I found this repo:
https://github.com/ClockZinc/vscode-copilot-chat-CN
It is an up to date fork of GHCP but with focus on support of local models and removal of the telemetry and github account requirements.
So no more "rate limited" local models, no requirement to send them your code anymore.
I wanted to make that myself, but this might just be the base we need to continue.
I had the 3 day evaluation of Qwen and Gemma models a while ago, that was quite interesting but it was still a "dry" test.
I did not really switch - the 200x price hike of June was more than a month away.
So I had my Pro+ license monthly reset a couple days ago, I'm at 3% Premium usage and after 5-6 prompts my weekly limit was at 80%. I did not want to use that up completely.
So I thought I'll give my Qwen agent a real-world try. This time PHP and C++ code, as well as very complex and nested CSS, javascript in a custom framework. (Millions of tokens of codebase).
I'm using a custom version of Qwen 27B, it's close to vanilla with removed safety boundaries.
Running in Q5 quantization and just 4 bit for the KV cache.
I am running this on a 5090 but I am running TWO agent slots (double the context) - I use ngram speculative decoding for a bit more performance.
---
First very positive shock:
So I used it to debug a really nasty problem on WSL linux, a very annoying issue with cmake cuda toolkit detection - it found the bug (a badly written sub detection algorithm that uses the location of a symlink instead of the actual binary) - it would have solved it in a minute if I had trusted it to execute the shell commands autonomously instead of waiting on each step (no regrets here).
That's at least Sonnet 4.5 level difficulty.
Second level:
So that was surprisingly good, I now let it refactor a C++ based complex custom scripting language into a PHP version. It produced a working PHP version. That's another very difficult task. Howver it did not refactor it properly, it invented a new version.
That's the biggest issue I found so far - it did not read the whole C++ file and deviated heavily from the original.
The result was so good that I didn't realize it for quite a while - still a real problem I encountered a few times.
Debugging:
I asked it to look into the framework, implement an automated way to connect to the remote server and investigate the data from that processed template.
It digged throught he framework, found a module that uses AUTH HASH based login - implemented that 1:1 into the templating modal for admin users, then used curl to test it, struggled a while with the return (I gave it a hint that there is a json version used by the frontend debugging modal) - found the json backend, got half a megabyte of json data back and analyzed it without pulling it all into context.
By request it followed up to document the new system with examples into the local readme.
All of that I'd normally have given to GPT 5.5 or Opus, or carefuly to Sonnet 4.6.
Third level:
Now I worked on the PHP framework and admin facing interface. I ran into an old bug that Opus 4.6 failed to solve in 4 attempts. I had given up, as it's just an inconvenience and didn't want to dig through the AI written javascripts and CSS.
In short: It's a interactive ajax populated diagnostic modal with 400kb of intricate data and various columns, it has a nested modal system for deeper information and some nested modals did not open up a second time. The javascript and css code is partly shared among different frontend parts - making it very difficult to see through.
The Qwen Agent identified the problem, fixed it in one single prompt and identified and fixed a surfacing second bug (the scroll parent modal location was saved but it had multiple independent scroll locations).
```
Actually, I'm realizing the scroll position might be on a different element than I thought. The diagnostic modal has a grid layout with .cycle-modal-columns inside .cycle-raw-modal, and I need to figure out which element is actually scrollable. When the nested modal closes and the innerHTML is restored, the scroll state could be lost if I'm not capturing it from the right element. I should trace through the CSS to see what's actually handling the overflow and scrolling.
```
It solved a bug Opus 4.6 failed to solve. And I asked that thing 3 or 4 times to fix it - each time it annoyed me - each time I postponed it while more important things are waiting.
My personal result
Local agents are not just a fallback - it solved bugs Opus didn't solve. It's faster than GPT 5 and Opus. I can run two sessions in parallel on a 5090 with high context.
All of this while NOT giving away all my data to a remote untrustworthy company - I've had not a single second thought giving it admin level hash keys.
The final endgame will be a mix, local agent for 90% of the work with the ability to call the best remote AI for dedicated help or as a expert subagent. That's something I'll work on at a later point.