Compact Context Command like Claude Code Or Something Similar

[ratacat]

The command for Claude Code is /compact and dear god is it ever useful, super helpful when you run out of context, or are starting to.

I'm guessing Claude Code goes through and does something like passing the context to to your model in 2 pieces, asking it to remove duplications within the context, remove log outputs, and to preserve files that may be in context. Then to go through and summarize anything else it can. Then it probably passes the rejoined pieces back in together and once more asks it to remove duplicates.

I think it wouldn't be super hard to do something like above, and it would be AMAZING feature for Roo Code. So freakin helpful.. there's definitely more complicated and robust ways you could go about it, keeping context pieces in sets of labeled containers, and rebuilding it as needed.

But I think the simple method would be a really good start.

[ratacat]

For anyone that finds this, if you need an interim solution, I will download the context, dump it in Google Gemini and ask Gemini to basically do above but without the need to split it into two. And then start a new task, and feed it the Compacted Context from Gemini. Slower then a single button click, but works reasonably well

[iamhenry]

This is great. Technically you can create a custom mode to do that and assign it the model of choice

Now that gem 2.5 is out, I hardly worry about context. I also use orchestrator by default so that helps immensely

[mrubens]

I noticed that too in Claude Code - cool feature!

[stevekstevek]

I was coming here to suggest a feature to be smarter about how context compaction works. I just looked at the code and it seemed pretty simple. I'm working on something right now where Roo is running commands that have a lot of output, but most of it is repetitive (the same error hundreds of times). When this happens, it fills the context, and then the current algorithm dumps the first part, and it loses track of what it's learned during the task. A smarter context compaction strategy could really improve the ability to continue on task.

[Spshulem]

Create a custom mode with this prompt:

Condence the current chat thread to something that I can start a new conversation with without losing any critical knowledge. That means describe what has been done, what has been tried, what has failed, what has worked, what the goal is, what we need to do next, where we left off, all the files that we touched, how things were done/thought through, while removing duplications within the context, remove log outputs, and to preserve files that may be in context. Then to go through and summarize anything else relevant to starting a new convo.

Seems to work well.

[iamhenry]

im gonna use this one, ty!

i made it into a structure step by step prompt. i find that i get better and more consistent results that way with a variety of models

https://gist.github.com/iamhenry/7ca015fa54fbf2e02fc9ac97023b24d7

[oOBigManOo]

The compact feature should work automatically as follows:

1 - Auto-Trigger Threshold
The compact feature SHALL automatically engage when context‐window usage ≥ 80%.
The threshold value MUST be user-configurable.
After user confirm accept of the compaction process.

2 - Model Selection for Compaction
The compact feature MUST allow selecting an AI model other than the default chat model.
This selection SHALL be configurable to minimize token costs.

3 - Configurable Compaction Instructions
All compaction directives (e.g., summarization depth, formatting rules) MUST be fully configurable via settings.

4 - Pre-Compaction Backup
Before any compaction operation is performed, a snapshot of the current context SHALL be saved as a file inside a directory named .context.
The backup behavior (e.g., backup path, file naming convention) SHOULD be configurable.

[Luigolas]

This have been done in Cline: cline#3086

[sandyleo26]

what about intelligent context condensing?