Duplicate code printed below correct code in same code block

[jeffaxup]

Type: Bug

• ask Gemini 2.5 Pro to generate new changes and to give the complete contents of the file with those changes.
• It gives the changes, but at the end it generates the filename, followed by a duplicate set of those changes. So if you copy/paste the code segment you have incorrect code.
• This has been happening for at least 6 mos or longer.
• There is a related problem where if Gemini 2.5 Pro has to edit multiple files at the same time, instead of putting each file in a separate code block it will put all the changes into one code block (along with comments, incorrectly formatted as code.)
• This is a serious ongoing issue regarding code quality, and it gets in the way of auto-applying changes accurately.
• I believe this is probably a Gemini 2.5 Pro specific issue, but this is the main model I use for everything.

Extension version: 0.35.3
VS Code version: Code 1.107.1 (Universal) (994fd12, 2025-12-17T14:15:14.850Z)
OS version: Darwin arm64 24.6.0
Modes:

System Info

Item	Value
CPUs	Apple M4 Max (16 x 2400)
GPU Status	2d_canvas: enabled direct_rendering_display_compositor: disabled_off_ok gpu_compositing: enabled multiple_raster_threads: enabled_on opengl: enabled_on rasterization: enabled raw_draw: disabled_off_ok skia_graphite: enabled_on trees_in_viz: disabled_off video_decode: enabled video_encode: enabled webgl: enabled webgl2: enabled webgpu: enabled webnn: disabled_off
Load (avg)	2, 2, 2
Memory (System)	64.00GB (0.35GB free)
Process Argv	--crash-reporter-id 5cad0c4e-247c-4a20-810a-cecab405b0ff
Screen Reader	no
VM	0%

A/B Experiments

vsliv368cf:30146710
binariesv615:30325510
nativeloc1:31344060
dwcopilot:31170013
dwoutputs:31242946
copilot_t_ci:31333650
e5gg6876:31282496
pythonrdcb7:31342333
6518g693:31436602
aj953862:31281341
6abeh943:31336334
envsdeactivate2:31353495
cloudbuttont:31379625
todos-1:31405332
3efgi100_wstrepl:31403338
use-responses-api:31390855
anonusage:31403420
je187915:31401257
3i15c187:31428712
ec5jj548:31422691
terminalsuggestenabled:31431119
cmp-ext-treat:31426748
cp_cls_c_966_ss:31442353
c0683394:31419495
478ah919:31426797
ge8j1254_inline_auto_hint_haiku:31431912
a5gib710:31434435
38bie571_auto:31440726
request_with_suggest:31435827
rename_enabled:31436409
7a04d226_do_not_restore_last_panel_session:31438103
anthropic_thinking_t:31432745
h0hdh950:31428394
9ab0j925:31438092
cp_cls_c_1081:31433293
copilot-nes-oct-trt:31432596
nes-jump-ctrl:31442063
nes-slash-models-on:31441832
nes-set-to-on:31441941

[jeffaxup]

[connor4312]

Gemini 2.5 Pro is not especially good at editing. Gemini 3 is much better. Please let me know if you repro the issue on Gemini 3, and share logs if so https://github.com/microsoft/vscode/wiki/Copilot-Issues#language-model-requests-and-responses

[jeffaxup]

I will try it and let you know.
I presumed that CoPilot has some capabilities to post-process what the model outputs in the chat window. i.e. it should be able to take the raw output and format it into different code boxes. It would also need to connect the target files listed in the output to the actual correct file in the IDE for applying changes. Is this not the case?

[jeffaxup]

Actually based off my first test of it Gemini 3 Pro looks pretty terrible.
It told me the edit the wrong file and didn't really seem to resolve the problem I was trying to fix. It also didn't format the result message properly (see image).
When I re-ran the same prompt with 2.5 it came up with an entirely different (and correct) answer, but once again, it is duplicating the code in nearly every response. Typically CoPilot seems to "post-process" this and only uses one version of the code to auto-apply the change to the file, but it is still very confusing to look at the duplicate code in the chat window, and sometimes it gets it wrong when auto-applying it.

[blazepsyki]

I have same issue.

In my case, these phenomena occurs during code explanation.
After explanation output, codes that I ask for explaining are printed but broken and duplicated.
(Some korean are included, but you will show this phenomenon.)

It is same with jeffaxup's case. If codes are not included, some blank lines are shown in the last paragraph.

In the debug log, all responses were normal. I attached the log.

panel_explain_eff7ff79.copilotmd

These phenomena only occur when using gemini series (include 2.5 pro and 3 pro, didn't test in gemini flash).
I wish this report will be helpful for solving problem.

[jeffaxup]

I think one of the sources of this problem is that it isn't formatting or understanding the difference between text that is a "message to the user" and text that is code (and should be in a code block). This results in everything being poorly formatted in the code block, and perhaps duplicate code that is not intended to be displayed to the user.
I think the CoPilot chat isn't parsing the response from the model correctly.

[Keavon]

I have been experiencing this problem for months (maybe half a year?) and got annoyed enough (and surprised it wasn't caught internally by Microsoft and fixed months ago) that I wrote up a detailed diagnosis that explains precisely what is happening: #286647. It is good to know that it happens for more people than just myself. Please see my issue for the added details it lays out.

[pwang347]

Hi folks, could you please upgrade to 0.36.0 for the copilot chat extension and check if this is still reproducing for you? A fix was checked in for Gemini models where content parts were being broken up by thinking parts, and it would be great to confirm if any cases above still reproduce. Thanks!