Curious about one thing: are you hitting the official DeepSeek API directly or using OpenRouter? Stability honestly depends on that more than the plugin. I’ve had issues with Continue for VS Code dropping the context window during long refactors, which was a huge letdown... but it's still the best for ghost text.

TL;DR: Continue is the top choice for completions, just watch your provider.

tbh I've tried many over the years and Continue extension for VS Code is highkey the winner for DeepSeek right now. It handles DeepSeek-V3 or R1 super well via API. It’s got that ghost text ur looking for plus actually solid RAG for indexing your codebase. I've found it way more stable than Roo with the context window... just plug in ur key and ur good. GL!!

Just sharing my experience: I went through this last month trying to ditch my $20 sub. I actually spent way too much time messing with RAG settings just to save a few cents on tokens lol... it’s definitely a balancing act.

Quick questions tho:
- Are you sticking to the official API or using a provider like OpenRouter?
- Whats your monthly budget for this?

Token costs can be a headache!

Jumping in here with a quick thought. To give you a better steer on which setup won't flake out on you after a week, I gotta ask a couple of things:

• Are you looking for something that handles RAG locally via vector embeddings on your own hardware, or are you cool with the extension managing that in the cloud?
• What is your typical project size like? Are we talking small microservices or a massive monorepo? The reason I ask is that stability usually comes down to how the extension manages the context window when the R1 model starts outputting massive chain-of-thought blocks. Some of the newer minimalist wrappers handle this way better than the big all-in-one plugins that try to do too much at once. Tbh, the ghost text performance depends heavily on whether you're hitting the DeepSeek API directly or using a traffic-managed provider.

Gonna try this over the weekend. Will report back if it works!

Regarding what #5 said about giving this a shot over the weekend... you are gonna have a blast! Just catching up on the thread now and it seems like the consensus is basically Continue for VS Code for those smooth completions and Roo Code Extension for the heavy logic. I’ve been doing the DIY route with my own keys for a while now and it’s honestly fantastic to finally ditch that $20 monthly sub. My advice is to stick with Continue for VS Code for your daily driving because the ghost text and indexing are super helpful for workflow. But definitely keep Roo Code Extension handy for when you need the R1 model to really think through a complex refactor! The way it handles the reasoning tokens is amazing compared to simpler wrappers. Also, pro tip: grab an OpenRouter API key as a backup for when the official DeepSeek API gets slammed... it happens sometimes during peak hours but the savings are totally worth the 5 minutes of setup tho!

Honestly, from a technical architecture standpoint, the market is splitting between simple 'wrappers' and 'agentic' frameworks. If you want the best DeepSeek integration, look at how the extension manages the context window—Roo Code is currently the leader for agentic tasks because it handles R1’s CoT (Chain of Thought) outputs without getting confused by the specialized tokens or formatting. For the RAG/indexing part, Aider is technically superior to most VS Code-specific plugins because it uses a 'repository map' to give the model a high-level view of your project's AST... basically prevents that mid-sentence cutoff by being way more efficient with token usage than the basic vector-search implementations. Most extensions are just competing on UI, but the ones optimizing the actual prompt-to-context ratio are where you’ll see the best stability with DeepSeek's API limits.