new "compaction" is completely broken

[dogmatic69]

Describe the bug

It does not trigger at 95%, have seen it at 100%.
It takes AGES to run, and when it finally finishes it has completely "forgotten" what the original request is so stops processing and asks what it should be doing...

paraphrasing:

"please implement authentication"
... works for 10 minutes...

...compacting..

"Can you clarify what it was you wanted to implement"

Affected version

0.0.374 Commit: 9fecd24

Steps to reproduce the behavior

1. setup a complex task
2. let it run till context is full
3. compaction runs
4. LLM has no idea what to do

Expected behavior

compaction should not completely wipe all memory, especially not user prompts

Additional context

No response

[dogmatic69]

It's constantly hitting compaction even after a short time. If I catch it early and end the session token usage is only at 70k which seems far too low to be doing compaction.

It's essentially unusable now. How can I disable compaction and have it how it was before

[dereklegenzoff]

Hi @dogmatic69 - thanks for flagging this. Could you share what model you're using?

It does not trigger at 95%, have seen it at 100%.

Are you referring to the indicator in the status bar that shows "Context XX% Full"? The "Context XX% Full" shows what percentage of the way you are to compaction happening so compaction happening when that indicator says 100% is expected. You can see more token details with the /context

It takes AGES to run

How long are you seeing it take? We'd expect it to be somewhere around 30-60 seconds

[dogmatic69]

model: claude-opus-4.5

I've seen the counter at over 100 seconds pretty much every time I've noticed.

Yes, I'm referencing the status bar %. As mentioned, when looking at /context it was happening at 70k tokens which is far too low. Sometimes I can run a big task with multiple agents and finish with 1.9m tokens and no compaction. It's very random..

One thing that is for certain, if it hits compaction, the session is dead and useless. All "memory" is gone including the original request.

As a side note, I've been experimenting with prompt including "do not use compaction at all" and seem to be able to avoid this all together 😂

[dereklegenzoff]

Hi @dogmatic69 - we made some updates to compaction to improve the behavior here. Compaction now runs in the background and we've made some other updates to prevent the model from "losing the plot" of the conversation.

[RyanHecht]

Closing this as we've made a number of compaction improvements -- feel free to open a new issue if you're still running into problems!