Azure GPT-5.4 requests use `max_tokens` on some paths, but the model requires `max_completion_tokens`

Description

Summary

When using opencode with liteLLM/azure/gpt-5.4, some requests fail with:

litellm.BadRequestError: AzureException BadRequestError - Unsupported parameter:     
'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.  
																				  
Short/simple prompts succeed, but longer / higher-context requests can fail.         
																				  
## Environment                                                                          
																				  
- opencode-ai version: 1.18.23                                                       
- platform: Windows x64                                                              
- model: liteLLM/azure/gpt-5.4                                                       
																				  
## Minimal observation                                                                  
																				  
This succeeds:                                                                       
																				  
opencode run -m "liteLLM/azure/gpt-5.4" --agent build --format json "Reply with      
exactly OK"                                                                          
																				  
But longer / higher-context requests can fail with the max_tokens vs                 
max_completion_tokens error above.                                                   
																				  
This also occurs outside of mission-driver / wrapper tooling, so it does not         
appear to be caused by an external runner.                                           
																				  
## Likely cause in source                                                               
																				  
From the current source tree:                                                        
																				  
- packages/llm/src/protocols/openai-chat.ts sets:                                    
max_tokens: generation?.maxTokens                                                  
- packages/llm/src/protocols/openai-responses.ts sets:                               
max_output_tokens: generation?.maxTokens                                           
- packages/llm/src/providers/azure.ts appears to choose between different request    
routes                                                                             
																				  
So it looks like some Azure GPT-5.4 calls are going through a chat/openai-           
compatible path that still emits max_tokens, while this model expects                
max_completion_tokens.                                                               
																				  
## Expected behavior                                                                    
																				  
For Azure GPT-5.4, opencode should avoid sending incompatible max_tokens and         
instead use the correct token field for the selected endpoint/model, or route the    
request through a compatible path.                                                   
																				  
## Relevant files                                                                       
																				  
- packages/llm/src/protocols/openai-chat.ts                                          
- packages/llm/src/protocols/openai-compatible-chat.ts                               
- packages/llm/src/protocols/openai-responses.ts                                     
- packages/llm/src/providers/azure.ts                                                
- packages/llm/src/providers/openai-compatible.ts    

### Plugins

_No response_

### OpenCode version

_No response_

### Steps to reproduce

_No response_

### Screenshot and/or share link

_No response_

### Operating System

_No response_

### Terminal

_No response_