Repository navigation
Replies: 2 comments 3 replies
|
Thanks, @Nepiii, for writing your workaround out so concretely. Yes, we use OpenRig's managed compaction on our own Claude seats. It's opt-in: it's off unless Your idea, compacting toward the next task rather than toward everything, is close to how that reading list works, except ours is ranked by the seat's role and yours by the next spec. What wasn't optimal for you? Which runtime and OpenRig version were you on, and what came back wrong or missing after a compaction? That would tell us the most. |
|
Same problem, a different angle — and I think "optimization" is the right lens, not "recovery." I've been exploring the idea that compaction is something you schedule, not just something that happens to you when the window fills. What I've been experimenting with is a pre-phase headroom gate. Instead of waiting for a threshold (the reactive path OpenRig's managed compaction already covers), we look before a phase boundary and ask: do I have enough headroom to run the next phase without a lossy compaction landing mid-phase?
So the compaction is optimized toward the next phase rather than "toward everything" — which is basically your instinct, "compact toward the next task, not toward everything," made structural and role-aware. I'm still working through it myself: whether the phase budget is the right unit versus a per-spec budget (your orchestrator compacts toward the next spec, we compact toward the next phase, and I'm not sure which generalizes better), and how it composes with OpenRig's managed compaction — my current read is the headroom gate is the when and the managed sequence is the how, but I haven't validated that yet. |
Uh oh!
There was an error while loading. Please reload this page.
So far, we have been taught to use AI in a context window that we need to throw away once we are done with the feature.
OpenRig create a new paradigm with this "continuity" of the seat mechanisms. But I found the compaction method isn't super optimal (maybe I missconfigured something) -> I currently workaround it with a rules my orchestrator agent got :
I feel this allow to get closer to the mental model I got where an agent = a human worker which have his own memory of previous work but won't necessarily use it for a task. It create a more focused agent.
Do you guys use others methods ?
Happy rigging !
All reactions