Fix xAI driver compatibility with OpenAI-compatible endpoints

[meirdick]

Problem

The xAI driver doesn't work with OpenAI-compatible /chat/completions endpoints like Cloudflare Workers AI's /compat endpoint. Two issues:

1. User messages always use array content format

MessageMap serializes all user messages as:

{"role": "user", "content": [{"type": "text", "text": "Hello"}]}

Some OpenAI-compatible endpoints only accept string content:

{"role": "user", "content": "Hello"}

The array format causes the model to receive garbled input and respond with nonsense, or reject the request with: "Type mismatch of '/messages/0/content', 'array' not in 'string'".

2. Structured handler crashes when content is returned as object

The structured handler expects choices.0.message.content to be a string (per the OpenAI spec), but some providers return it as a parsed object/array:

// Expected (xAI/OpenAI):
{"content": "{\"title\":\"Inception\"}", "parsed": {"title": "Inception"}}

// Actual (Workers AI):
{"content": {"title": "Inception"}}

This crashes on new AssistantMessage($content) with: TypeError: Argument #1 ($content) must be of type string, array given.

Fix

1. MessageMap: Use plain string content when no images are present, array format only when images exist. Both formats are valid per the OpenAI spec.

2. Structured handler: Normalize array content to a JSON string via json_encode(), and use it as the parsed data when no parsed field is present.

Both changes are backward-compatible with xAI's native API — verified by all existing tests passing.

Testing

• All 49 existing xAI tests pass
• Added new test: it('handles content returned as object instead of string') covering the Workers AI response format
• Updated MessageMapTest to expect string content for text-only messages

Context

Relates to #520. The xAI driver is commonly used with Cloudflare Workers AI (via the /compat endpoint) because it sends to /chat/completions rather than /responses. These two fixes make that combination work reliably for text generation and structured output.

[meirdick]

Note: The Text handler (Text.php:78, :155) also reads choices.0.message.content and passes it to AssistantMessage. In testing, text generation responses from OpenAI-compatible endpoints always return content as a string, so this isn't currently an issue. But if a provider ever returns array content for non-structured responses, it would hit the same TypeError. Happy to add the same is_array() guard there if you'd like — kept the PR minimal for now.

[SergkeiM]

Hi @meirdick as I understood current workaround is your package: /prism-workers-ai (thank you)

But is definetly nice to have this built-in in prism

[meirdick]

Hi @SergkeiM, yes that is correct. The package provides the full driver for complete worker-ai compatibility. It also makes for a clean automated implementation via this skill:
https://github.com/meirdick/laravel-cloudflare-ai-gateway