What is the point of the tree functionality?

[superDuperCyberTechno]

The changed files aren't rolled back when going back to an earlier message, so what exactly is the tree supposed to accomplish?

1. You prompt the AI.
2. Changes are made.
3. You prompt the AI.
4. Changes are made.
5. You access the tree and go back to no. 1.

Now you have the no. 2 and no. 4 tasks written to disk, but the AI is just no aware anymore. I'm not being sardonic here, I genuinely don't understand what it's supposed to achieve.

I'm half assuming I misunderstood it completely.

[D7x7z49]

tbh pi's agent seems like an assistant, not a workhorse. ur agent should know what it's doing — if the tree can roll back disk changes, then the human just watches and can't really take part. pi's take is that pi is more like tab completion, not vibe coding.

the whole point of tree is to make it easier to manage your own agent interactions, for branching and evaluating. been using pi for a bit and the caching is surprisingly solid. sometimes simple is best.

[RicardoRyn]

You can use pi-rewind to expand the function of /tree.

[NonlinearFruit]

I think /tree is a tool for controlling the agent's context window. In my experience, there are a couple reasons I might branch the session:

• The agent fell down a rabbit hole
  • A back and forth ironing out a change the agent made
  • Agent fixes it (or I fix it by hand)
  • /tree back in session history to the last helpful message
• Series of tasks that all need a similar context window
  • Load a skill, plus some requirements, maybe a /grill-me (from Matt Pocock)
  • Then using /tree, branch at this point with 'Do {A,B,C,D}'

[shuowu]

I mainly use Pi for reading and research, and /tree is one of my favorite features for deep dives.

When reading a book, I'll branch at the root for each question — every branch gets full context of the source material but stays focused on one thread. If I go too deep into a tangent, I jump back and branch again. The key is it manages your conversation context, not your files — when you navigate back, the AI forgets the later messages and starts fresh from that point.

/tree works fine for navigation, but I wanted my family to use this workflow too and they're not terminal people. So I built pi-tree — a web UI that wraps Pi's session tree into a visual, browser-based reading experience.

[jinglebiscuits]

Similar to other posts, I use the feature when trying to follow a multi-step process. For example, I need help setting up a docker container on my home server for hosting models through ollama. The guide it gives is several steps, and most of the steps need to be broken down further and understood by me. Also some of them don't go as smoothly as hoped. So, I go back and forth with step 1 until I'm satisfied, then use tree to get back to the original step guide and start all over with step 2. I do this as many times as I need until I'm finished. It saves money and prevents context bloat.