Outline tree view

[pcuci]

Nest Tool Calls and follow-on Assistant responses so they show up in the Outline view for easier navigation. This would match LangSmith's threads view and would feel familiar, hence improved DX

Next level: imagine some kind of MCP/Tool that, in the background, summarizes each message

lol, ok, dumb idea

you actually rely on the specific header to be exactly # %% assistant|tool_execute|user

[rusiaaman]

The first visualisation you shared of grouping tool calls and follow-on Assistant responses looks great in the outline tree view.

This generally makes sense to me, as long as there isn't a hard restriction while parsing on which level a message should be.

The parsing should just ignore the level of nesting and work as it works currently. Additionally, we should emit "## %% assistant" or "## %% tool_execute" instead of first level heading, when sending tool results or when executing tools respectively.

I'll pick it up unless you want to give it a shot.

[pcuci]

oh, I see, so the LLMs don't actually "see" the hierarchical view, to them each message is either a system message, a tool call, or a human message.

this is more like a nice to have, I haven't gotten into a situation yet where multiple recurrent tool calls get made to fulfil a user request, in chat.md

feel free to have a go at it, otherwise I'll pick it up next time it frustrates me 😄

what I find interesting with the 1M context window Gemini 2.5 pro is that it changed my workflow, once again. This thing can now reason and suggest edits across multiple files reliably. So their web UI is quite horrible when I preload it with say https://microsoft.github.io/genaiscript/llms-full.txt or https://langchain-ai.github.io/langgraph/llms-full.txt which are about 300K tokens each. The browser can't handle so much text as easily as vscode can. Now imagine 10M context window

so I was thinking of creating an MCP that would preload these llm.txt in chat.md as a tool call, right at the start of a new chat.md thread. but maybe it should be an MCP prompt(?) instead

I might be going offtopic here, the idea is, sometimes I dump large amount of text in a chat.md thread to bias the machine a certain way, then I hate scrolling up and down 600 pages, so I gravitated toward the outline view to hop around. I'm even spending time to learn and practice more vim motions in vs code, to navigate around large quantities of text more easily...

Now imagine when 10M context window models start to not suck

[rusiaaman]

so I was thinking of creating an MCP that would preload these llm.txt in chat.md as a tool call, right at the start of a new chat.md thread. but maybe it should be an MCP prompt(?) instead

It should definitely be an MCP prompt.

Or if you don't find it much hassle, you can paste link to the file on your local disk like [reference1](./langgraph/llms-full.txt) (just in case you didn't know if that's possible) to attach a big text. In fact you can drag and drop any text or image file from file explorer pressing shift . This could also solve your issue of scrolling through a lot of text.

I also feel the need of MCP prompt. So that's a feature that should be of bigger priority. I'll implement it this weekend itself.

[pcuci]

actually, I think this one might already work as such? "Vibe Coding" LangGraph apps with llms.txt and MCP

perhaps an idea worth forking

[rusiaaman]

actually, I think this one might already work as such?

Tools to surface documentation is good for when you have dynamic list of documents and it's upto LLM to decide what document to fetch. MCP prompts excel when you know what exactly to prompt / what document to attach.

I've added MCP prompts feature in the release v0.2.1. When you hover over status bar it'll show prompts to attach that you can click, it'll get inserted at cursor.

You can have an MCP server with both tools to fetch documentation like in the video and also expose prompts to attach the documents manually.