Skip to content
Notifications
Clear all

Switched from Cursor back to vanilla VS Code + Copilot. Miss the chat, not the bugs.

5 Posts
5 Users
0 Reactions
86 Views
(@data_shipper_joe)
Prominent Member
Joined: 6 months ago
Posts: 680
Topic starter   [#13313]

Hey folks, been lurking for a bit but wanted to share my recent experience. I've been using Cursor for about three months, drawn in by the promise of the agentic chat and better context handling. As a data engineer, I was hoping it'd be a game-changer for writing pipeline scripts and debugging connector code.

And honestly, I *really* miss the chat interface. The ability to just paste a chunk of Airbyte config or a problematic Fivetran log and have a conversation about it was fantastic. It felt like pairing with a junior dev who never sleeps. For example, troubleshooting a tricky `dbt` model with nested CTEs was so smooth in chat.

But... I had to switch back to vanilla VS Code with the GitHub Copilot extension. The bugs just became too much of a friction point for my daily work. The most frequent and frustrating one was the editor's context would seemingly "lose" my project structure. I'd ask it to "create a script to sync Salesforce contacts to Snowflake based on the existing pattern in `scripts/`" and it would generate code referencing non-existent modules or the wrong file paths. It felt like its awareness would randomly reset.

Another pain point was with code actions. It would suggest edits that looked correct in the diff preview, but applying them would sometimes corrupt the syntax, especially in YAML files for pipeline configurations. I ended up having to double-check every single change, which defeated the purpose.

Here's a tame example of a config it mangled:
```yaml
# Original
source:
postgres:
host: ${PG_HOST}
port: 5432

# Cursor's suggested "fix" (applied on action)
source:
postgres:
host: ${PG_HOST}
port: 5432 # Indentation broken, YAML becomes invalid
```

I'm back to Copilot now. The completions are less "conversational," but they're reliable. I don't have to second-guess if the editor is going to hallucinate my project layout. For the chat capability, I'm supplementing with the ChatGPT sidebar, but it's not as seamlessly integrated.

Curious if anyone else has hit similar stability issues, especially on larger codebases? I'd consider going back if a future version addresses some of this, because that chat experience was truly special for problem-solving.

ship it


ship it


   
Quote
(@juliap)
Estimable Member
Joined: 4 months ago
Posts: 100
 

I'm a senior technical procurement lead at a 250-person fintech, and our ~50 engineers deploy with vanilla VS Code + Copilot across all environments - I review the invoices and the support tickets.

**Real cost**: Cursor's Pro plan is $20/user/month on annual billing, and you pay for every seat. VS Code is free, and GitHub Copilot for Business is $19/user/month, but you get the entire MSFT ecosystem behind it (Teams, Azure, etc.). The hidden cost with Cursor was the time lost to bugs, which isn't on their pricing page.
**Integration friction**: Cursor required our team to adopt its fork of VS Code, which broke a few key extensions for our monitoring and internal linting tools. Vanilla VS Code + Copilot was a toggle in our existing setup. The migration back took about half a day per engineer to clean up settings.
**Stability & support**: When we trialed Cursor, we hit the same context-resetting bug OP mentioned, usually after 2-3 hours of sustained use. Support took 72 hours to respond with a boilerplate "we're working on it." Copilot has occasional hiccups, but we have a named Azure rep and issues are typically acknowledged within the SLA (4 business hours).
**Target audience fit**: Cursor's chat-centric model is brilliant for solo devs or small teams (15 that needs predictable, auditable tooling integrated into a CI/CD pipeline, the vanilla + Copilot combo is the boring, safe bet.

I'd pick vanilla VS Code + Copilot for any team where shipping reliably matters more than cutting-edge chat. If you're a solo contractor or in a research-focused role, Cursor's chat might be worth the bugs. To decide, tell us your team size and whether you're maintaining legacy code or building mostly new systems.


Your free trial ends today.


   
ReplyQuote
(@infra_switcher)
Reputable Member
Joined: 5 months ago
Posts: 320
 

You're hitting on the critical scaling point that individual users often miss: the support and integration tax. That "half a day per engineer to clean up settings" is a brutal, real cost that doesn't show up in any vendor's quote. It's lost velocity.

I've seen teams get locked into forked editors or niche tools that break their established workflows, exactly like your monitoring extensions. The minute your team can't run the internal linter or the custom observability plugin, you're blind. You're trading core capabilities for shiny features. The 72-hour support cycle on a blocking bug is a deal-breaker at scale; that's multiple engineer-weeks of wasted time waiting.

Your point about the MSFT ecosystem is practical, not just marketing. Having a path to a named rep and an SLA, even if it's just for Copilot, means the problem is now someone else's job to escalate. For a 50-engineer team, that's the difference between a managed service and a hobbyist tool. The chat interface is nice, but stability is the foundation.


Been there, migrated that


   
ReplyQuote
(@consultant_carl_42)
Reputable Member
Joined: 6 months ago
Posts: 381
 

Exactly. The vendor lock-in isn't just about the contract, it's about the workflow. You end up bending your entire development posture to accommodate the tool's quirks, and the cost of that posture change is invisible until you have to reverse it. I've watched teams standardize on a 'better' editor only to find their custom debugging scripts are now incompatible, turning a 10-minute check into a half-day archaeology project.

Your point about the SLA is the whole ballgame for any team bigger than a garage band. A "hobbyist tool" might have a brilliant feature, but when it breaks - and it will - you're just another username in a Discord channel. That 72-hour support cycle isn't an abstraction; it's your engineers twiddling their thumbs on day three because a critical autocomplete feature started hallucinating SQL syntax. The bigger the org, the more you're paying for the ticket number, not just the feature list.

And let's be honest, the chat is a parlor trick compared to stability. You can train a team on a slightly less magical workflow. You can't train them to work around a tool that randomly loses its context window or corrupts local settings. The shiny feature becomes the thing you apologize for in stand-up.


Test the migration.


   
ReplyQuote
(@eval_rookie_42)
Honorable Member
Joined: 8 months ago
Posts: 445
 

This is really clicking for me as I'm comparing CRMs for my team. That "workflow lock-in" cost is exactly what we're worried about. You don't see it on the demo.

When you say a "10-minute check into a half-day archaeology project," are you mostly talking about custom internal tooling breaking, or does it also happen with common plugins from the marketplace that the forked editor just can't run?



   
ReplyQuote