dsh-workbuddy2api

Manually wiring a workbuddy2api gateway into DeepSeek Harness used to mean hand-copying every model field into settings.yaml, babysitting a long-running background process, installing a full Go toolchain to cross-compile the binary, and running a login script that requires both bash and python3. This plugin strips all of that out: the gateway is spawned automatically when dsh starts and reaped on exit, an already-healthy instance on the same port is reused without a duplicate process, and the entire model catalog--context window, output ceiling, reasoning tiers--is read live from the gateway /v1/models endpoint so no static metadata table is ever maintained. A single /wb2api-setup command downloads the pre-built binary with SHA256 verification, prepares the runtime directory, walks through OAuth login, and syncs the model list, all in pure Node with zero dependence on Go, Python, or bash.

The plugin is deliberately thin. Protocol adaptation, account pooling, weighted rotation, tiered circuit-breaking, and token refresh all stay inside the Go gateway; the plugin only owns process lifecycle and model-directory automation. Six management commands cover status, start, restart, login, and account switching. Switching accounts works by renaming credential files rather than patching gateway code, so every upstream release stays conflict-free. Request serialization enforces two hard contracts--tool results must carry role:tool and user-message content must be verbatim-stable within a turn--so the gateway per-turn RequestID aggregation key does not silently drift at every step.

It is built for users who already run a workbuddy2api gateway and want it cleanly plugged into dsh chat: no Go install, no hand-copied model metadata, no lingering processes after exit, and a one-command rebuild on a fresh machine. Distributed as an npm package, it is a self-contained, auditable module that never touches gateway source code.