dsh-connect-workbuddy

What it solves

You are signed in to WorkBuddy on your machine, but you write code in DSH, and switching back and forth is tedious. This plugin wires that existing local session into the DSH model picker: models become selectable, each labelled with the credit multiplier the upstream actually reports (e.g. GLM-5.3 · x0.79), and promotional windows show the discounted rate automatically. It does not supply accounts, share credentials or bypass billing — every call spends credits from your own account.

Who it’s for

• People already using WorkBuddy who want its models available inside DSH;
• Users who need to manage the China and global editions separately (accounts, model selection, sessions);
• Anyone who cares whether a model is free right now, how long the queue is, or when a rate limit resets;
• Users who want image generation callable by the agent (opt-in).

Highlights

• Two providers in parallel: the China edition registers as workbuddy, the global edition as workbuddy-global; both can coexist and different sessions can pick different sides. Regions without an account list no models, so you never click a model that cannot serve a request.
• Model management: refresh the catalogue from upstream and tick models individually, saved as a draft; image input is decided by a vendor verification table, with unknown cases left unverified rather than guessed; each row has its own “thinking can be turned off” toggle, with a review list of models that answer HTTP 400 for that mode and therefore cannot be enabled.
• Four-tab settings card: Accounts / Models / Capabilities / Diagnostics. Switching tabs neither refetches nor discards unsaved drafts; the diagnostics tab holds only read-only facts you can paste into an issue verbatim.
• Promo badges and busy levels: badges reuse the catalogue data already fetched during refresh (no extra requests); the China edition additionally gets a read-only busy probe rendered as per-model icons, while the global gateway simply has no such endpoint — it degrades silently instead of guessing.
• Honest queue and rate-limit feedback: queue codes produce a “you are #N in line” message, with no account switching and no automatic retry; rate limits parse the reset time out of the 429 response body, and if nothing states it, the plugin says so instead of inventing a countdown.
• Model availability test: probes with a realistic payload size (about 25k input tokens) because this gateway throttles by request size; it is per-model and manual only, since it really does spend credits.
• workbuddy_draw_image (China edition, off by default): lets the agent call text-to-image on its own.

Caveats

The plugin relies on the WorkBuddy client’s own endpoints rather than an official public API, so upstream tightening can break it. Assess the account risk yourself; using the official WorkBuddy client remains the safest option.