Agent mode support

[sapphirepro]

Describe the need of your request

Please add agent mode support to allow AI self compile, find errors etc. That would speed up work a lot with much less manual prompt. Maybe also make clangd errors visible to AI directly.

Proposed solution

No response

Additional context

No response

[Samir-Toumi]

@carlrobertoh: When do you plan to release the agent mode?

[carlrobertoh]

The actual release is currently blocked by our cloud pricing model changes.

However, I will be publishing a pre-release version sometime early next week, and I'm looking for some feedback on the initial version. It can be used with all providers and models except our cloud one, as our existing plan isn't yet suitable for agentic use.

I will add a link here once ready.

[carlrobertoh]

Small sneak peek:

[sapphirepro]

Small sneak peek:

OMG! Woow! Awesome! Can't wait for release. And count me in as early tester. GitHub as myself I will not use, since commercial projects. But if it can access local files itself, watch errors and maybe self hosted Gitlab will be amazing! :)

[carlrobertoh]

Awesome! May I ask which provider and model you are using? Also, have you used or are you currently using any other agentic tools?

[sapphirepro]

Awesome! May I ask which provider and model you are using? Also, have you used or are you currently using any other agentic tools?

I use Grok and OpenAI. Grok for tiny stuff, OpenAI for serious things.

I have compiled version from github and MCP is awesome! Only thing is to consider for future, maybe to use something like ".aiignore" or something like original JetBrains AI does, to blacklist folders/files from access and exclude asking each step permission to read. Write permission makes huge sense through. (tested with current git, there is mcp, but no yet agent, but it's good as for now).

I was using JetBrains AI tool, but honestly it is trash. Tried a few extra and ended up with this plugin. It is very good!

What would be very needed is support of responses api for ChatGPT (I use Custom OpenAI). completions are less effective in terms of token caching and agent mode. But for that needed either JSON object support like reasoning : { "effort" : "medium" } or just a field to drop plain JSON pieces there.

[carlrobertoh]

Hey! As promised, I uploaded the bundled version here: https://github.com/carlrobertoh/ProxyAI/releases/tag/3.7.0-pre. I'll soon push the changes to master as well.

It's basically a deep agent with several built-in tools (you can prompt the agent to output which tools are available). It supports Anthropic, OpenAI, Custom OpenAI, Google, Mistral, and Ollama providers. It follows instructions from the PROXYAI.md file. At the moment, it lacks logic for context compression, history, rollbacks and a few other important features. You'll probably also notice some UI/UX-related issues, which are most likely known but not yet addressed (and who knows what else I might have broken).

FYI: This hasn't been tested by anyone before, so I can't promise it will work on the first try. If it works, use with caution! 😄

[sapphirepro]

Hey! As promised, I uploaded the bundled version here: https://github.com/carlrobertoh/ProxyAI/releases/tag/3.7.0-pre. I'll soon push the changes to master as well.

It's basically a deep agent with several built-in tools (you can prompt the agent to output which tools are available). It supports Anthropic, OpenAI, Custom OpenAI, Google, Mistral, and Ollama providers. It follows instructions from the PROXYAI.md file. At the moment, it lacks logic for context compression, history, rollbacks and a few other important features. You'll probably also notice some UI/UX-related issues, which are most likely known but not yet addressed (and who knows what else I might have broken).

FYI: This hasn't been tested by anyone before, so I can't promise it will work on the first try. If it works, use with caution! 😄

Thanks! Will be testing now :-)

[carlrobertoh]

Awesome! Please let me know if something isn't working or if you need some additional information. Also, I'll be updating the same bundled version over the week. I will ping here once I do that.

[sapphirepro]

Awesome! Please let me know if something isn't working or if you need some additional information. Also, I'll be updating the same bundled version over the week. I will ping here once I do that.

I am very sorry for late reply. Got sick. Now I tested new plugin and ready to give summary.

Now regarding plugin. In overall agent mode does something, but:
Model selector broken (uses incorrect model, supposed to use Custom Open AI - GPT 5.2, used some GPT 5.1 instead, config is correct, model use incorrect).
Remaining tokens counter went -40% which is weird. MCP tools in chat mode working, but chat unable to apply diffs.

Reading/writing files etc works, but result was not like expected, huge token growth. Critical is model selector broken and in chat mode unable to apply diffs at all. As a alpha version it is good progress! :-) But fixed required.

