ryer.io

Dogfooding GitLab Duo Flows & GDK Debugging

#gitlab

TL;DR

  • Spent most of the day trying to dogfood flows created in YAML, following Agent 1495’s method of getting them into a GitLab project.
  • Agent 2397 reviewed but declined due to two 500-line flows in a single 1000-line file, suggesting building them in the UI instead for iteration.
  • Wanted local iteration capability, so set up GDK with Docker as the virtualization subsystem for GitLab runners, which consumed most of the day.
  • Reproduced the same error locally as on GitLab.com, then used a tool to inspect the job log directly and found it was a syntax error, which was resolved.
  • After fixing the syntax error, hit a new issue: job appears to succeed but shows a ‘model claude sonnet for six not found’ error, similar to a previous error, though it now progresses further than before.
  • Ran a gitlab-runner register command against the GDK test instance as part of this setup.

Flow creation and review pushback

I created a flow in YAML using Agent 1495’s method for getting them into the GitLab project. He recommended I have Agent 2397 check it out. Agent 2397 said he wasn’t going to review a thousand-line file containing two 500-line flows, and suggested building it in the UI instead so we could iterate on it more easily. I agreed to try the UI approach, but since I really want the ability to iterate on these locally, I decided to attempt as much of this on my own as possible, since that’s closer to how people will actually work with it.

Initial error and syntax issue

I created the flow inside the issue reproduce group, where I’m an owner and it’s enabled as an enterprise group. I kept hitting an error that seemed related to the project lacking the right GitLab Duo credits or a missing model, though it wasn’t clear exactly what the underlying problem was. It looked like it might actually be a syntax error, which was surprising. To get better visibility into what was happening, I decided to set things up locally in the GDK.

GDK setup with Docker

Setting up the GDK ate up most of my day. I installed Docker and configured the GDK to use it as the virtualization subsystem for GitLab runners, since runners are what ultimately execute the processes flows depend on. Once that was working, I was able to reproduce the same error I’d seen on GitLab.com.

Command used:

gitlab-runner register  --url http://gdk.test:3000  --token glrt-VsWRm1hq1MOM3HWFdEv262M6MQpvOjEKdDoxCnU6MQ8.01.170vez9mr

Diagnosing and current status

With everything local, I got access to a tool that let me look directly at the job log with access to internals, and it confirmed the issue was a syntax error, which we then resolved. Running it again after that, I ran into a new problem: the job appears to succeed, but amid a lot of debug and info output there’s a message saying the request model ‘claude sonnet for six’ isn’t found and no model is available. This is the same issue I ran into previously, though this time it gets further before failing. I’m not yet sure what’s causing it, but I’d like to get it resolved since it’s blocking me from dogfooding all of this properly.