He Built an AI Software Factory on Kubernetes, and Issues Turn Directly into PRs
Piotr Mińkowski runs a self-hosted AI software factory on Kubernetes: GitHub issues or PR comments trigger Tekton pipelines; the agent in the pipeline calls a vLLM model inside the cluster, so code never leaves the cluster. The project is open source and comes with a video and architecture diagram.
Piotr Mińkowski posted a video on X: he built an AI software factory on Kubernetes. Open an issue on GitHub, or leave a comment in a PR, and Tekton will spin up a pipeline; an OpenCode agent runs inside that pipeline and, within the same cluster, calls a Qwen model hosted by vLLM. Code and issue text never leave the cluster—the only outbound call is pushing the PR back to GitHub.
The project is open source, and the repository is at github.com/piomin/ai-software-factory. The author drew an overview diagram:
The pipeline looks like this:
GitHub (issue, comment, review)
│ webhook
▼
OpenShift Route ──► EventListener (Tekton Triggers)
├─ github interceptor (HMAC signature check)
└─ cel interceptor (who / what / which event is allowed)
│
TriggerBinding ─► TriggerTemplate ─► PipelineRun
│
┌─────────────────────────────────────────┘
▼
print-event ─► fetch-skills ─► prepare ─► plan ─► agent ─► verify ─► publish
There are three pipelines in total:
issue-agent: Triggered by a new issue. It first posts a plan checklist comment, then implements it on a new branch, opens a PR, enables auto-merge, and finally replies with a summary.issue-comment-agent: Triggered by an issue comment that starts with@ai. If it is a question, it answers it; if code needs changing, it produces a plan and a branch.pr-comment-agent: Triggered by a review comment on a PR. It reads the entire PR conversation, then pushes changes to the PR branch or answers the question.
Trigger conditions are strict: only repository OWNERs and MEMBERs can trigger it; bot events are ignored; forks and closed issues/PRs are skipped outright.
On the technology choices, the author's reasoning is worth reading:
- Tekton: Kubernetes-native CI; Triggers can turn GitHub webhooks directly into PipelineRuns, with built-in signature verification and CEL filtering. Each run is an object that can be retried, has logs, and can be timed.
- vLLM: high-throughput inference; continuous batching lets multiple parallel agents share one GPU deployment.
- OpenCode: an open-source coding agent that supports any OpenAI-compatible endpoint and is not tied to a model vendor.
- Qwen: an open-source model that runs in your own cluster, has no per-token cost, and only requires a configuration change to switch models.
- GitHub App: short-lived installation tokens, so no personal access token is needed.
- OCI artifact distribution of skills: skills and subagents are packaged as OCI images and pulled from the same registry.
The security model is designed in considerable detail. The agent step has no GitHub credentials at all; even git commit and git push are disabled in the configuration. The plan step receives only a token valid for one hour and can only post gh issue comment. The publish step then signs a fresh token, because the agent may run for more than an hour.
The model is deployed in a separate ai namespace, uses NVFP4 quantization, and requires 2 GPUs, 180Gi of memory, and a 300Gi cache volume. By default replicas: 0; scale it up only when needed to avoid occupying the GPUs.
Below is a screenshot of a PipelineRun in the OpenShift console:

This project also sparked some discussion online. One user asked: why hand deterministic processes over to AI? Another user commented that the design of running the model in your own cluster is well suited to companies that do not want to send code to external APIs. But everything the agent knows comes from the repository and issues; it does not know that this is a payment service, nor whom to contact if the service has problems. That "SDLC context" layer has to be added separately. Another user shared a more complex approach called Chantier, which uses a "Chief of Staff" to orchestrate Change Requests and outputs a series of stacked PRs instead of a single PR.
Currently the repository has no license file; the author said in a reply that he will add one.
If you want to see it in action, the repository includes a showcase repo where you can see the bot's complete record of posting plans in issues, opening PRs, waiting for review, and then auto-merging. The README documents all the steps from installing GPU drivers to connecting webhooks; follow it and you can run it.
发布时间: 2026-11-21 03:56