API efficiency depends on context; without it, you're basically burning money. In my experience, Continue for VS Code handles DeepSeek-V3 reasoning tokens smoothly without editor lag. Just watch ur config!

Quick question - before I give advice, are you mainly worried about the monthly API costs or are you more focused on the initial setup difficulty? I ask cuz some of these take a bit of tinkering to get the FIM working right, especially with DeepSeek-V3. Anyway, adding my two cents... I've been super happy with my current setup lately. Honestly, moving away from Copilot saved me a bunch of money already. Basically, I'm using: - Continue extension for VS Code: It's open-source and free, so you only pay for what you use via the API. It handles those reasoning tokens much better now without lagging the whole UI.
- Roo Code extension: This one is highkey the winner for folder context. It actually "sees" the structure instead of just the file you're in.
- DeepSeek-R1 via Continue: I've found that setting this up feels just as snappy as Copilot once you've got the config right. I'm still kinda new to this but I've had no complaints so far! It works well and my monthly bill is way lower. What's ur budget looking like? Cheers!

Big if true

Huh interesting. I had no idea. The more you know I guess 🤷

Tbh, I’m gonna have to go against the grain a bit here. Everyone’s talking about Continue, which is cool and all, but if you really like the DIY approach and want the model to actually *do* stuff rather than just chat, you should check out Roo Code. I’m still pretty new to the whole agentic thing, but it feels way more powerful for Python/TS projects because it can actually read your whole folder and execute commands. Comparison-wise:
- Continue is basically like a better, cheaper version of Copilot. It's simple and snappy, but mostly stays in its lane as a side-panel assistant.
- Roo Code is more of a 'power user' DIY tool. It uses those DeepSeek-R1 reasoning tokens to actually plan out changes across multiple files. It’s a bit more 'manual' to set up, but way more rewarding!!! Wait, has anyone noticed if the API latency is better on one versus the other? I’m still trying to get my head around the 'Fill-in-the-middle' settings... it’s kinda confusing for a beginner like me lol. But yeah, if you want that DIY feel and total control, maybe skip the basic extensions and try the agentic ones?

> I’m looking for something that supports robust context-sharing—I want the AI to actually understand my folder structure and not just the file I have open. Yeah, I've been tweaking my DeepSeek setup for months now - basically since the V2 days - and honestly, long-term reliability is the real hurdle. I think IIRC there was a specific issue where indexing large monorepos would basically hang the UI once those heavy reasoning tokens started flowing. I’m always a bit cautious about how much "full project context" I let these tools grab – maybe it's just me, but I've had a few close calls with memory leaks during long dev sessions. I’m not 100% sure if the latest extension updates have completely solved that "reasoning lag" yet, but in my experience, keeping the setup as lean as possible is usually safer. Sometimes the ghost text gets out of sync with the actual API response if the latency spikes, which can be super annoying when you're in the zone. Tbh, I'd suggest just testing one setup at a time for a few days to see how your RAM holds up before fully committing. Not sure if that helps, but stability > features for me every time.

Can vouch for this

Honestly, I'd suggest an extension that lets you toggle 'stream' mode for those reasoning models. My current setup stopped the editor lag cuz it handles the 'thinking' in a separate panel. It's way more stable than the default stuff! Also, enabling 'FIM' support made the ghost text feel like Copilot but way cheaper. It actually understands my whole project structure now too lol. gl!

Interested in this too

Noted!