In my experience, context loss is a mess. Try Aider with DeepSeek-V2.5 API:
- Pennies per task
- Safer local code It actually handles deep refactors! gl

Sooo I've been thinking about your post for a bit because I went through this last year. Story time: I'm still kinda new to the deeper side of microservices, but I was recently thrown into a project that was honestly a mess of legacy code. I was terrified of breaking the whole infra, so I was being super cautious about which AI to trust. I didnt want to be the one who took down the entire staging environment!! Just sharing my experience: I ended up trying a few things. First was Sourcegraph Cody Pro, and later I experimented with a local setup using the Anthropic Claude 3.5 Sonnet API through a command-line tool called Aider AI pair programmer. Comparing them was realy interesting. Sourcegraph Cody Pro felt like a safe, professional choice cuz it realy focuses on that repository-wide context you're talking about. It helped me map out how a change in our auth service would ripple through the billing system without me having to manually feed it every file. On the other hand, the Aider AI pair programmer setup felt more powerful for actually executing the refactors across multiple files at once, but it felt a bit more "wild west" and I had to double-check everything twice so I didnt cause a total meltdown lol. I was realy satisfied with how Cody handled the architectural questions tho. It didnt hallucinate nearly as much as the standard web chats I used before. It works well for someone like me who wants to move slow and not break things... it's been a real life saver. Anyway, that was just my journey with it... good luck with the migration!! peace

> Any updates on this? @Reply #7 - good point! It seems like the thread has mostly covered indexing tools like Cursor, but if you're doing a massive migration, you gotta be extra careful about reliability. In my experience, even the best context windows can fail if the model gets lazy or loses the thread during a deep refactor. A few different setups stand out for heavy lifting lately:

• Continue.dev IDE Extension: This is my favorite for safety. Pairing it with DeepSeek Coder V2 works well for complex logic. Pros: You have full control over what code is shared and which files are indexed. Cons: It requires more manual effort to manage the context than an auto-indexer.
• GitHub Copilot Workspace: This is more of an agentic approach that thinks in terms of entire pull requests. Pros: It actually tries to understand the whole issue and creates a plan before writing any code. Cons: It can feel a bit like a black box, which makes me nervous when dealing with mission-critical legacy infra. Honestly, when you're dealing with deep call stacks and microservices, no AI is gonna be perfect. Ive found that keeping a human in the loop with a more manual tool like Continue is safer than trusting a fully automated agent... just my two cents tho.

In my experience, Cursor AI Code Editor is AMAZING for $20!! it basically indexes your whole repo which is way better than chat... i love it sooo much!

Re: "Exactly what I was thinking" - basically, the context window is the whole game here! Everyone has made some fantastic points so far about indexing and keeping things local. Saw this earlier but just now responding. If youre doing heavy architectural stuff, these two have been absolute game changers for my workflow and offer a different vibe than Cursor:

• Windsurf Editor by Codeium. I love it!! It uses their Flow feature which is just incredible for following those deep call stacks you mentioned. It feels way more proactive about understanding how a change in one microservice hits the others compared to standard tools. Its fantastic for big picture stuff.
• Continue.dev IDE Extension paired with Mistral Large 2. This combo is amazing because you can custom-tune the context providers. Mistral Large 2 has been fantastic for reasoning through complex logic without hallucinating as much as GPT-4o often does when the codebase gets messy. Honestly, if youre doing a massive migration, look into a more DIY setup where you control the RAG. Having that deep repo knowledge makes a massive difference for those senior-level design decisions... nothing beats a tool that actually knows your infra inside and out.

Following this thread

Solid advice 👍

Exactly what I was thinking

Any updates on this?