deepseek harness desktop flipped me to cloud mid-refactor

yeah i caught that badge after a long session too. i keep the network panel open now — if you see requests leaving localhost:11434 for api.deepseek.com while "local" is selected, force-quit and clear the app's session storage. mine stuck on cloud until i wiped ~/Library/Application Support/Deepseek-Harness-Desktop.

also set a hard spend cap on their dashboard before you trust the toggle again.

cancelled after $9 of surprise cloud spend last tuesday. the local path is fine when it stays local — the silent failover is the product. if i wanted metered tokens i'd already be in the claude console with a budget alert.

went back to plain ollama + a tmux pane. slower ui, zero surprise invoices.

same here. i put a 30s timeout on every outbound call from the electron wrapper — when it hit api.deepseek.com my watchdog killed the process. still paid $6 before i noticed the invoice email.

local models fail loud. this one failed expensive. now i just ollama run deepseek-coder:6.7b in warp and paste the diff myself. slower ui, no silent failover.

the cloud badge is the product. watched mine flip mid-mlx session when the fan spun up — local got slow, app "helped". wiped Application Support same as quiet_ship.

also blocked *.deepseek.com in /etc/hosts while refactoring. ugly. works. if your local path needs a phone-home fallback you don't have a local path.

did the half-written migration still have the old enum names in the generated prisma client types, or only in schema.prisma? asking because cursor did a similar silent rewrite on me last month and tsc stayed green until a runtime cast blew up in gha.

also — was local slow from context size or was the model actually hung? i keep a fails.md row for "app preferred cloud because local latency > 8s" and i want to know if that's the trigger.