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.

From the Backlog to the Build Log

Week of January 2 - January 8, 2026

What I Worked On

The honest headline this week: I pointed four different AI setups at the same job and none of them finished it cleanly. I ran four DevNitro CLI instances side by side, Claude Code on Opus, Claude Code on Sonnet, OpenAI Codex, and a local Ollama model, and had each one build the same site.

Here is how it went. Sonnet did the work but persisted the todo items in local storage instead of the database. Opus ran out of context because it tried to migrate and the schema information was too large. Codex hit its token limit inside the five-hour window. Ollama got stuck calling the same tool over and over. Four models, four different failure modes.

That is provider independence in practice. Supporting many models is not a checkbox, it is four sets of quirks you design around. The real lesson was that I did not have enough visibility into what the subagents were sending, which is why they burned so much context.

So most of the week went into subagents and testing. I built parallel subagent execution, cancellation that works even while a subagent is mid-run, and conversation filtering so the console shows the right agent's output. I also enforced read-before-write with persistent file tracking, so an agent cannot accidentally overwrite a file it has not read.

What the AI Did Well

When it worked, it really worked. I generated a WorkSmarterBook site, several game-backlog sites, and a complete site end to end, including running the database migrations and the tests. Sharing the dependency-injection setup between the CLI and the Visual Studio extension was a clean refactor. And updating the prompts with an explicit planning-first workflow made the agents stop diving into edits before they understood the task.

What Needed Correction

Beyond the four-provider mess, the subagent plumbing needed real work. Getting parallel subagents to display correctly in the chat UI, filtering their conversations, and making sure a cancel actually stops a running child took a couple of days. I also fixed file-locking on the local conversation files so reading them stopped blocking writes.

BlankSite needed cleanup too. I restructured it into feature folders, worked through Blazor persistence and an SSR-to-WASM waitlist admin, and chased a logout error and a dark-mode theme that would not apply without a reload.

What Shipped

DevNitro runs subagents in parallel, cancels them cleanly, and tracks read-before-write so it cannot clobber unread files. BlankSite is reorganized into feature folders with better docs and tokenized templates. Payments require Stripe now, and I removed the old Authorize.NET path from DevNitro itself (still available in GCT [and used by some client projects]). The cost-analysis tool is updated. And I ran a real multi-provider bake-off, even if the results were less than ideal. It was really about being able to have a bake-off in the first place.

I also read through Steve Yegge's Gas Town writeup on long-running agents. I started planning how to bring issue tracking directly into DevNitro, tied to sessions and conversations. That one is going to matter. I've been using Beads (also from Yegge) for a issue tracking for a while and I like it, but there are real issues with having a team utilize it with each developer having their own agents that need to claim issues. I decided to bring issue tracking into DevNitro itself, tied to sessions and conversations, so the agents can claim and work on issues without needing a separate tool.

The Numbers

Category	Hours
Product / Features	33h 18m
Testing	16h 29m
Content	9h 27m
Business	3h 54m
Community	2h 14m
Total	65h 22m

A big week: 33 hours of product, 16 of testing, plus content and a few hours refining the pitch and the cost analysis behind it.

Multi-provider AI agents sound like a feature and behave like four different products wearing the same interface. This week proved both the value of supporting all of them and the cost of pretending they behave the same way.

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

January 9, 2026 Build Log: Four Models, One Build