dsh-ark-plan

Calling DeepSeek V4 Flash through the official API in DeepSeek Harness is not cheap. The Volcano Ark plan subscription offers the same model at up to ten times lower token prices during off-peak hours, but the built-in deepseek provider points at api.deepseek.com by default, and naively swapping the baseURL tends to create provider conflicts, silently drop reasoning_effort, or return 401 errors because of key-family mismatches.

dsh-ark-plan takes a deliberately thin approach: no custom provider, no re-declared models. It simply overrides the baseURL of the existing deepseek provider to the Ark plan API so that all model knowledge (1M context, 384K output, effort=max) carries over untouched. Two lanes coexist out of the box; the default model stays on the official route, and when you want to burn budget plan tokens you switch the model picker to the same provider without any additional wiring. A built-in ark_plan_doctor command reads the active config, flags the known footguns, and fires a live test against the Ark chat/completions endpoint with thinking.type=enabled and reasoning_effort=max so you can verify in one step.

Built for users who already run inference workloads on DeepSeek Harness, hold an ark- prefixed API key, and want to slash off-peak token costs without maintaining a separate provider stack. No extra dependencies, no provider re-configuration, and plugin defaults never overwrite user settings.