background compaction just landed in OpenCode as a experimental setting go try it out and tell us what you think

@opencode how does it work?

@opencode @Teknium this would be nice for hermes

@opencode There goes my 55 second break

@opencode Finally its here, this was in github copilot and I really wanted this in opencode as im experimenting cool stuff with it

@opencode Gotta update my notes already:

@opencode that's the one of parts in my own fork i struggled often, found the solution as forks. letting each agent consume their own context eg one question+5 subagents starting from a single feature list, a common goal set by user bring the fork feature

@opencode is there any relation of compaction with cache?? I wanna understand if compaction is killing the cache

@opencode no it's not killing the cache, its working as it used to but faster

@opencode I saw this in Pi first in the Discord server, so I'm so glad to see more harnesses adopting this!

@opencode finally it forgets stuff in the background like a normal person

@opencode @thsottiaux you know

@opencode woopencode

@opencode Does the agent keep working while it compacts, or does the turn wait until it's done?

@opencode When a tool finishes while the summary is being written, does its result stay verbatim after the switch?

@opencode Would be nice to be possible to configure thresholds per context size or per model or agent.

@opencode I hope @ajambrosino and @thsottiaux are seeing this? 🥲

@opencode a connection is made 🎵

@opencode Is it intentional that openai native compaction only works if you use the auth methods marked "legacy" was confused why mine stopped working when I had to reauth today

@opencode Was trying to figure out other methods of compaction for a while now.. this one looks very interesting

@opencode We do this in Buddy! It's a huge QoL win.

@opencode Hell yeah

@opencode also, this is fast fast

@opencode What about caching? Do you keep the system prompt + tools stable so they stay cached or each compaction just invalidate the whole cache?

@opencode Wish codex team implement this

@opencode Letting compaction run quietly is a welcome improvement for focus.

@opencode Whats the logic? is it configurable? if cacheExpired && context.size() > 80% ??

@opencode FYI a lot of people are asking why DeepSeek v4.1 flash is so slow and doesn’t feel the same as the official api

@opencode The point is to never need compaction so I hope it stays opt in

@opencode compacting in the background instead of freezing the session mid-task is a nice quality of life change. long sessions are where that stop and wait feeling hurts most.

@opencode If it preserves the current step and tool results, background compaction could make long coding sessions way less stop-start.

@opencode compaction in the middle of a task is exactly where mine forget what they were doing will give it a try

@opencode That looks so cool! I hope codex and claude do the same it soon.

@opencode 后台自己压上下文，人才不用盯窗口水位。 这截该外挂出去——别自己手动清 session。

@opencode what do you think about omp's snapcompact ?

@opencode does background mode still wait until the context is full to kick in?