---

name: models-providers description: "Configure OpenAI Agents Python model settings, OpenAI providers, Responses websocket transport, third-party adapters, retries, and multi-provider routing." disable-model-invocation: true metadata: disco-role: operating license: MIT

Models & Providers

Use this sub-skill when configuring model selection, provider routing, OpenAI Responses vs Chat Completions, Responses websocket transport, ModelSettings, model retries, LiteLLM, any-llm, or OpenAI-compatible endpoints.

Read First

• Start with references/provider-reference.md for the model/provider classes, settings fields, environment variables, transport choices, and optional extras.
• Use references/workflows.md for copyable setup patterns: default OpenAI, websocket reuse, custom OpenAI-compatible endpoints, MultiProvider prefix modes, LiteLLM/any-llm, retries, and feature validation.
• Use references/troubleshooting.md for API key errors, wrong API path, websocket keepalive/message-size issues, Chat Completions feature gaps, namespaced model IDs, missing optional extras, and reasoning-content replay.
• Run scripts/inspect_model_provider.py for a local no-API-call import/configuration check that masks secret values and reports optional adapter availability.

Routing Boundaries

• Stay here for ModelSettings, OpenAIProvider, OpenAIResponsesModel, OpenAIChatCompletionsModel, responses_websocket_session(), MultiProvider, LiteLLM, any-llm, model retries, feature validation, and reasoning-content model adapter caveats.
• Route realtime model sessions, RealtimeAgent, RealtimeRunner, audio, and voice optional extras to ../realtime-voice/SKILL.md.
• Route runner loop behavior, handoffs, sessions, RunState, guardrails, tool execution, and turn limits to ../core-runtime/SKILL.md.
• Route tracing processors, trace export keys, trace metadata, and sensitive trace payload controls to ../tracing-observability/SKILL.md.

Fast Decisions

• Prefer the default OpenAI Responses path for OpenAI models and Responses-only tools.
• Use Chat Completions only for providers/endpoints that do not support Responses, and enable strict feature validation while migrating.
• Use responses_websocket_session() for multi-turn Responses websocket workflows that should reuse one websocket-capable provider/run config.
• Use MultiProvider(openai_prefix_mode="model_id", unknown_prefix_mode="model_id") when an OpenAI-compatible endpoint expects literal namespaced model IDs.
• Use LiteLLM or any-llm only when their adapter coverage/routing is needed; validate the exact upstream provider feature set before relying on structured outputs, tools, usage, or Responses-only behavior.