Regarding what #2 said about Roo code setting credits on fire, unfortunately, that is a common theme with most agentic tools right now. I have seen the same behavior where these extensions fail to leverage the deepseek-v3 prompt caching, leading to massive overhead. I have been testing a few others and honestly, Aider AI Pair Programming Tool CLI handles the architectural changes much better than most GUI extensions because it is more precise with its map-to-code logic. Another one to look at is Cursor AI Code Editor Desktop using your own API key. It is more polished than VS Code plugins, but even there, I have had issues where the deepseek integration feels slower than their native Claude implementation. To summarize the thread, you are basically choosing between a credit-hungry agent or a manual context manager. If you are on a budget, aider is the way to go because it is optimized for keeping token counts low, even if the learning curve is steeper.

@Reply #1 - good point! In my experience, Roo just set my credits on fire. Switched to Void Editor Open Source AI IDE because it lets you hand-pick context so DeepSeek-v3 wont loop.

> im mainly torn between continue and roo code. continue seems really solid for general chat and autocomplete but i keep hearing that roo code is way better for actual agentic tasks I've spent way too much time testing these and unfortunately, neither is the magic bullet people claim. I tried using Continue VS Code Extension v0.8 for a complex Next.js migration and it was a bit of a letdown. It struggled with the larger context window when I fed it multiple components, and the token usage spiked because it didn't seem to handle the prompt caching on DeepSeek V3 API as efficiently as I'd hoped. Basically wasted money on redundant input tokens. Then I moved to Roo Code VS Code Extension for the agentic stuff. Not as good as expected, honestly. I had it trying to fix a styling bug across three files and it got stuck in a weird loop re-reading the package.json for no reason. I lost about 2 million tokens in a few minutes because the agent just wouldn't quit. It was a total mess and pretty disappointing for a tool that's supposed to be smarter than a standard chat box. If you're trying to save every cent before Friday, maybe look at Aider AI Coding Tool via the terminal instead? It feels more precise with how it edits files compared to the VS Code wrappers. But if you must stay in the IDE, stick with Continue for the basic chat to stay safe. Agentic tools are just too risky when you're counting pennies, tho the workflow is nice when it actually works. Just keep an eye on that terminal output... I learned that the hard way.