Super User Daily: July 24, 2026

User Voice

The single most repeated ask is control over model selection. People do not want the harness quietly deciding for them: @SebastianRoehl wants a cheap model like Sonnet for trivial commit-and-push commands instead of burning the top model on them, @jpschroeder wants the auto-fallback that silently downgrades Opus 4.8 either turned off or at least made visible, and @emollick wants to hand-assign which model runs which type of task.

Quota anxiety has hardened into churn. @rezoundous and @bojie_li describe constantly refreshing the usage meter, @ecoleai says subscription cost is pushing them toward local models, and @mikiane is walking to Codex over how Fable 5 limits are handled. The cost conversation is now a retention problem, not a footnote.

Big jobs still fall off the context cliff. @svpino and @rohanpaul_ai both hit the wall on migrations and multi-service refactors that do not fit one agent's window, and worse, the tool cannot cleanly resume where it stopped; @VaibhavSisinty adds that it wastes 30-50K tokens re-exploring the same codebase at the start of every conversation.

Skills are a security surface nobody signed up for. @dr_cintas points out that a Claude Code skill is just executable code running with your full access, a real supply-chain risk with no vetting story, and @jsnnsa says the new Security plugin itself burned roughly $600 in tokens and bricked the session when it finished.

The agents also act with too little restraint. @VictorTaelin flags a destructive keybinding where the key that clears the input box also irreversibly kills every running agent and its context, while @nobi and @socialwithaayan describe agents over-consuming credits, making wrong assumptions, and editing unrelated code without being asked.