spec.ts declares all capabilities, accessors, and hooks:

export const XIAOMI_PROVIDER_SPEC: ProviderSpec<
  BridgeChatCompletionCreateRequest,
  ChatCompletion,
  ChatCompletionChunk,
  ChatCompletionRequest
> = {
  name: XIAOMI_PROVIDER_NAME,
  protocol: CHAT_COMPLETIONS_PROTOCOL,
  capabilities: XIAOMI_SPEC_CAPABILITIES,
  endpoint: { defaultBaseURL: DEFAULT_XIAOMI_BASE_URL },
  auth: BEARER_AUTH,
  toolName: DEFAULT_TOOL_NAME_CODEC,
  response: {
    firstChoice: xiaomiFirstChoice,
    finishReason: xiaomiFinishReason,
    outputText: xiaomiOutputText,
    usage: xiaomiResponseUsage,
  },
  stream: { deltas: xiaomiStreamDeltas },
  hooks: { patchRequest: xiaomiPatchRequest },
};

Capabilities

const XIAOMI_SPEC_CAPABILITIES: ProviderCapabilities = {
  parameters: {
    supported: new Set([
      "stream", "temperature", "top_p", "max_output_tokens",
      "user", "reasoning", "text.format",
    ]),
  },
  tools: {
    supported: new Set([
      "function", "web_search",
      "local_shell", "shell", "apply_patch",
      "custom", "namespace",
    ]),
    degraded: new Map([
      ["web_search_2025_08_26", "web_search"],
      ["web_search_preview", "web_search"],
      ["web_search_preview_2025_03_11", "web_search"],
      ["local_shell", "function"],
      ["shell", "function"],
      ["apply_patch", "function"],
      ["custom", "function"],
      ["namespace", "function"],
    ]),
    maxTools: 128,
  },
  toolChoice: { supported: new Set(["auto"]) },
  responseFormats: { supported: new Set(["text", "json_object"]) },
  reasoning: { effort: "boolean" },
  streaming: { usage: true },
};

Notable capability details:

• Boolean reasoning: Xiaomi uses reasoning.effort: "boolean", so the bridge maps reasoning effort to thinking: { type: "enabled"/"disabled" }.
• Tool choice limited: Only auto is supported for tool_choice.
• Tool degradation: Codex built-in tools (local_shell, shell, apply_patch, custom, namespace) are degraded to function type; versioned web_search variants degrade to web_search.
• Web search support: Xiaomi natively supports the web_search tool type.
• Max 128 tools: Xiaomi accepts up to 128 tools per request.
• Cached tokens: Xiaomi reports prompt_tokens_details.cached_tokens.
• Reasoning tokens: Xiaomi reports completion_tokens_details.reasoning_tokens.

Hooks

xiaomiPatchRequest

Transforms the bridge's Chat Completions request for Xiaomi-specific behavior with three priority levels:

1. Historical reasoning content — When reasoning_content exists in message history, forces thinking: { type: "enabled" } to preserve multi-turn reasoning continuity.
2. Bridge-set thinking — When the bridge has already set thinking (via boolean effort mapping), preserves it unchanged.
3. Default disabled — When no reasoning was requested and no thinking field exists, sets thinking: { type: "disabled" }.

Additionally strips reasoning_effort (not supported by Xiaomi) and maps max_tokens → max_completion_tokens.

xiaomiStreamDeltas

Extracts typed deltas from Xiaomi's SSE chunks:

• content → text delta
• tool_calls → tool call deltas (via shared mapCommonChatStreamDelta)
• usage → usage delta (with cached token and reasoning token mapping)
• finish_reason → finish reason delta

mapXiaomiUsage

Maps Xiaomi's usage format to the standard ResponseUsage type, including:

• prompt_tokens_details.cached_tokens → input_tokens_details.cached_tokens
• completion_tokens_details.reasoning_tokens → output_tokens_details.reasoning_tokens

Both fields are validated with assertFiniteNumber — malformed values throw ProviderError.

Client Construction

The createXiaomiProviderEdge() function in client.ts builds the ProviderEdge:

1. Creates a ChatProviderClient with the provider's base URL, API key, and timeout.
2. Uses createProviderEdge() from the bridge kernel to wrap the spec, config, and HTTP methods.

Registration

The provider is registered in src/providers/builtin.ts:

export const XIAOMI_PROVIDER_DEFINITION = createProviderDefinition(
  XIAOMI_PROVIDER_NAME,
  createXiaomiProviderEdge,
);

The default Xiaomi base URL is https://api.xiaomimimo.com/v1. The default model is mimo-v2.5-pro. The API key is configured via the MIMO_API_KEY environment variable.

Content Handling

Xiaomi may return message.content as either a string or an array of content parts. The xiaomiOutputText accessor handles both cases, extracting text from type: "text" parts in array content. When content is null or an unrecognized type, it returns an empty string.

Models

Model ID	Description
mimo-v2.5-pro	Default model, most capable
mimo-v2.5	High quality model
mimo-v2.5-tts	Text-to-speech model
mimo-v2-pro	Previous generation pro model
mimo-v2-omni	Previous generation multimodal model
mimo-v2-flash	Fast inference model

MiniMax ReferenceZhipu ReferenceDeepSeek Reference