www.digitalmars.com         C & C++   DMDScript  

digitalmars.D.learn - opensource ai agents that airnt useless with d

reply monkyyy <crazymonkyyy gmail.com> writes:
The "zed" free trail is able to write a small d program by check 
if it runs and it will compile... but its a free trail and it 
auto ran formatting and its a fucking bloated ide, that makes my 
fans spin, that is half broken scaling. Etc.

So its probably time to start poking at the open source tooling 
and ripping out the stupidity and giving it the d docs directly.

Who's doing what? Whats handles d to some degree?
Oct 16
next sibling parent reply Sergey <kornburn yandex.ru> writes:
On Thursday, 16 October 2025 at 20:19:56 UTC, monkyyy wrote:

 So its probably time to start poking at the open source tooling 
 and ripping out the stupidity and giving it the d docs directly.

 Who's doing what? Whats handles d to some degree?


Hey monkyyy.

I've made a D port of Python project that is taking the whole 
github source and putting it into one huge formatted document, so 
AI agents will be able to use it effectively [1]

I think it is possible to use similar approach with Ddocs - 
generated for the project and directly shared into the 
context/system prompt.

I've seen similar idea was done by one rust dev [2], [3].
Especially as some Agentic CLI tools are becoming more popular - 
so no bloated IDE will be required in theory.

References:
1. https://github.com/cyrusmsk/rendergit-d
2. https://github.com/cortesi/ruskel
3. https://github.com/cortesi/dox
Oct 16
parent reply monkyyy <crazymonkyyy gmail.com> writes:
On Thursday, 16 October 2025 at 21:02:15 UTC, Sergey wrote:

 On Thursday, 16 October 2025 at 20:19:56 UTC, monkyyy wrote:

 So its probably time to start poking at the open source 
 tooling and ripping out the stupidity and giving it the d docs 
 directly.

 Who's doing what? Whats handles d to some degree?


 Hey monkyyy.

 I've made a D port of Python project that is taking the whole 
 github source and putting it into one huge formatted document, 
 so AI agents will be able to use it effectively [1]

 I think it is possible to use similar approach with Ddocs - 
 generated for the project and directly shared into the 
 context/system prompt.

 I've seen similar idea was done by one rust dev [2], [3].
 Especially as some Agentic CLI tools are becoming more popular 
 - so no bloated IDE will be required in theory.

 References:
 1. https://github.com/cyrusmsk/rendergit-d
 2. https://github.com/cortesi/ruskel
 3. https://github.com/cortesi/dox


why html, markdown exists?
and your not answering what ai models your using
Oct 16
parent Serg Gini <kornburn yandex.ru> writes:
On Thursday, 16 October 2025 at 23:01:14 UTC, monkyyy wrote:

 On Thursday, 16 October 2025 at 21:02:15 UTC, Sergey wrote:
 why html, markdown exists?
 and your not answering what ai models your using


you can use any model with this approach.
This is helping to add source code of your dependencies to the 
context and ask questions with this info.
I've tried Gemini from the big one.

For the local models and open source - I've played only with 
Llama3
But now there are better models are available. I've heard many 
good things about Qwen.

But they are popping up every week.. So most of the inference 
engines will be able to run most of them.
Oct 17
prev sibling next sibling parent reply jmh530 <john.michael.hall gmail.com> writes:
On Thursday, 16 October 2025 at 20:19:56 UTC, monkyyy wrote:

 The "zed" free trail is able to write a small d program by 
 check if it runs and it will compile... but its a free trail 
 and it auto ran formatting and its a fucking bloated ide, that 
 makes my fans spin, that is half broken scaling. Etc.

 So its probably time to start poking at the open source tooling 
 and ripping out the stupidity and giving it the d docs directly.

 Who's doing what? Whats handles d to some degree?


I want to create a custom GPT by uploading the D spec and a few 
other resources and see if that was any good. But haven’t had the 
time. My initial enthusiasm for ChatGPT 5 has waned a bit over 
the past few weeks since I’m getting similar kinds of mistakes I 
had on GPT 4 with certain types of questions.

The weird thing is that it’s so good at a lot of things, but is 
still dumb on others. Still gotta be careful.
Oct 16
parent Lance Bachmeier <no spam.net> writes:
On Friday, 17 October 2025 at 00:36:21 UTC, jmh530 wrote:


 The weird thing is that it’s so good at a lot of things, but is 
 still dumb on others. Still gotta be careful.


That will be true 30 years from now. I'm a big believer in their 
usefulness, but don't expect they'll ever stop surprising us. 
That's a good thing because it forces us to understand what's 
going on.
Oct 16
prev sibling next sibling parent reply Mark Davies <mark.davies trap10.net> writes:
On Thursday, 16 October 2025 at 20:19:56 UTC, monkyyy wrote:

 The "zed" free trail is able to write a small d program by 
 check if it runs and it will compile... but its a free trail 
 and it auto ran formatting and its a fucking bloated ide, that 
 makes my fans spin, that is half broken scaling. Etc.

 So its probably time to start poking at the open source tooling 
 and ripping out the stupidity and giving it the d docs directly.

