• First test of Claude Sonnet 4.5 for an agent that backports a patch
• More reliable agents
• Backporting upstream patches with Code Assistants

There was one backport that served as “the final boss”: CVE-2025-59375 in expat. The upstream fix has 19 commits and changes almost 1000 lines. The models back then couldn’t handle this for Stream 9 where the drift in code was severe. I recall the backporting process halting several times with a message saying “this is too complex, a human developer with an IDE should do this work”.

It’s March 2026, let’s try to do this with Claude Code, imperfect prompt, and Opus 4.6.

Keep reading

For quite some time I’ve wanted to test how prone agentic tools are to prompt injection. Let’s go.

I’ll be using Claude Code 2.1.5, 4.5 Opus in various different sessions.

…

Keep reading

This is a followup to my previous post about how I tried Ambient Code.

TL;DR: I opened a PR (#specfile/508) and had to take over in the end.

Where did we end?

Claude Code proposed a solid implementation plan to us, Stella (the staff engineer agent) approved it.

Keep reading

Keep reading

One of the daily tasks we have when developing AI agents is to review their runs. We have to read dozens of decisions so we can evaluate if the agents did the right thing. If not, we have to adjust our user prompts, system prompts, and tools.

Let’s review how Sonnet 4.5 performs while backporting a complex patch (with multiple conflicts).

…

Keep reading

Over the last two weeks, we’ve spent time guiding our agents to perform more advanced workflows.

It was rough. For several days I was truly frustrated, because the results were atrocious.

…

Keep reading

This is a follow-up to my previous post about Claude Code.

We are building a tool that can backport upstream git-commits into CentOS Stream autonomously using AI coding assistants.

Keep reading

I am writing this blog post as Claude Code is working on upsint, a tool we worked on many years back. I haven’t touched upsint’s codebase for some time. It worked just fine all those years but recently I started getting 401 and 403 while creating pull requests, probably due to my API token expiring. I have never implemented any serious error handling in the tool so it was hard to diagnose the issue quickly:

requests.exceptions.RetryError: HTTPSConnectionPool(host='api.github.com',
port=443): Max retries exceeded with url: /repos/packit/ai-workflows/pulls
(Caused by ResponseError('too many 403 error responses'))

Keep reading

Log Detective service is live for more than two weeks now. Running an LLM inference server in production is a challenge.

We started with llama-cpp-python’s server initialy but switched over to llama-cpp server because of its parallel execution feature. I still need to benchmark it to see how much speedup we are getting.

This blog post highlights a few common challenges you might face when operating an inference server.

Keep reading

In Log Detective, we’re struggling with scalability right now. We are running an LLM serving service in the background using llama-cpp. Since users will interact with it, we need to make sure they’ll get a solid experience and won’t need to wait minutes to get an answer. Or even worse, see nasty errors.

What’s going to happen when 5, 15 or 10000 people try Log Detective service at the same time?

Let’s start the research.

…

Keep reading