[sapphirepro]

Model selector broken (uses incorrect model, supposed to use Custom Open AI - GPT 5.2, used some GPT 5.1 instead, config is correct, model use incorrect).

Sorry, my bad, model use correct one. So agent basically able to perform read/write actions. What it strange in both Chat and Agent mode despite of provided files as context (attahed) it read them again itself. Result got a bit broken logic, but that's probably due to not using proxyai.md. Good part is that files it reads itself are all "cached" tokens.

Chat mode also reads all files even files provided, but as mentioned output fully broken, it is unable to apply diff/patch.

[sapphirepro]

Forgot to mention, that agent start windows has links for some configs, but none of it working (probably known issue).

And tokens counter needs a fix in both chat and agent mode. Those "%" is completely unclear what it is and how it counts, numbers would be preferred instead. GPT models have 400K overall (128K output, the rest input).

But in general good work👍👍👍

PS. For big project something like ".aiignore" file needed for filesystem MCP to prevent access to unwanted area, because too much files drives AI crazy, also it is useful to limit area where AI supposed to act

[sapphirepro]

After intensive usage, I came to amazing idea. This seems how it should work.

This is how it is now. This is good for non-MCP and non-agent modes. But to do so, that if mode=agent or mode=chat and MCP (file_system) attached, instead of posting context (which AI then ignores and still grabs itself), make a reference, meaning something like file://Libraries/some_header.h, file://Binaries/some_exec.cpp etc. So that instead of full file context AI gets list of file paths to watch as context. It would save both time and money for end users and that way is how Agent/MCP should be optimal for work. I mean chat mode same way if file server mcp enabled.

[carlrobertoh]

Many thanks for your feedback! I've compiled a new version which hopefully addresses some of the issues you mentioned, plus some new cool additions: https://github.com/carlrobertoh/ProxyAI/releases/tag/3.7.0-pre

Model selector broken (uses incorrect model, supposed to use Custom Open AI - GPT 5.2, used some GPT 5.1 instead, config is correct, model use incorrect).

There is still a pending issue where updating the model from the general settings isn't being reflected in the Agent toolwindow. Currently, the workaround is opening a new agent tab.

Remaining tokens counter went -40% which is weird. MCP tools in chat mode working, but chat unable to apply diffs.

Hopefully, this should be fixed in the new version attached above. The calculation is as follows: remaining tokens = model max size - max output size - current prompt size.

What it strange in both Chat and Agent mode despite of provided files as context (attahed) it read them again itself.

I'd advise not testing the tags right now. Manually added external context isn't being connected to the Agent (yet).

Good part is that files it reads itself are all "cached" tokens.

Can you give me more details on this one? I believe the provider caches the tokens themselves, because at this point ProxyAI doesn't handle caching on its own. I know that many providers cache input automatically, but not all of them (e.g., Anthropic).

Chat mode also reads all files even files provided, but as mentioned output fully broken, it is unable to apply diff/patch.

I will take a look at the Apply logic in more depth. I might have broken something with the new changes.

Forgot to mention, that agent start windows has links for some configs, but none of it working (probably known issue).

Links are still mocked but will be addressed in the final version i.e., support for configuring rules and subagents. This is still work in progress.

PS. For big project something like ".aiignore" file needed for filesystem MCP to prevent access to unwanted area, because too much files drives AI crazy, also it is useful to limit area where AI supposed to act

This will be handled via project-level rules .proxyai/settings.json. Again, something I'm still working on.

This is how it is now. This is good for non-MCP and non-agent modes. But to do so, that if mode=agent or mode=chat and MCP (file_system) attached, instead of posting context (which AI then ignores and still grabs itself), make a reference, meaning something like file://Libraries/some_header.h, file://Binaries/some_exec.cpp etc. So that instead of full file context AI gets list of file paths to watch as context. It would save both time and money for end users and that way is how Agent/MCP should be optimal for work. I mean chat mode same way if file server mcp enabled.

I generally like this idea. However, I think it has some negative effects on the initial prompt cache. The problem is that when you replace full file contents with file references, you're moving content out of the cacheable prompt and into the dynamic conversation flow. Because of that, the overall token efficiency might worsen (might be wrong here tho).