Hi all, Bolt.diy is not modifying the project files and only displaying in the chat window responses to my prompts. Is anyone seeing the same problem? I’m running Bolt.diy in Docker and happens with Deepseek R1/V3 and Google 2.0 Flash.
1 Like
Hi @fairchild,
can you post some screenshots of your bolt and also please check terminal and DEV-Console for Errors.
I am having the same issue, it responds in the chat, but no files are modified or edited.
Running on Windows 11 if that matters.
I got the same problem. At first it worked, but after a while, it stopped updating the files. Just wrote the code. I tried the new experimental settings and it errors.
Are you on the main brach or stable? If stable, then try to switch to main temporarely, cause there are some fixes reagarding context optimization etc.
(turn on “Context optimization and use Optimized Prompt”
Same issue. I ask Bolt.diy to perform a simple task (e.g. fix a broken logo image link, remove a component, add a call to action based on an attached image.) Responds with the code and tells me that "you should be able to see…) but the change is not there. npm run dev instruction does not make the change appear. Refreshing the browser no good either.
Thanks to Laurence Creates, I found that adding the prompt " Please provide me with full complete accurate code only. Do not truncate any of the code you provide." Helps with this.
1 Like
Hi all, has anyone figured this out? I’m running into the same issue both on main and stable branch. Tried different LLMs both local and cloud, it just writes the updated code in chat.
Changing the model to Anthropic’s claude-3.7-sonnet fixed it for me.
I had the same issue, i started with chatgpt-4o-latest selected, and it only gave me a basic astro project, like the blank template and didn’t add any of the features i asked for.
When i asked it to actually make the changes I asked for, none of the files actually changed even though it said it was implementing them. The response was just chat, like the agent didn’t have any tools to access.
I tried a few other OpenAI models and refreshing the page but the same problem persisted, so I eventually clicked the rewind button in my chat history, changed the model to claude-3.7-sonnet and gave it the exact same follow up message about actually implementing my features and BOOM, claude started actually running the terminal commands it needed, started modifying files etc.
My guess is there is something about the openai provider that can’t run agentic workflows or something? I’m new to this project so just a guess based on observation and an hour or so of testing manually.
Its just that the current System Prompt, doesnt work well for all LLMs out there. Anthropic is one of the Models where it works best. In the future we want to provide the possibillty to choose from more system prompts or give the users the abillity to create their own with a “Plugin/Extension” mechanism. So maybe then there is a community extenslibrary where users can provide their working system prompts for different models.
That makes a lot of sense, maybe some recommendations for newbies on starting out with anthropic models, and then letting people know they can branch out to other models to experiment with, but with varying results. Right now, i watched some youtube tutorials including Cole’s, but I didn’t have the video open when i was going through the motions myself, and i didn’t try and reproduce word for word his prompts, i was trying an idea i had for what i thought was a basic app. So I didn’t have any context as to why other models might be good or not, there are so many to keep track of these days.
Anyways, just a simple suggestion, maybe an opportunity to smooth out a friction point that almost had me ditching bolt entirely, I’m glad i stuck with it and tried claude though, if i hadn’t i would have been missing out.
1 Like
Let me know where you would expect such an Info so new users find it, and I bring it in 
Honestly, If it were me, I would look for that sort of information in the README.md right below the section.
Check the bolt.diy Docs for more offical installation instructions and more informations.
I would also look for it right in the “Run Application” section of those docs.
or somewhere before there.
thanks, I have a PR for updating some stuff in readme open and added this for now like this:
At all I am not very happy with the docs/readme, cause there is so many stuff and it feels overload. Think for the future we should try to keep the Readme.md clean and short and move most of them in the “bolt.diy Docs”-Page.
Hi @leex279
I have updated API Key for Google Gemini, and after chatting with bolt all it does is respond to the prompt and nothing happens in the code window. No code, No Files, just an empty box. Is there something that needs to be done that is not there in the documentation for the code to actually be generated ?
@jacklearner no, nothing normally todo.
Which model exactly? Can you export the chat and provide it please, with also a screenshot of your bolt
here’s the screenshot from inside the app. If you see the loader in the workbench thingy → it just keeps on loading and nothing happens, when I enter more chats it just responds and keeps loading with the right side empty all this while. While the screenshot is with qwen I am observing same behaviour with Google Gemini 2.0 as well.
ok but you did not choose google gemini as I can see and you wrote above 
You use Qwen 2 7B, which is a very small model most likely not working with bolt.diy.
Also the Dropdown styling looks a bit weird. what browser you are using? (should be normal chrome)
Sorry. I just edited my comment above, I am actually seeing the same behaviour with Google Gemini Flash (attaching the screenshot). I am using Safari on Mac.
try pls with directly from google. We got the same as a git issue where someone tried this model over openrouter and got the same problem.
I testet it with google provider, same model, without any problems
(its also free via google)