Examples

Provider Switching

A pattern for using multiple model providers without rewriting product workflows.

Provider switching belongs at the model selection boundary, not throughout product workflow code. Runners should receive a model with the capabilities they need; they should not know every provider-specific option.

Scenario

Support chat uses OpenAI gpt-5.5 by default. Internal escalation uses Anthropic claude-opus-4.8. A request with a PDF must choose a model path that supports document file input before building the prompt.

Flow

Step	Boundary
create provider clients	application startup
map use cases to models	model registry
check capabilities before prompt build	runner
record selected provider/model	trace metadata
test each configured model path	provider smoke tests

Example

import { AnthropicClient } from "@anvia/anthropic";
import { OpenAIClient } from "@anvia/openai";

const openai = new OpenAIClient({
  apiKey: process.env.OPENAI_API_KEY,
});

const anthropic = new AnthropicClient({
  apiKey: process.env.ANTHROPIC_API_KEY,
});

export const models = {
  supportDefault: openai.completionModel("gpt-5.5"),
  documentCapable: openai.completionModel("gpt-5.5"),
  escalationJudge: anthropic.completionModel("claude-opus-4.8"),
};

export function selectSupportModel(input: SupportModelInput) {
  if (input.attachments.some((file) => file.mediaType === "application/pdf")) {
    return models.documentCapable;
  }

  if (input.channel === "internal-escalation") {
    return models.escalationJudge;
  }

  return models.supportDefault;
}

The runner does not change shape:

const model = selectSupportModel({
  channel: input.channel,
  attachments: input.attachments,
});

if (input.attachments.length > 0 && !model.capabilities.documentInput) {
  return { ok: false, error: "model_does_not_support_documents" };
}

const agent = createSupportAgent({
  model,
  user,
  services: input.services,
});

const response = await agent
  .prompt(input.message)
  .withTrace({
    name: "support-chat",
    userId: user.id,
    metadata: {
      provider: model.provider,
      model: model.defaultModel,
      channel: input.channel,
    },
  })
  .send();

Production Checks

Provider-specific params are set at model construction or request boundary, not scattered through business logic.
Capability checks happen before adding unsupported files, schemas, or tools.
Smoke tests cover streaming, tools, schemas, document input, and fallback behavior for each configured model id.
Traces record selected provider and model id.

Failure Modes