OpenAI Responses Adapter

​

Key features

• Model defaults: Reads the model name from the constructor (model=) or the OPENAI_MODEL environment variable. Override per message by passing model in the chat payload.
• Session context defaults: Creates AgentSessionContext(system_role=None) so system/developer prompts are driven by the caller or wrapper agent.
• Tool bundling: Converts the supplied toolkits into OpenAI tool definitions (both standard JSON function tools and OpenAI-native tools like computer_use_preview, web_search_preview, image_generation).
• Streaming support: Consumes the streaming response API, emitting events such as reasoning summaries, partial content, and tool call updates.
• Parallel tool calls: Optionally enables OpenAI’s parallel_tool_calls setting (disabled automatically for models that do not support it).
• Structured output: If output_schema is provided to create_response(), requests JSON schema output and validates the result.
• Automatic compaction: Uses OpenAI Responses auto-compaction (context_management) by default.

​

Constructor parameters

OpenAIResponsesAdapter(
    model: str = "gpt-5.2",
    parallel_tool_calls: Optional[bool] = None,
    client: Optional[AsyncOpenAI] = None,
    response_options: Optional[dict] = None,
    reasoning_effort: Optional[str] = None,
    provider: str = "openai",
    log_requests: bool = False,
    max_output_tokens: Optional[int] = 32000,
    context_management: Literal["auto", "standalone", "none"] = "auto",
    compaction_threshold: int = 200000,
    base_url: Optional[str] = None,
)

• model – default model name; can be overridden per message.
• parallel_tool_calls – request parallel tool execution when supported.
• client – reuse an existing AsyncOpenAI client; otherwise the adapter creates one via meshagent.openai.proxy.get_client.
• base_url – override the provider base URL used when the adapter creates its own client. Defaults to OPENAI_BASE_URL when omitted.
• response_options – extra parameters passed to responses.create.
• reasoning_effort – populates the Responses API reasoning options.
• provider – label emitted in telemetry and logs.
• log_requests – when true, logs HTTP requests for debugging.
• max_output_tokens – cap output tokens per response; also used when deciding whether to compact the context.
• context_management – controls compaction behavior:
  • auto attaches context_management to each request and lets Responses handle compaction.
  • standalone uses manual responses.compact preflight in the adapter.
  • none disables both auto and manual compaction.
• compaction_threshold – threshold used for compaction (compact_threshold in Responses context_management entries and manual preflight trigger in standalone mode).

​

Tool provider integration

• Image generation – ImageGenerationTool
• Shell execution – ShellTool
• MCP – MCPServer, MCPTool
• Web search preview – WebSearchTool
• File Search - FileSearchTool
• Code Interpreter - CodeInterpreterTool
• Reasoning - ReasoningTool

​

Handling a turn

1. Bundles tools - Collects the tools from your toolkits and packages them for OpenAI’s API
2. Calls the model - Sends messages and tools to OpenAI’s API
3. Handles responses - Processes text, tool calls, or structured output
4. Executes tools - When the model requests tools, executes them and formats results
5. Loops - Continues calling the model with tool results until it produces a final answer
6. Returns Result - Gives you the final output (text or structured data)

​

Context compaction

• context_management="standalone" to use manual responses.compact before a turn when usage crosses the threshold.
• context_management="none" to disable compaction management.

​

Related Topics

• Adapters Overview: Understand LLMAdapters and ToolResponseAdapters
• MeshAgent LLM Proxy: How requests are routed and metered inside rooms
• OpenAI Tool Response Adapter: How tool outputs are rendered back into the chat transcript.
• Process Agents Overview: Shows where the adapter is invoked in the recommended overall agent flow.