xAI Responses adapter

Unreleased source addition after 0.1.0. The published @agenticdriver/sdk@0.1.0 package uses Chat Completions through xai(). Use a reviewed development build for the new xaiResponses() factory. No live xAI account or model has been selected for certification.

xAI's Responses endpoint supports function calls, streamed text and encrypted reasoning continuation. Its storage defaults to enabled; this adapter explicitly sends store: false and carries the returned output items into later tool steps. It does not use server-stored response IDs. Responses reference, local continuation and storage.

Configure an explicit API account#

import { AgenticDriver } from "@agenticdriver/sdk";
import { xaiResponses } from "@agenticdriver/sdk/providers";

const driver = new AgenticDriver({
  providers: [
    xaiResponses({
      id: "company-grok",
      apiKey: () => readSecret("company-xai-key"),
      models: [selectedModel],
    }),
  ],
  usage: {
    hostId: "company-driver",
    accounts: { "company-grok": "company-xai-account" },
  },
});

readSecret and selectedModel are supplied by the application. The default endpoint is https://api.x.ai/v1/responses; an operator can configure a different HTTPS base URL. Model IDs and account bindings are explicit. Catalog discovery uses the selected endpoint's model route without generation.

For the development host CLI:

agenticdriver init --config ./driver/config.json \
  --provider xai-responses --provider-id company-grok \
  --account-id company-xai-account --model YOUR_SELECTED_MODEL

The default key reference is XAI_API_KEY; configuration stores that environment variable's name, never its value. Existing host configurations can opt in by changing kind: "xai" to kind: "xai-responses". Preserve id, accountId, models and apiKeyRef only while they still refer to the same selected account and model. Start a new conversation when changing endpoint dialects; private Chat Completions state is not a Responses continuation record.

The old xai() factory and xai host kind retain their existing route. There is no automatic endpoint, provider, model or billing fallback.

Execution and evidence#

• Only selected application function tools are included. Tool arguments still pass schema validation, host authorization and approval before execution. Provider-hosted search, code execution, remote MCP and other server tools are not enabled by this adapter.
• Text deltas stream before the terminal response. Reasoning/tool events can report actual progress without exposing reasoning as visible text. Both run and inactivity timeouts remain disabled by default; caller cancellation aborts the provider request.
• Input, output, cached-input and reasoning token counts map to the existing account-scoped usage record when reported. Missing measurements remain unknown. Existing Usagestat identity and storage contracts are reused.
• Encrypted reasoning stays private during tool steps and in explicitly enabled native sessions. Public results, exported session history and usage records do not contain that state.
• HTTP auth/rate failures, provider error events, malformed usage and unfinished streams fail through normalized SDK errors. Truncated tool calls do not execute.

The adapter shares the tested Responses parser with OpenAI while keeping its own vendor, endpoint, usage identity and function declaration format. Its explicit media policy currently excludes direct PDF input; PDF retrieval can use the existing ingestion and source-scoped context workflow.

Historical parser checks are not current live qualification. The old simulated endpoint was removed. AD-020 remains open until an explicitly selected xAI account/model passes real checks. The Grok Build subscription adapter has separate native isolation requirements.