o3-mini is now in Langtail 🚀

Currently, state transitions are defined in code (imperatively) rather than declared upfront. This is similar to LangGraph's 'new Command' API, where the visualization tools also can't automatically generate diagrams because they can't predict the possible transitions. There are two potential solutions to this: 1. We could require developers to manually define all possible state transitions upfront. However, this adds extra work and the definitions could easily get out of sync with the actual code implementation. 2. We could create an automatic visualization tool that uses an LLM to analyze the code and generate Mermaid diagrams. This would understand the possible transitions by reading the code itself. What would you prefer?

🚀 Check it out on GitHub:

@pavelsvitek_ @Macaly My bad. Fixed. 🙏

@itsmechase Every time I open the Sidebar example from shadcn and want to make completely basic changes, it fails. It's a change like 'change the text of this button'. This doesn't happen to me with Bolt.

@lgrammel hey 👋, I keep running into situations where I'd like to imperatively append a message to the streamText. For example, with execute in a tool, I'd like to return the tool result and append a user message to steer the model where I need it. I can do it now, but it's not very elegant. Any chance to support it more natively?

@levelsio There are alternatives like EFTA, which focuses on business. The EU is unreformable.

@oromancov I don’t think I can influence this much. As far as I know, they don’t have a ‘reasoning_effort’ setting.

@AravSrinivas It's uncensored, but still with a Chinese narrative, right? Or did you fine-tune it to be American?

Things that I critically miss in and have to do complicated workarounds for. 1. Must be able to open private GitHub repository 2. Changes made in Bolt must be able to be sent back as PR. Like now I'm downloading zip and doing it manually like in the FTP days. Wen? @boltdotnew

It turns out evals are all you need.

I think it would be nice if there was a way to simulate tool calls for reasoning models that currently don't support them, like Deepseek or Gemini Flash Thinking. Especially for streamText, where it's quite complicated to simulate tools and execute api. I would probably be satisfied if I could parse something myself in onChunk and imperatively append another message, which would simulate toolCall/execute and maxSteps.

It seems that people are quite interested.