Some Ideias from my diry Windows port.

[kn0xx01]

I have a very dirty port of Late to Windows ( i used Opencode closed git to make the first prototype re-style )
Late is very good at beeing effective, but i found that, there was a lot of permissions, and some "what the heck are the subagents doing or did ?".

So I added some features. Like:
session modified files changelog, and what agent did it( consult the subagent session ).
Allow permission per session: each agent requests permissions. specially editing and creating.
Session history from Orchestrator and subagents.
Subagents status.
Session switch in-app.
Session compact triggers ....
tool calling grouping.
and some fluff

this is my very very dirty port/butchered,

I think some of these features can inspire the project.

[mlhher]

Thanks for sharing this! Looks like you've put a lot of work into this port, always nice to see.

session modified files changelog, and what agent did it

I can see it in your screenshot, it's a nice visual, though it slightly conflicts with my vision for Late. I see Late primarily as a tool to architect code rather than read/write it line-by-line, which I feel is best left to an IDE. Though note (as said below) I am open to any kind of feature request though they need to be evaluated individually and there needs to be an identifiable usecase for it.

Allow permission per session: each agent requests permissions

The permission system is actually in the middle of being completely overhauled right now. The current setup can feel like it is asking too much though if we went further with the strings based parser it is simply too irresponsible due to bash being turing complete. An AST based parser is in the making which should resolve most issues immediately while providing the safety.

Session history from Orchestrator and subagents.

Currently, the subagent history is intentionally discarded because the system doesn't use it anywhere. However, I have seen this requested multiple times, so we can look into retaining it.

Subagents status.

Since subagents run sequentially right now, I am curious about the use case for this. Is it primarily for visual feedback, or does it serve another specific workflow purpose?

Session switch in-app.

Can you explain what the advantage is of this over just spinning up a new instance? I was thinking about adding a command palette for unrelated features, but I'm curious how/why people would use in-app switching over just starting a new terminal/instance. I did see someone else talk about this though so I might implement it.

Session compact triggers ....

I've chosen not to implement this because it significantly reduces the intelligence and context of the model. I don't want to do things behind the scenes that degrade output quality without users knowing. This might be added to a specific command that a user can execute or manually set.

tool calling grouping.

Could you expand on this I am not sure I understand?

I see your focus here leans heavily into UI/UX. While I am working on bringing up the UI myself, I want to keep the interface as clean and readable as possible and fitting to the usecase. I specifically want to avoid falling into the same pitfalls as OpenCode and all the other infinite copies. For example, displaying raw code output combined with streaming can make things harder to read and clutter the terminal, and to me personally does not seem to have a real use case.

Thanks again for the post and for taking the time to experiment with Late. If you want to push any of these forward, feel free to open an issue for any specific thing you might like to see and we can see how it goes.

[kn0xx01]

Hi there,
sorry, i just put out the ideias without actually explaining.

• Sessions and sub-agent sessions.
  Late uses fast json history, its efective, In my particular case its not very pratical. my flow is: I run the app on the folder ( windows folder -> terminal here -> late ) and then switch the session, because I am not always with localserver running, neither coding.. I need a quick "continue" session, instead of digging on json files of Late, and figure which session is, and such, i just open late inside folder, pop the session inside late. I already added in previous tests to late an Options modal that uses the json history:

simple and manageable.

sub-agents sessions historys is something that was added by flow not by intent per-se, because on those images of the re-styled Late, it uses Sqlite backend, to keep the history of all sessions. the ideia of the Sqlite Backend, is to also provide later, Memory and knowledge base of the project. Late original view is much more raw, and thats ok. no complains.

sub-agent status: since we are on Orchestrator view, subagents live-and-die, its just to keep track of how many agents where/are active.
for example, with my port, Orchestrator sent 2 sub-agents to 2 diferent files, one of them got stalled ( no parallel jobs in my llama-cpp). its a thing adapted to my specific needs.

Permission per-agent on Orchestrator view.
Currently Late uses a visual indicator when a subagent needs a permission, we need to Tab, see the permision and aprove. on my port, the "permission pop-up" shows on orchestrator view, information of which agent is calling and what is calling and how is calling. then Allow only, allow for the session and deny selections.

I need the compact triggers, running on very very tight room on system ram. I need to compact the session.
UI UX is just something that I am exploring with TUI's implementations, its more for testing and knowledge gathering. performance wise can get really "polluted" very fast.

[mlhher]

Thank you for the thorough explanation!
Note that session management is available natively you might want to also check out the Quickstart Guide.
I will investigate what could/should/might be done specifically in regards to the vision I have for the project.

If you have anything you might want to see yourself in Late or have a suggestion on how to optimize features further (e.g. the session management) feel free to open an issue.