fix: encode empty tool arguments as JSON object instead of array

[MaximeWillinger]

Problem

When using multi-step tool calling with tools that have no parameters, the second request fails because empty tool arguments are encoded as a JSON array ("[]") instead of a JSON object ("{}").

This causes provider-side validation errors. For example, OpenRouter (which translates to Anthropic format internally) returns:

OpenRouter Bad Request: messages.1.content.0.tool_use.input: Input should be a valid dictionary

Reproduction

This happens in any multi-step tool calling scenario where a tool has no parameters:

$tool = Tool::as('get_current_time')
    ->for('Get the current server time')
    ->using(fn (): string => now()->toISOString());

$response = Prism::text()
    ->using(Provider::OpenRouter, 'anthropic/claude-sonnet-4')
    ->withTools([$tool])
    ->withMaxSteps(3)
    ->withPrompt('What time is it?')
    ->asStream();

When the model calls this tool during streaming, Prism's Stream handler:

1. Receives arguments: "{}" from the model
2. Decodes it with json_decode("{}", true) which returns [] (empty PHP array)
3. Stores it in the ToolCall value object
4. On the next step, MessageMap re-encodes it with json_encode([]) which produces "[]"

The provider then rejects "[]" because function arguments must be a JSON object, not an array.

Root cause

This is a PHP-specific ambiguity: an empty PHP array [] is both a valid list and a valid dictionary. json_encode([]) produces "[]" (JSON array), but the OpenAI API specification requires function arguments to always be a JSON object string.

Solution

Use json_encode($toolCall->arguments() ?: (object) []) to ensure empty arguments are encoded as "{}". When arguments are non-empty, the ?: operator short-circuits and behavior is unchanged.

Affected providers

The same pattern existed in all providers that encode tool call arguments in MessageMap:

• OpenRouter
• OpenAI
• Groq
• XAI
• DeepSeek
• Mistral

All six have been fixed with the same one-line change, along with a corresponding test for each.