seek://map/dsh-llm-hub
dsh-llm-hub — Models-Page Management Suite
#llm #models #balance
TL;DR
Completes the half of DSH's models page the official adapter leaves undone: fetch the on-sale model list, resident balance with threshold warnings, gateway probes with a health board, evidence-only hiding of unusable providers, and registration of locally installed codex/claude/agy CLIs as subagents. Zero runtime dependencies, maintained by webkubor.
A models-page companion suite: where the shipped DeepSeek adapter stops at a hand-typed model list, dsh-llm-hub adds live model discovery, a resident balance row with threshold warnings, measured gateway latency with a health board, month-to-date token usage with CSV export, and evidence-only hiding of confirmed-unusable providers from the composer dropdown. It also registers locally installed codex/claude/agy CLIs as subagent providers. Zero runtime dependencies; the filtering is subtractive and reverts on unload.
Facts
- install
- dsh plugin --profile web add @dsh-plugins/dsh-llm-hub
- license
- MIT
- note
- 还需把包名手动加入 web profile 的 dsh.profile.bundles 数组并重启——boot graph 变更热载不生效。
Key points
- Takes the unoccupied llm-deepseek discovery slot — per its README the official adapter never registers one and the registry allows a single registration per settings namespace — and live-fetches the api.deepseek.com model list for checkbox adoption
- Mounts a balance row under the DeepSeek card that keeps the amounts as the strings upstream returns; the route answers GET/HEAD only and refuses cross-site reads
- Probes pi-ai gateways for reachability, latency and catalog size; per its README, a modelgo catalog pull measured 71 models where hand-typing listed 11
- Dropdown availability filtering is evidence-only and fail-open — unresolvable credentials, gateway 401/402/403, provably exhausted balance, runtime failures — with automatic recovery once the key is fixed or the account is topped up
- Detects locally installed codex/claude/agy CLIs through the kernel's own executable resolver and registers them as subagent providers; what isn't installed never shows up
FAQ
Why doesn't the official direct-connect models page show balance or the on-sale models?
Per its README, the official adapter never registers model discovery (true of both versions it tested, 0.1.2-rc.1 and 0.1.5-rc.2) and never exposes the balance endpoint; the discovery registry allows one registration per settings namespace, the llm-deepseek slot sits unoccupied, and DSH's slot-contract.d.ts states those extension slots exist for plugins distributed outside the repository.
Why did a provider vanish from the model dropdown? Does it come back?
Only hard evidence hides one: an apiKeyEnv variable that resolves nowhere, the gateway answering the catalog endpoint with 401/402/403, a provably exhausted balance, or a real request failing on credentials or quota — fail-open, anything inconclusive stays. Fix the key or top up and the cache invalidates and re-probes automatically; the card stays in Settings with a red hidden-from-dropdown chip, and Re-check all forces a full re-probe.
How is it different from TokenTracker, the usage tracker on this map?
TokenTracker reads coding tools' session histories on your machine and reckons both token usage and cost; this project reads nothing outside DSH — it tallies the month's in/out/cache tokens inside the Settings page with CSV export, and its README is explicit that it does not compute money: open-source DSH exposes no text-pricing interface, so conversion is left to you against the official price list.
Why can't a deepseek-flash added via the button accept images?
A limit of the harness's discovery contract itself: the llm service keeps only id/name/contextWindow/maxTokens, so discovery candidates cannot carry inputModalities and the entry lands as text-only. The README's fix is to patch the llm-deepseek models config by hand, setting inputModalities to text and image; the official pi-ai route has the same limit and no plugin can fix it.
Official references