400 Item 'XX_XX' of type 'reasoning' was provided without its required following item.

[VivekRepo1]

• Have you read the docs? Yes - Agents SDK docs
• Have you searched for related issues? Yes - searched for similar issues but couldn't find a match for this specific error. Most of the issue are around stateless use or suggest managing the conversation your self

Describe the bug

We're experiencing a random 400 error when using the OpenAI Agents SDK with OpenAIConversationsSession to store conversation context. The error message is:

400 Item 'XX_XX' of type 'reasoning' was provided without its required following item.

This error occurs intermittently between chat conversations, typically after several turns in a conversation.

The error suggests that a reasoning item is being included in the conversation input without its required following item (which should be an assistant response). This appears to be a validation error from the OpenAI API when processing the conversation history stored in OpenAIConversationsSession.

Debug information

• Agents SDK version: @openai/agents@^0.4.4
• Runtime environment: Node.js 22.19.0
• OpenAI SDK version: openai@^6.15.0
• Model settings: Using reasoning: { effort: 'medium', summary: 'concise' } in ModelSettings
• Streaming: Enabled (stream: true)
• Model: gpt-5-mini-2025-08-07

Repro steps

The issue occurs when:
Using OpenAIConversationsSession to persist conversation context across multiple turns

Here's a minimal reproduction scenario:

import { Agent, OpenAIConversationsSession, Runner, ModelSettings } from "@openai/agents";

const defaultModelSettings: ModelSettings = {
  parallelToolCalls: false,
  truncation: 'auto',
  store: true,
  reasoning: {
    effort: 'medium',
    summary: 'concise'
  },
  text: {
    verbosity: 'low'
  },
};

// Usage
const agent = new Agent({
  name: "TestAgent",
  prompt: { promptId: "your-prompt-id" },
  modelSettings: defaultModelSettings,
  tools: [/* User-specific tools that make DB calls */],
});

const runner = new Runner({
  traceId: "test-trace",
  groupId: "conversation-id",
});

// After several turns in a conversation
const session = new OpenAIConversationsSession({
  conversationId: "existing-conversation-id",
});

const result = await runner.run(agent, "User message", {
  session: session,
  context: {
    userId: "user-id",
  },
  stream: true,
});

// Error occurs randomly when processing the stream
for await (const value of result) {
  // Error: "400 Item 'XX_XX' of type 'reasoning' was provided without its required following item."
}

Key observations:

• The error is random and doesn't occur on every request

Additional context

Our implementation details:

• We're using OpenAIConversationsSession to persist conversation state across multiple turns
• We're using reasoning: { effort: 'medium', summary: 'concise' } in model settings
• Streaming is enabled (stream: true)
• We're using custom tools that make database calls based on user authentication:
  • Tools query user-specific data (routes, organizations, documents, etc.) based on the userId passed in the context parameter
  • The context includes userId.
  • These tools are called during agent execution and return data specific to the logged-in user
• The agent uses direct function tools that access our database

Questions:

1. What is the exact requirement for reasoning items? Must they always be immediately followed by an assistant response?
2. How should reasoning items be handled when using OpenAIConversationsSession? Should the SDK automatically ensure reasoning items are always paired with their required following items?
3. Could tool calls interactions affect how reasoning items are stored or retrieved from the conversation session?

Workaround

Currently, we're working around this by:

• Catching the error and retrying the request with a new conversationId (basically starting a fresh conversation)
• However, this is not ideal as it doesn't prevent the error from occurring

We would appreciate guidance on the correct way to handle reasoning items when using OpenAIConversationsSession for persistent conversation state.