Fix incorrect formatting in OpenAI responses API for tool

[noorbhatia]

Fix #124

[Copilot]

Pull request overview

Fixes incorrect OpenAI Responses API request-body formatting for tool-call round trips in OpenAILanguageModel (Fix #124), aligning it with Responses API conventions rather than Chat Completions conventions.

Changes:

• Formats function_call_output.output as a direct value (plain string for .text, JSON array for .blocks) instead of JSON-stringifying “content blocks”.
• Flattens image_url in tool output blocks to a string URL (Responses format).
• Converts Chat Completions-style raw assistant tool_calls messages into Responses function_call input items.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[mattt]

@noorbhatia Thanks for your work on this. This isn't the first regression we've had with OpenAI request formatting, so I spent some time attempting to contrive a failing test case. Problem is, the API appears to be too flexible in what it expects (or else, I'm not clever enough to find a counter-example). I added what I came up with in c7ee6be. Until we fold in Replay and test the exact shape of request objects, this will have to do, I suppose.

[mattt]

This is now available in 0.7.1