Claude context length limits causing incomplete code generation

[dhx20150812]

Hi, @mnluzimu, thanks for your great project!

When testing bolt.diy with Claude models, I frequently hit context length limitations that cause:

• Code generation stopping mid-way
• Incomplete/truncated files
• Missing parts of the generated website

This happens especially with complex website generation tasks in WebGen-Bench evaluation.

For example, as you can see in the below picture, the bolt.diy agent got stuck when generating the ChartDisplay.tsx file because of the limited context length of Claude3.7-Sonnet.

Questions

1. Have you encountered this issue? Any known workarounds for Claude's context limits?
2. Does bolt.diy have chunking mechanisms to break large tasks into smaller pieces?
3. Any configuration options or alternative models you'd recommend?

Thank you very much in advance for your help!

[dhx20150812]

[mnluzimu]

Thank you for the question. You can increase the tokens allowed in https://github.com/stackblitz-labs/bolt.diy/blob/main/app/lib/modules/llm/providers/openai-like.ts:

      return res.data.map((model) => ({
        name: model.id,
        label: model.id,
        provider: this.name,
        maxTokenAllowed: 8000,
      }));

However, the problem of client getting stuck in the process of writing a file is not always due to context limit. It has been observed that bolt.diy's web container client tend to behave unstably, especially when the system is busy or a lot of tasks have been cached.