Migrate an existing OpenAI Chat Completions integration to the Responses API: convert messages to input, read output_text instead of choices[0].message.content, and update tool and structured-output schemas

Documented steps

1. Change the endpoint from POST https://api.openai.com/v1/chat/completions to POST https://api.openai.com/v1/responses, keeping the same 'Authorization: Bearer <API_KEY>' and 'Content-Type: application/json' headers.
2. Replace the 'messages' array with 'input': pass a plain string for simple prompts, or an array of role/content Items for multi-turn context.
3. Move the system prompt out of the messages array into the top-level 'instructions' parameter.
4. Update response parsing: read the convenience field 'output_text' instead of choices[0].message.content. For structured access, iterate the typed 'output' array, whose items may be of type message, reasoning, function_call, function_call_output, web_search_call, etc.
5. Update tool definitions to the flat shape: {"type":"function","name":...,"description":...,"parameters":{...}} at the root, replacing Chat Completions' nested {"type":"function","function":{...}} wrapper.
6. Note that tools are strict by default on Responses; pass "strict": false per tool if you need the older non-strict validation behavior.
7. When returning a tool result, emit a 'function_call_output' item carrying 'call_id' to correlate with the originating 'function_call' item. Chat Completions used tool_call_id inside a role:tool message.
8. Move structured-output config from response_format:{type:'json_schema',json_schema:{...}} to text:{format:{type:'json_schema',name:...,schema:{...},strict:true}}.
9. Replace hand-written wrappers for hosted capabilities with native tool entries such as {"type":"web_search"}, {"type":"file_search"}, or {"type":"code_interpreter"} in the tools array.
10. Drop the 'n' parameter: multiple generations per request are not supported; issue separate requests instead.
11. Choose a state strategy: previous_response_id to chain server-stored responses, the Conversations API for durable state, or manual replay of prior output items with store:false.
12. Verify billing assumptions: store defaults to true and responses persist 30 days; set store:false for zero-data-retention environments.