HISE Transformation to the new age

[Christoph Hart]

@Lindon we will not let Ralph loose on the HISE source code in the foreseeable future (or ever) - any kind of architectural change I tried in the last weeks had me code review every single line of code multiple times. Plus the ralph loop needs a validator step where it can check if it‘s done, and what would that be for complex bugs in HISE?

But you guys raise an interesting an realistic opportunity that has arisen with AI: the time that it takes you to wait for me fixing trivial to easy bugs should be gone once we setup a workflow that somehow leverages this distinction. I‘m thinking of this flow of events: 1. user finds a bug. 2. user starts a verified and sophisticated analysis process that evaluates the severity and complexity of the problem. 3a. if the complexity is below a threshold to be defined, user launches an agent to fix it. 4. user confirms that the bug is fixed and then makes a pull request. I‘ll review all prs regularly and merge the ones that I approve, the other ones will be discussed or rejected. 3b) if the complexity is too high to confidently exclude any side effect that fix could introduce then the agent will create a thorough analysis of the problem that then can be passed on to me in a format that let‘s me easily review and think about a proper solution.

The „easy route“ has been verified to be working effectively with the fixes that Dan did 2 months ago, so I think we just need to setup a formalized procedure that you can follow that ensures that I‘m not bombarded with PRs that I need to reevaluate each time for side effects, etc

[Lindon]

@Christoph-Hart This sounds like a very good approach. But what about the issue that I think we all need to get on a HISE commit that we can use as a basis for this "new world order". Perhaps the bug I pointed at which seems to fit in 3b, would be a good test case for you to try this new approach?

[Christoph Hart]

@Lindon let me cycle back and forth with claude about this idea, it‘s pretty good at „generalizing“ problems like this and come up with a formalized procedure. I can imagine that my terrible branch policy of stuffing everything in the develop branch might not survive this process ;)

I think what could help me is to look in the issues for open bugs and do a naive assessment yourself whether you would flag this as 3a or 3b, then just list them here. This could offer a valuable insight for the baseline of how much the analysis needs to do - if we agree there 100% of the time, then perfect, otherwise there needs to be some kind of AI assisted step.

[Lindon]

@Christoph-Hart Ok let me head off to GitHub then, how do you want me to report back?

[Christoph Hart]

@Lindon yup, just post 10 issues or so with a assessment of whether you would expect the AI to be able to perfectly oneshot the fix.

You can be sneaky AF and include issues that you like to be resolved first in this list, this kind of training data poisoning is harmless and well deserved…

[Lindon]

@Christoph-Hart said in HISE Transformation to the new age:

@Lindon yup, just post 10 issues or so with a assessment of whether you would expect the AI to be able to perfectly oneshot the fix.

You can be sneaky AF and include issues that you like to be resolved first in this list, this kind of training data poisoning is harmless and well deserved…

OK, well I'll def. include the one I care about most at the moment...

[Lindon]

OK here we go:

Easy-peasy one-shot Claude issues (perhaps):

https://github.com/christophhart/HISE/issues/869
https://github.com/christophhart/HISE/issues/739
https://github.com/christophhart/HISE/issues/762
https://github.com/christophhart/HISE/issues/761
https://github.com/christophhart/HISE/issues/625
https://github.com/christophhart/HISE/issues/863
https://github.com/christophhart/HISE/issues/769

Woah, smart brain of good looking German man(who like musical theatre) required:

https://forum.hise.audio/topic/14165/latest-version-of-hise-breaks-existing-presets/40?_=1771154413502
==== damn it I cant find the github issue number....possibly because there isnt one - so one of us is (to quote the Sex Pistols) a "Lazy Sod"...sorry

https://github.com/christophhart/HISE/issues/868
https://github.com/christophhart/HISE/issues/867
https://github.com/christophhart/HISE/issues/782
https://github.com/christophhart/HISE/issues/713

[Christoph Hart]

@Lindon yup good stuff, thanks. I'll try to explain Claude our plan, then give it this list of issues and let it categorize it into 3a or 3b without telling it your assessment. Then I'll do an assessment myself and we have three different evaluations per issue which should be a good enough "dataset" to distill a solution for step 2.

@David-Healey or anyone else reading this, you could also chime in with other findings of the bottomless pit of open issues from GH, the more assessments we have the better, it also expands the first column of the training data from Lindon's brain to a more generic "HISE developer" perspective - no pun intended Lindon :).

In the end we can also compare your assessments against mine and the AI suggestions and create a leaderboard in the who-can-take-over-if-I-get-hit-by-a-bus-category and a "who is actually a LLM disguised in a humanoid lifeform"-category...

[dannytaurus]

@Christoph-Hart When I fixed those bugs, and added small features or enhancements, I went for stuff that [1] had small surface area, [2] I was familiar with the feature, and [3] had been mentioned recently.

Bonus if there was a reproducible example included in the issue or a forum post.

It also helped that a lot of them were useful for my current work.

If something started to get complex or far-reaching I backed off and moved on to another issue. If I recall, some of the ScriptNode bugs were like this.

[dannytaurus]

@David-Healey said in HISE Transformation to the new age:

It's about time we got back to the master I think.

Agreed!

[Lindon]

@Christoph-Hart said in HISE Transformation to the new age:

"who is actually a LLM disguised in a humanoid lifeform"-category

oh I hope Im this.....!!

[Christoph Hart]

Alright, I just found out a feature of the CLI I'm using - whenever I said I use Claude Code I was lying. In fact I was using Opencode, which is the same thing but not tied to a certain model provider. Doesn't matter in the long run, but I just found a feature that we can somehow tie into that entire process: you can easily share the entire conversation history with the agent.

Eg, here's an example session of one of the earlier feature requests I added to HISE:

https://opncd.ai/share/aKO5yUQn

I can imagine that adding this link to a PR will help me track down the steps much better than just reading the final file changes, so for anything that isn't as trivial as this example this might be a very helpful piece of knowledge.

@dannytaurus I haven't touched Claude Code, but is there a similar feature over there in proprietary land?

[David Healey]

@Christoph-Hart said in HISE Transformation to the new age:

In fact I was using Opencode,

Which model are you using? I tried to use Claude with it but it seems anthropic are blocking it.

[Christoph Hart]

@David-Healey switched between all of the anthropic ones. I started using them through OpenRouter with the API key, but I blew through about 80USD worth of tokens within 2-3 hours, then went my way up the Claude subscription ladder up to the 200USD Claude Max subscription. It hasn't complained once about usage through a 3rd party client, but I also saw the headlines a few weeks back so maybe they just rolled this back to prevent backlash.

[David Healey]

@Christoph-Hart I get this message whenever I add a prompt

I'll check their github see if there is an issue there.

[Christoph Hart]

@David-Healey have you authorized opencode? Type opencode auth login, then select anthropic (there are two options, one for regular API usage and one for the plans), pick the one for the plan, paste the link in the browser, authorize it with your Claude subscription account, then paste back the stuff back into the terminal.