 Who's doing what? Whats handles d to some degree?


I know it's not an agent but this is how I use AI.

I am using llama.cpp locally and I run

Qwen3-Coder-30B-A3B-Instruct-Q5_K_M
gpt-oss-20b-Q5_K_M.gguf
gpt-oss-20b-Q4_K_M.gguf
Qwen3-30B-A3B-Thinking-2507-Q6_K.gguf

For my uses, the Qwen models have been the best for D, I aim to 
try most models released on huggingface.

But most models do not know D well. It's easy for them to go off 
on a tangent and start saying the compiler is too old or the 
library is wrong.

Would be really nice to get a fine tuned model on the D language 
and library. Generating the documentation in plain text form to 
aid training would help.

I use codeblocks for editing and project management, fossil for 
source code control.

My workflow is to prepare the query in a codeblocks window and 
then paste into llama html query page. If I have a D module that 
works and I know it doesn't need modifying, I generate simple 
documentation (comments, function signatures) for that module and 
paste that instead of the source to reduce context use.

It takes seconds to copy the result paste into codeblocks and hit 
build.

I have projects generated that have more than 3k lines of code 
over multiple modules. Even if it's simple code the speed at 
which the AI can generate the code beats typing.

Regards,
Mark.
Oct 16
parent monkyyy <crazymonkyyy gmail.com> writes:
On Friday, 17 October 2025 at 06:38:34 UTC, Mark Davies wrote:

 
 Qwen3-Coder-30B-A3B-Instruct-Q5_K_M
 gpt-oss-20b-Q5_K_M.gguf
 gpt-oss-20b-Q4_K_M.gguf
 Qwen3-30B-A3B-Thinking-2507-Q6_K.gguf


qwen cli seems capable if standard verbose and over confident in 
stupid take
Oct 18
prev sibling parent reply Mark Davies <mark.davies trap10.net> writes:
On Thursday, 16 October 2025 at 20:19:56 UTC, monkyyy wrote:

 The "zed" free trail is able to write a small d program by 
 check if it runs and it will compile... but its a free trail 
 and it auto ran formatting and its a fucking bloated ide, that 
 makes my fans spin, that is half broken scaling. Etc.

 So its probably time to start poking at the open source tooling 
 and ripping out the stupidity and giving it the d docs directly.

 Who's doing what? Whats handles d to some degree?


I should have mentioned in my previous reply but one of the code 
blobs I generated is the core of an agent written in DLang, it 
has websocket interface to the html front end and openai 
interface to llama.cpp. It has tool calling support and tracks 
conversations. I've kind of put it on the back burner as I can't 
get tool calling working in the way that I want.

All the code was generated by AI models.

Mark.
Oct 17
parent reply Serg Gini <kornburn yandex.ru> writes:
On Friday, 17 October 2025 at 07:53:58 UTC, Mark Davies wrote:

 I should have mentioned in my previous reply but one of the 
 code blobs I generated is the core of an agent written in 
 DLang, it has websocket interface to the html front end and 
 openai interface to llama.cpp. It has tool calling support and 
 tracks conversations. I've kind of put it on the back burner as 
 I can't get tool calling working in the way that I want.

 All the code was generated by AI models.

 Mark.


Sounds nice.

How do you handle async tasks for agents?
Oct 17
parent Mark Davies <mark.davies trap10.net> writes:
On Friday, 17 October 2025 at 08:38:11 UTC, Serg Gini wrote:

 On Friday, 17 October 2025 at 07:53:58 UTC, Mark Davies wrote:

 I should have mentioned in my previous reply but one of the 
 code blobs I generated is the core of an agent written in 
 DLang, it has websocket interface to the html front end and 
 openai interface to llama.cpp. It has tool calling support and 
 tracks conversations. I've kind of put it on the back burner 
 as I can't get tool calling working in the way that I want.

 All the code was generated by AI models.

 Mark.


 Sounds nice.

 How do you handle async tasks for agents?


The websocket, web interface and openai interface run as separate 
threads with message passing by an internal mail system, tool 
calls are intercepted by the openai interface and external tasks 
are spawned as shell commands via pipe processes. There is a sub 
channel in the websocket messaging system for passing messages to 
and from the front end for task/internal status while text from 
the AI gets sent plain. This isn't fully complete as I was trying 
to resolve issues with tool calling before putting more effort 
into this area.

My plan was to write an internal built in editor for the AI that 
could read and write to internal buffers to reduce context use, 
having the ability to load a file into a buffer then scan and 
modify, then compile and run, capturing output in another buffer. 
This would allow context pruning while maintaining AI 
understanding of the task.

I can get it to use the shell interface quite well, I have had it 
scan a directory and analyse the files to generate what it thinks 
the files do. I can get it to use wget to retrieve information 
(bitcoin price, exchange rates etc.)

But I just can't get it to use the editor interface it gets 
totally confused. There was an issue with Qwen models and tool 
calling when using the openai streaming interface but they say 
they have fixed that. But it doesn't seemed to have helped in my 
case. I have put the code to one side until I can resolve this 
issue as that is the crux of what I wanted to do.
Oct 17