deepseek harness desktop ate 6gb then hung on my M1
tried the electron wrapper for deepseek harness last night. installed fine from the github release, dropped my api key in the settings panel.
opened a 3-file refactor prompt, watched activity monitor climb to ~6.1gb, then the UI froze with the spinner still spinning. no crash dialog. had to force quit.
curious if anyone actually uses this vs just hitting the api from a terminal + ollama for the cheap bits. the harness itself seems fine; the desktop shell feels like it's doing something expensive in the background and not telling me.
4 comments
Join the discussion
Log in to comment.
same vibes on m2 air. electron + any long-running agent loop is rough.
i ended up running the harness cli in a tmux pane and only opening the desktop when i want a transcript viewer. still leaks a bit but at least i can kill the pane without losing the shell session.
one thing that helped: set DEEPSEEK_MAX_CONCURRENCY=1. without it it spun two child processes and ram jumped twice as fast.
tmux tip tracks. on my m1 air the desktop also left a helper process after force quit that kept chewing ~900mb until i
pkill -f deepseek.DEEPSEEK_MAX_CONCURRENCY=1 helped. still wondering why an electron shell needs its own embedding cache when the cli already has one — anyone measured rss with the ui closed?
honestly i bounced after day two. the wrapper is cute but i was already paying for deepseek api + claude for the hard paths, and an electron shell that ghosts mid-refactor is not a feature i want to bill a client for.
plain curl + a tiny python runner has been more boring and more reliable. sticky note still says ~$18 this month which i can live with.
if they ship a lighter native client i'll look again. until then terminal stays.
same. billed a client three hours debugging a hung electron wrapper when the api call itself was fine.
sticky note on my monitor now: if the desktop ghosts, the invoice still doesn't. cli + curl until they ship something that doesn't eat 6gb staring at a spinner.