alt Hacker NewsSol and Fable are not doers, they are thinkers. Using them to "copy a few dozen HTML/JS/CSS snippets properly into a slightly obscure codebase, despite the reference being fully featured" is one of the most ineffective possible usage of those model. What you described is a job for luna.
Also I suspect you overprompt, with GPT-5.6 (sol, terra and luna), a lot's of what's used to be mandatory in the AGENT.md is redondant at best and frequantly actively harmful.
I also tried Luna (is my first line for such task in general because of the speed and efficiency, but not because Fable or Sol can’t be “doers”, in my evals both are superior in well scoped task adherence and implementation vs Luna and the like, have yet to find a task Luna can that Sol or Fable can’t) along with Opus 5 (cause the Max sub gives a lot of that I can freely use after Fable runs out for the week so I wasted three runs on that) and Ox Alpha/GLM-5.3 Flash (as that was free at the time). Sol (on High) and Fable (on Medium) were acts of 12h desperation, before I did it manually, after the three other models failed, first with a very simple prompt, failure, revert commits, then a very specific implementation plan with clearly scoped verification, failure, thus manual in 45 minutes. It was a major port of a new settings UX for a Firefox fork, all test baked, all with 1:1 replacement targets in the reference and the original. It was merely massive context wise and only 5.4 has ever behaved well across compactions in a manner that I could validate in evals. No model before or since gave me confidence in this regard and I’ll retroactively test with 5.4, somehow forgot to try that one.
I do not have an AGENTS.md in any of my repos at all, stopped using them around 5.2. I was a bit ahead of the curve with that, my evals very early showcased how limiting those tend to be, same why I never used MCP, were useless context bloat.
Will maybe make an edit explaining a bit further what the task was once I get to my computer.