From the Backlog to the Build Log

Week of September 12 - September 18, 2025

A note about this build log entry: I published this after the week it covers. DevNitro reconstructed a draft from work-related entries in my journal, Git commit history, and time tracking. I reviewed, edited, and approved the final post. The story reflects what I knew and built at the time.

This was a plumbing week. The GlobalCove Technologies Framework (GCT), the platform every DevNitro-generated app sits on, got a multi-database .NET architecture refactor so it can run on something other than SQL Server, and the NuGet packaging that ships it got automated, versioned, and obfuscated. Less glamorous than a tool call firing, but this is the layer the generated apps stand on.

What I Worked On

GCT was welded to SQL Server. If DevNitro is going to generate apps for anyone, the platform underneath cannot assume one database. So I added a DatabaseProvider setting that lives beside the connection string, introduced an IUnitOfWorkFactory, and moved the stores off IConfiguration and onto the factory. SQLite and PostgreSQL are stubbed in behind it. The T-SQL-specific SQL still has to be generalized, but the seam is cut.

The other half of the week was packaging. I automated NuGet package creation off the current git branch with a Directory.Build.targets script, fixed main versioning after it first spat out packages named with a branch-and-timestamp instead of 1.0.1, and rolled the new packages out across the consumer apps: DevNitro itself, BlankSite, my client sites, and a stack of smaller sites.

On the product side, I added lifecycle tools to the Workbench so Core can start and stop the app inside a container, with matching endpoints on the Workspace Agent, and wrote a README describing the architecture.

What the AI Did Well

The multi-database refactor is wide, mechanical work: change every store, rewire DI, keep the tests compiling. That is where AI-assisted .NET development earns its keep, and the bulk of it came out clean. I had @copilot run a peer review on the GCT pull request, which caught the kind of thing you skim past at hour six.

At the end of the week I had Claude Code dedupe and categorize a mess of open browser tabs and research links. Small task, but it cleared my head and my machine.

What Needed Correction

The packaging script bit me. After I merged Blazor back into main and pulled it down, the build produced NuGet packages named with a branch-and-date-time string instead of 1.0.1. I had to fix the pack-all script so main versions the way it should before the rollout could go anywhere.

The architecture question is still open. Right now one central LLM talks to each Docker instance. That may not hold. The alternatives are putting the model inside the container agent (Claude Code SDK style), or pushing messages onto a queue with several LLM processes pulling from it. I wrote the current design down knowing I may tear it up when the bottleneck shows up.

What Shipped

The GCT platform packages now ship as automated, versioned NuGet packages, and the consumer apps consume them. In prototype: the Workbench can control the app lifecycle inside a container. Nothing customer-facing, but the package supply chain the product depends on is real now.

The Numbers

Category Hours
Content 25h 16m
Product / Features 17h 40m
Community 1h 7m
Business 0h 56m
Total 44h 59m

Almost thirteen hours on NuGet in one week. That is the price of turning a pile of projects into a versioned package pipeline every generated app can pull from.

None of this shows up in a demo. It is the reason the demo will work later: one platform, many databases, shipped as packages the agent can install.

Last week Core and the container learned to talk. Two weeks back the first MCP tool call fired. Open thread carried forward: generalize the T-SQL so SQLite and PostgreSQL actually run, and decide where the model lives once one LLM per container stops scaling.

Curious how this site was built?
Watch the 5-minute walkthrough at devnitro.com →

An AI built this blog (but not this post)

This entire blog system (domain models, database, services, UI, auth, etc.) was built by DevNitro on its own in 95 minutes. I recorded every second.

    We respect your privacy. Unsubscribe at any time.

    Built with Kit