Question
Hey,
i used opencode with different open source models and i found out, that agentic coding often does not understand the big (or small) picture, which results in dump mistakes or coding errors.
For example :
-
1st example: the coding agent was fixing a PHP unittest file from a laravel application. The fix was done and the php tests did nothing. The reason: opencode added 2 methods with the same name in one php file.
-
2nd example. The command for opencode was to safely remove one methode from a larger PHP file. The agents removes the method and also the imports used by the method (imported php moduels, that was called on the top of the file using usestatement). The agent did not realize, that many other methods also needed those imports, that was removed - so the application was broken.
I had different coding-mistakes like that, but they never happened with claude code. So first of all i thought about a bad or poor model in background. So i tried different models, like Qwen 3.5 122B, MiniMax M2 or GLM 4.7.
Than I found out, that the agents often only reads the files partial, to find out the important part. But by reading only a part of a file, the model will never be able to understand the big picture.
So it now looks like, that this is not a question of the right model, it seams more to be a question of: how agent uses the "read-file"-tool. In this case, this would be an option for an opnencode improvement, right?
Does anyone has the same problems?
Thank you, B.
Question
Hey,
i used opencode with different open source models and i found out, that agentic coding often does not understand the big (or small) picture, which results in dump mistakes or coding errors.
For example :
1st example: the coding agent was fixing a PHP unittest file from a laravel application. The fix was done and the php tests did nothing. The reason: opencode added 2 methods with the same name in one php file.
2nd example. The command for opencode was to safely remove one methode from a larger PHP file. The agents removes the method and also the imports used by the method (imported php moduels, that was called on the top of the file using
usestatement). The agent did not realize, that many other methods also needed those imports, that was removed - so the application was broken.I had different coding-mistakes like that, but they never happened with
claude code. So first of all i thought about a bad or poor model in background. So i tried different models, like Qwen 3.5 122B, MiniMax M2 or GLM 4.7.Than I found out, that the agents often only reads the files partial, to find out the important part. But by reading only a part of a file, the model will never be able to understand the big picture.
So it now looks like, that this is not a question of the right model, it seams more to be a question of: how agent uses the "read-file"-tool. In this case, this would be an option for an opnencode improvement, right?
Does anyone has the same problems?
Thank you, B.