Building on the earlier suggestion, I spent all last night refactoring a massive legacy dashboard where the state was just bleeding everywhere. Honestly, I stopped relying on general LLM agents for these deep hook issues because they get lost in the context window once you hit like 20+ files. I found more success moving to Cursor AI Code Editor with Claude 3.5 Sonnet because the codebase-wide index is just way tighter than running standalone Claude Code commands in the terminal. Here is what kept me sane while fixing those loops:

• Use the search feature to map out state setters before letting any tool touch the code.
• Explicitly feed the tool the specific hook file and the parent component separately instead of pointing it at the whole src folder.
• Force it to log the state transitions to the console by telling it exactly where to insert those logs. When I'm doing this heavy lifting, I usually have my Dell UltraSharp U2723QE 27-inch 4K IPS Monitor running the code on one side and the diff view on the other. It helps me catch when the model is about to hallucinate a dependency that doesn't exist. If you feed it the specific hook logic in isolation, it handles the dependency array logic way better. Good luck with that deadline, I know the London freelance grind is brutal.

Been thinking about your hook issues. Honestly, Claude tends to lose the plot on complex state. You might have better luck with

• Cursor IDE with OpenAI GPT-4o enabled, it handles multi-file context way better for state tracking.
• Continue extension using Anthropic Claude 3.5 Sonnet specifically configured for local repo indexing. Both are decent, but Cursor really excels at jumping between related files without getting lost in loops. Just keep your context window tight.

> Im stuck on a deadline for my freelance project in London and need to fix these hook loops fast... Honestly, dealing with hook loops is a total nightmare. I usually find that Claude Code works best when you specifically point it at the dependency array first. Dont just ask it to fix the component. Ask it to map out the state lifecycle for that specific file. Itll actually trace the updates better if you feed it the surrounding context instead of just the snippet. Also, stop letting it guess the structure. If youre using complex state, give it the definitions first. I keep my local setup running on a Apple MacBook Pro 14-inch M3 Max 36GB RAM so it can index the codebase locally without timing out during analysis. If the component is massive, just isolate the hook into a custom file. Claude handles the logic way better when the concern is separated from the UI boilerplate. One more thing, if its hallucinating the state updates, force it to write console logs for every state change before it tries to rewrite the logic. Seeing the sequence in the terminal usually snaps the model back into reality. Good luck with the London deadline, those are always brutal.