AI-powered web development platform

I have previously had two similar projects with their own pros and cons. I need to dig up both of them and put together a new template based on a haih agent.

Expand template

I've been thinking a lot about different mechanics in this direction, and so far using a TypeScript server seems like the most promising approach. Why? Because then we can operate not just with the files themselves and read them individually, but also analyze their dependencies. For example, we tell the agent to examine some component. It can simply use a special tool to get the entities that are there (variables, functions, imports, exports, etc.), as well as easily request the additional dependencies it needs without worrying about where they are located. But if it needs the final code of the required dependency, it can easily find out which file it is in and read it.&#x20;

Along the way, we probably need tools for reading files with a specified line range. After all, a code file can be large, and if TS immediately gives us information not only on where it is located, but also in which line range, we don't necessarily have to read the entire file.

Add code analysis features

It is necessary to provide a minimal project lifecycle:

* Creation from scratch or cloning of repository/repositories
* Launching separate environments (dev/stage/prod)
* Maintaining a single project knowledge base

Add minimal project management functionality

Need to see the project state, which files have been modified, commit, etc.
Without this, the agent simply cannot know what has already been done for the task and what hasn't.&#x20;

Need to add a Git integration resolver

Currently, the agent often forgets to log workloads, including what was done, how, and for which task, or records very little information. Instead, it simply tells the user about it verbally.

This is especially critical in research tasks, where it has to perform experimental actions and it is necessary to know exactly what it did, along with example queries, etc.

We haven't decided yet which exact tool to use for this, but there are several options. The most promising seems to be a post-agent that, after the response is completed, will look at what the user said, what the agent did, and log a complete report. However, this still leaves room for inaction. Alternatively, we could add stronger debugging directly into the tools themselves. We already have a reasoning field there where the agent indicates what and why it is doing, i.e., explaining why it is executing this tool. We could also add a task ID field, and if it is specified, automatically log the results. But this might turn out to be redundant.

We can also rely on the fact that the URL of the page from which the write is being made is passed in every request. If we are on a task page, this is an additional indicator, especially when we are working with multiple tasks at once. In principle, we could add a message-sending field directly to the task itself.&#x20;

But these are just ideas.&#x20;

Improve workload tracking

Сейчас у нас выводится подобная информация:&#x20;


⏳ Calling Read Concepts...

⏳ Calling GraphQL Request...

⏳ Calling Shell Execute...

⏳ Calling Shell Execute...

Думаю, хорошо было бы видеть какие именно запросы и команды агент пишет и какие там ответы прилетают.&#x20;

Проработать более подробное логирование того, что выполняет агент

Currently, there is a resolver for creating a project folder, but none for getting a list of folders, deletion, etc.

Add resolvers for working with project directories

We need the ability to generate both overall layouts and individual elements and components. We also need infographics with text, because text is traditionally problematic for image generation. However, it seems the latest Qwen was boasting about this capability.