Roché Compaan (@rochecompaan) header
Roché Compaan (@rochecompaan) profile picture

Turning ideas into software since the last century

Port Elizabeth, South Africa
Joined September 2010
Photo by Roché Compaan (@rochecompaan)Photo by Roché Compaan (@rochecompaan)Photo by Roché Compaan (@rochecompaan)Photo by Roché Compaan (@rochecompaan)Photo by Roché Compaan (@rochecompaan)Photo by Roché Compaan (@rochecompaan)Photo by Roché Compaan (@rochecompaan)Photo by Roché Compaan (@rochecompaan)Photo by Roché Compaan (@rochecompaan)Photo by Roché Compaan (@rochecompaan)
I just released Patchmill v1.0 It is a complete rebuild. The old workflow engine is gone, and a small kernel takes its place: it starts Pi, claims issues, reads git and the issue host, and checks what the agent actually did. Everything that needs judgment, such as how to plan, code, review, and write, now lives in skills you can read and change. The production code went from 53,542 lines to 4,300. On top of that kernel, this release adds two things we've wanted for a long time. You can now plan an issue with Patchmill in a chat, approve the spec from a dialog, and let a background worker build it. And you can run Patchmill without touching your tracker's labels at all. The skill pack also grew: frontend design with real HTML mockups, an independent frontend review, plain-English writing rules, and an implementation skill that always gets a fresh reviewer. This is a breaking release. There is no upgrade path from 0.20 configuration or saved state. Read Upgrading below before you run it on an existing repository. github.com/rochecompaan/patc…
1
2
69
Here's my vision for a Patchmill web interface, where developers and AI agents work on issues together, in the open. The idea is to let the whole team collaborate with agents transparently, using the same skills, models, and tools across our development process. I had Claude Opus 5.5 make this walkthrough video, and I was super impressed by how natural the Kokoro text-to-speech model sounds. If you want to help build this or just follow along, check out patchmill.dev.
45
Most agent harnesses handle a full context window by compacting it: summarize the old history and hope the summary kept what matters. It usually doesn't. Context paging does something simpler. The model sees a rolling window that always stays under a token budget, set below the "dumb zone". The full history stays on disk as exact text, and the agent can search it and load back exactly what it needs. No summaries, and no mid-task resets. It's now obvious to me that this is a great solution to context management, and I'm surprised more harnesses aren't doing it. Or maybe, for once, I'm just ahead of the curve! I published pi-context-paging a few days ago. It's experimental, but it's working surprisingly well, and I've just enabled it globally across all my Pi sessions. npmjs.com/package/@rochecomp… Curious if you experimented with this yet @pidotdev?
2
84
Imagine your coding agent never enters the dumb zone because its context is full. And imagine you never have to reset a productive session just to get a fresh context window. I've been experimenting with a fairly radical approach to this. Coding agents keep accumulating conversation until they eventually have to compact it. What if we do almost the opposite? Keep the working context tiny. Aggressively remove old conversation. Keep the complete history outside the context window and give the agent tools to search, browse and page exact pieces of it back in when needed. My experiment deliberately takes this to an extreme: the coding agent gets its current task state and only the last two model turns in working memory. Everything older is gone unless the agent goes looking for it. The experiment has progressed far enough that I've now reached a slightly more bonkers stage: I'm using pi-autoresearch to optimise the context policy itself. So instead of me deciding exactly how an agent's working memory should be structured, I'm giving another agent an experimental environment in which it can discover better ways to structure it. I've found related research around demand-paged LLM context, adaptive memory and self-improving agent harnesses, but I haven't yet found anyone applying an autoresearch-style optimisation loop specifically to context management like this. If you are doing something similar, or know of related work, please reach out! #CodingAgents #ContextEngineering #Autoresearch
3
1
53
Patchmill 0.20.0 improves progress reporting and terminal output. It also starts an experiment to determine which established code-quality checks can improve agent-produced code. We are dogfooding a broader quality framework against Patchmill itself: - Static analysis and strict type checking - Production coverage and CRAP scores - Property-based testing - Architecture checks - Mutation testing These are standalone checks in 0.20.0 while we study the results. The goal is to bring the useful checks into Patchmill issue runs, providing implementation and review agents with objective feedback as they work. This experiment was inspired in part by Uncle Bob’s work on Clean Code and his recent experiments with coding agents. Thanks to @mattpocockuk for unpacking those ideas in his interview with Uncle Bob. Documentation: patchmill.dev Release notes: github.com/rochecompaan/patc… #softwarefactory #aidevs #aidevelopment
34
You sandboxed your coding agent, right? Nice! Most developers I work with don't even get that far! Did you give it your SSH key so it could push to GitHub? Please don't do that. You know you'll regret it at some point. Just imagine it creating issues in a repo you don't own while sharing sensitive client IP or embarrassing you with slop. I'm working on a new project called RepoWolf to make it easy to manage agent access to GitHub locally, without the hassle of creating fine-grained access tokens for every repo and agent. RepoWolf keeps your GitHub credentials and SSH keys out of the agent sandbox. A broker you control holds the real credentials and enforces which repositories the agent is allowed to access. The sandbox gets only a scoped RepoWolf token. RepoWolf could use some real-world testing. Give it a try and let me know where you run into trouble, whether that's the setup, the docs, or the approach itself. github.com/rochecompaan/repo…
1
44
Love that I can just tell Pi to build me an extension to auto-handoff before we enter the dumb zone. This extension summarises the current state and automatically proceeds in a fresh session before we hit 150k tokens. github.com/rochecompaan/pi-c…
1
21
I’m seriously impressed by the amount of innovation in the new DeepSeek harness. This feels like the kind of release that lifts the game for every harness, open or closed. youtube.com/watch?v=1NyOG9z9…
1
92
New marathon PR with a 23 minute improvement over last year.
1
2
79
What started as an experiment might become a lifestyle change. I reached a hundred days of zero alcohol use exactly two weeks ago. The impact on my energy levels, quality of sleep, and resting heart rate has been significant! De-alcoholized wine doesn't taste too bad, either!
3
127
Do you have a ticket for PWC yet? Don't miss out on some awesome talks, including mine! I'm excited to share how you can use Kubernetes to level up your developer experience. 2023.pythonwebconf.com/
1
4
5
334
If Kubernetes can manage it for you, I see no point in paying Amazon 2 to 4 x times what a service is worth just because they manage it for you. Even their ec2 instances are ridiculously overpriced. A self-managed server with 4 CPUs and 16 GB RAM costs $50/month.
1
102
ec2 with the same spec costs $156/month. Linode is $120/month. Wait! Hetzner cloud instances with 8 CPUs and 16 GB RAM are $26/month and are now in the US! So a highly available k8s cluster with 3 nodes on Hetzner cloud will cost me $78/month!
1
2
110
I've been bumping my head against the infra-provisioning part of a Kubernetes cluster for a long time. Before you can be productive on an AWS EKS cluster, you need to provision the cluster itself, a service account with the appropriate policies, subnet groups, etc. #Kubernetes
2
1
123
"It uses the API directly so it's hilariously fast. While something like AWS Cloudformation might take you 10min to deploy and Terraform isn't that fast either, Crossplane will make the necessary changes in under 1 second.
1
73
It also constantly monitors and self-heals so configuration drift doesn't happen (someone goes and manually clicks shit in the console)." reddit.com/r/devops/comments… I LOVE the fact that self-healing snubs out configuration drift!!!
42