For anyone using LM Studio instead of OpenAI API. I had to drop using the OPenAI compatible API for LM Studio to get structured output working. Nothing would fail but it would not recognize the formatted output I specified. Doing this does require change a lot of the structure of payloads though. I used AI to help me do that quickly.

18:49 since I'm using LM Studio with a Gemma 4 model that has reasoning, mine actually already works for this multi-file use case lol. Since I never know when I will just get a response back or it will reason first. Though the first time I ran this and said "read the contents of package.json and composer.json " it got the package.json but failed to find the composer.json

Here is my code for that

while(true) {
            $prompt = text('Talk with Gemma!', required: true);

            $this->history[] = [
                'role' => 'user',
                'content' => $prompt
            ];

            ray($this->history);

            while(true) {
                $response = spin(
                    message: 'Hmmmmmm.....',
                    callback: fn () => $this->runModel()
                );

                $this->history = [...$this->history, ...$response['output']];

                $lastResponse = end($response['output']);

                if ($lastResponse['type'] == 'function_call') {
                    $tool = $lastResponse['name'];

                    if ($tool == 'get_current_time') {
                        $time = now()->toIso8601String();

                        ray($time);
                        $this->history[] = [
                            'type' => 'function_call_output',
                            'call_id' => $lastResponse['call_id'],
                            'output' => $time,
                        ];
                    }

                    if ($tool == 'read_file') {
                        $this->history[] = [
                            'type' => 'function_call_output',
                            'call_id' => $lastResponse['call_id'],
                            'output' => file_get_contents(
                                base_path(json_decode($lastResponse['arguments'])->path)
                            ),
                        ];
                    }
                } else {
                    info($lastResponse['content'][0]['text']);
                    break;
                }
            }

            ray($response)->purple();



            ray($this->history);
        }

Though of course Jeff's way is a lot cleaner that what I had :D

@willtomlinson think they realized that model was not going to work