Hacker News

points by kadoban 4 days ago | 0 comments

DeepSeek v4.1 flash seems to do this thing where it starts saying that we should wrap up the current work and start again "next session" (whatever that means), that the "context budget" is almost up, or some other wording I'm forgetting, as the context gets longer.

I've taken to just manually making it write a handoff file and clearing context completely. I'm pretty much manually compacting instead of my harness doing it. Feels weird, but it does work.

[–] plasticchris 4 days ago

I think it’s trained into them so they emit a stop token with enough window left to generate a compaction summary

[–] kqr 4 days ago

I would assume it's the harness sending the model a notification toward the end. At least Claude Code is designed to attach reminders with tool call output as a way for the harness itself to steer the model slightly.

[–] kadoban 3 days ago

Relatively sure it's not. I'm using pi and I can't find any place it's doing any such thing. My current theory is that it's behavior trained into the model either accidentally or intentionally, probably during post-training in some harness that _does_ give hints of that sort?

[–] kadoban 4 days ago

Any idea what the ~rule is or how it's meant to work? I assume it'd work better if I can wire up my harness to handle it, but no idea. It didn't seem to know itself, from it's training. I didn't search too much yet though.

[–] ptkachykby 3 days ago

You could also increase compaction.reserveTokens in ~/.pi/agent/settings.json to make pi compact more often. Also you can use /compact <instructions> to specify what facts should be included in the summary.

One hard part is that from a handoff file you only have information for 1 session, so you have to manually determine next week which facts are still true. At Substrate we solve this problem using our memory: we have 1 memory outside of an agent and track the source of every fact. https://trysubstrate.co/l/a2hn