Agent mode inserts entire file content at point

[timo-abele]

I am using Qwen-3 in Agent mode in IntelliJ and VS Code. I select a piece of code, hit Ctrl+J (L) for Chat and ask the agent to change something. The model generates its explanation and a preview of the changed file in the chat window. I click accept, so the changes are applied to the actual file. But now the entire preview file is inserted instead of just the differing lines:

Accepting this change then, of course breaks the build, because now I have the changed file inserted into the original.
Am I using continue wrong? I don't see how else to do this...

Edit: here is a recording of what I am doing:

And 'Edit' instead of 'Chat' gives the same result:

Is this a bug in continue? Or is the model somehow supplying the output wrong?

[bdougie]

I attempted to recreate and could not. Your screenshot looks likely to be Jetbrains and if so, we would be open to taking a contribution on this as we are no longer prioritizing Jetbrains and asking the community to help support it.

#8606

My attempt was in VScode.

[timo-abele]

Thank you for taking a look! In VS Code the behavior seems not as clear cut indeed. I had another look at VS Code, for edit I can reproduce something similar, the upper half (until the selected line) of a file is inserted:

For chat I cannot reproduce the behavior described above, but still something unexpected. I get two different results here:

1. the chat shows me the correct solution in a large preview block (the entire file), does not actually apply it, then claims to have applied it 🤔

2. the chat shows me the correct solution in a (appropriately) small preview block, but applies a raw explanation (the very first part of the model response, before the stuff we actually want? ) into the code, which breaks the build.