I’m the type of guy who is now adorning the new term “vibe coder”. I have a personal app that uses the Gemini API that I have been working on since last year, and now I’m recently aware of the new GenAI SDK. It looks like soon I need to migrate to the new SDK library, but as a vibe coder, I heavily rely on Gemini to help me code, and I don’t really know how to do that.
However, it seems like Gemini 2.5 Pro is still not aware of its new SDK and keeps pointing me back to the old SDK ways.
Does anyone from Google have any idea when this will be solved? I don’t think there’s much I can do about this until Google gives Gemini a knowledge upgrade, but I’m raising it here in the hope that Google sees more and more of their customer desires to focus on the issues.
If anyone has any workarounds, I would love to hear about it too. I’m using Visual Code + Gemini Code Assist + Copilot (so far, Claude 3.7, Sonnet, and GPT 4.1 have no clue either).
Correct it still refers to itself as 1.5 Pro, as a work around I provide context with urls to new documentation and tutorials to guide me, but this again is problematic each time I come across a new bug it does not know how to trouble shoot and leads me down rabbit holes.
I’m still trying to build my local python adk to vertex ai utilising the new Memory Bank to fingers crossed finally eliminate this context issue. Alas my third day and another new bug due to the new update. I have all but given up lol tomorrow is a new day..
Googles Documentation does not have one source of truth either, you will read multiple iterations regarding set up too.
Thanks for bringing this up.. As of today Gemini Code Assist doesn’t have grounding and hence purely relies on the information it was trained upon which clearly justifies these issues.. I will escalate this issue to concerned team..
Late reply, maybe you found a solution since then. I’m a solo founder and whilst not exactly a vibe coder, I faced the same issue as you did when using coding agents, and here are my potential fixes (both free ofc) :
Use an MCP that gives relevant context to your model of choice. A free option that works decently is context7. from the guys at upstash. It’s quite effective : you tell your coding agent to use context7, and its going to call the MCP with the right tool, and its pretty much simply going to inject a set amount of documentation, relevant to your specific needs, with a token-limit you can decide.