From the Backlog to the Build Log

GPT-6.1 Terra vs Opus on a Real Bug: October 30 - November 5, 2026

What I Worked On

On Saturday I gave the same real bug to GPT-6.1 Terra and Opus 5.5, first in Codex and Claude Code and then in DevNitro. I wanted to see how each model handled an existing codebase, and how DevNitro compared with running each model in its own tool. Feeding what the runs showed straight back into DevNitro was this week's lesson in Intelligent Engineering.

The bug lived in my HowdyChad blog. After a fresh load, the first Next click showed "Page 2" with page 1's posts. A top comment on a big AI coding video asked for model tests on existing codebases, so I gave all four runs one 703-character prompt in copies of the repo at the same commit.

Like last week, it was a light week for DevNitro itself (four commits, just under four hours), and most of my time went to outreach, which started Tuesday.

What the AI Did Well

All four runs found the same fix, a one-boolean guard so the prerendered [PersistentState] result only gets used on the first load, and all four built cleanly. Only the Opus runs explained why. In .NET 10, the framework hands the restored value back on the next parameter update, even after the Blazor component sets it to null. Claude Code proved it by decompiling the framework, and DevNitro with Opus pulled the .NET source from GitHub. I committed the Claude Code version, the smallest change with the clearest comment.

DevNitro with Terra was the only run that proved the fix in a browser, clicking Next and Previous and checking the posts on each page.

Then I had DevNitro write up the four runs from the session logs and recommend changes. The biggest one was about how DevNitro runs Terra. Codex batches tool calls and reuses its prompt cache well, so DevNitro now drives the Codex app-server, the engine behind the Codex app, while every edit still goes through DevNitro's own tools. On the rerun, Terra fixed the bug with a 91.8% cache hit, in line with Codex alone at 91.4%.

Friday night the agent added prompt caching for OpenAI's newer models and xAI. I fixed caching for Claude in mid-October, but OpenAI and xAI work differently. OpenAI's automatic cache marker would have written the whole prompt to the cache again on every call, at a higher rate, and xAI only reuses whole messages that match from the start. The agent added a table of how each provider caches to DevNitro's agent docs, so a later session doesn't assume they all work alike.

What Needed Correction

The agent's first setup for the Codex app-server was read-only, so that every edit would have to go through DevNitro's tools. Codex then told the model the filesystem was read-only, so the model found the cause of the bug and only proposed a diff. The agent switched to a writable workspace with Codex's own shell, patch tool, and approvals turned off, and the next run made all its edits through DevNitro's tools.

What Shipped

Saturday night I published "GPT-6.1 Terra vs Opus 5.5 on a Real Bug in My Existing Codebase." The video covers Claude Code and Codex, and my AI video studio did the edit. Wednesday I published another video.

I committed the DevNitro changes but haven't released them. I'll test and deploy them once outreach is running.

Outreach Started

Go-to-market took 26 hours. Tuesday I finished the three campaigns, and the first emails went out. By Wednesday morning too many of them had bounced because one data source gave me bad addresses, so I paused everything and cleaned the lists. Thursday the next batch went out as plain text with open tracking off.

The Numbers

Category Hours
Business 26h 19m
Content 18h 36m
Testing 3h 56m
Total 48h 51m

DevNitro time came to 3h 56m, all logged as tests, close to last week's 3h 44m. Go-to-market rose from 22h 7m to 26h 19m, and 7h 14m of the content time went to the studio.

Where This Fits in Intelligent Engineering

The model is only part of the result. How many calls it makes, what it reuses from the cache, and whether anything proves the fix all depend on what surrounds it. Intelligent Engineering puts your rules, tools, and checks in one layer that every model works through. That's why DevNitro turned off Codex's own ways to edit files when it moved GPT-6.1 Terra onto the Codex app-server. Terra's edits now go through the same DevNitro tools that Opus uses, so the same rules apply whichever model does the work.

Once outreach is running, the plan for DevNitro is to make the app-server path the default for GPT-6.1 Terra. If you want to see DevNitro in action, grab a spot on my calendar and I'll give you the tour. If you want help getting more out of the AI tools your team already uses, I can help with that too.