dsh-model-hub

A plugin for DSH (DeepSeek Harness 0.2.2-rc.1) that turns it into a model hub: provider sign-ins, extra model routes, rocco GPU dispatch, per-model cost and spend caps, prompt-cache checks, and a dsh-hub CLI. (Formerly dsh-anthropic-membership; the Claude Pro/Max pieces are still here.)

Public copy (DSH dsh-plugin topic): https://github.com/antonio-mastropaolo/dsh-model-hub — the old dsh-anthropic-membership URL redirects there. DSH's node_modules are found portably (lib/dsh-nm.mjs): set DSH_NODE_MODULES if your dsh is not a Homebrew//usr/local npm global.

What it does

• Accounts for Claude (Pro/Max), Codex (ChatGPT), GitHub Copilot, Kimi, OpenRouter and xAI: health, expiry, and a watchdog that refreshes grants ahead of expiry through DSH's own credential seam.
• Routes in the live settings: Google, DashScope, rocco (OrcaRouter rocco-qwen38-q4 first, Mac static mac-qwen38-q4, plus roster models). Default model: rocco / rocco-qwen38-q4.
• rocco GPU dispatch: see which of the 4 L40S are busy and who owns them, load a model on chosen GPUs, unload, return to automatic placement. If OrcaRouter is unreachable or starved, the shim serves the same model from the Mac (x-rocco-failover: mac-qwen38-q4).
• Cost + caps: every model step is priced and appended to a ledger; metered requests over a cap are stopped before they are sent.
• Cache check: tracks prompt-cache hit ratio per session and alerts when it collapses.

Sign in to Claude and Codex (DSH native flow)

1. Open DSH, Settings -> Models -> Model hub -> Accounts.
2. Click Sign in (or Sign in again) next to Claude or Codex. The panel opens the provider link, shows a device code or a paste box if the provider asks for one, and can be cancelled. The grant is stored by DSH itself at llm-pi-ai/<provider> in ~/.dsh/.credentials.yaml and refreshed under DSH's cross-process lock.
3. The dot turns green once the watchdog confirms the grant.

Why the Codex CLI login is not imported: OpenAI refresh tokens rotate. If DSH and the Codex CLI shared one grant (~/.codex/auth.json), whichever refreshed first would invalidate the other's token and log it out. DSH holds its own, separate Codex grant, and this plugin never reads ~/.codex/auth.json. The same goes for Claude: DSH has its own grant. The old login.mjs / logout.mjs scripts and the /login / /logout routes were removed because they wrote the credentials file without DSH's lock and could lose a rotated refresh token; sign-in and sign-out now exist only through the native flow. This plugin never writes llm-pi-ai/* records.

dsh-hub CLI

Install: ~/.local/bin/dsh-hub -> bin/dsh-hub.mjs. Every command takes --json.

dsh-hub status                      host API health, plugin features, plan
dsh-hub accounts                    sign-in / expiry / watchdog state (exit 1 if a refresh failed)
dsh-hub gpus                        GPU cards, owners, loaded model, stale reservations, last failover
dsh-hub switch <model> [--gpus 2,3 | --auto] [--force]
dsh-hub unload                      unload the rocco model now
dsh-hub spend [--since today|7d|30d|all] [--by provider|model|session|day]
dsh-hub prices <provider/model>...  price rows (litellm / override / unpriced)
dsh-hub cache [--session <id>]      prompt-cache hit ratios
dsh-hub doctor                      end-to-end health check

GPU dispatch

• switch <model> loads a model and lets the manager choose GPUs. --gpus 2,3 pins it to exactly those GPUs (the count must match the model's profile; the pin is refused, not guessed, if they are not free).
• A pin survives chat-triggered reloads of the same model. switch <model> --auto clears the pin (the loaded model keeps running where it is until you unload); the UI has the same Return to auto button.
• The shim holds a 300 s lease after any request, so a switch right after a chat answers 409. Wait, or pass --force.
• unload stops the model now; nothing reloads until the next request.
• A stale reservation (for example study-dsv41-hold) is shown with a stale tag. The plugin never deletes it.
• Other users' GPU processes are shown but never touched.

Caps file and ledger

• Ledger: ~/.dsh/model-hub/spend.jsonl (0600, append-only, rotated at 20 MB). One row per model step with tokens, usd (what you are charged), apiEquivUsd (what the API would cost), price source and billing kind. No prompt or completion text is stored.
• Caps: ~/.dsh/model-hub/caps.json, defaults {"dailyUsd":40,"sessionUsd":15,"perProviderDailyUsd":{},"warnPct":50}. A cap of 0 turns that cap off. Only metered providers (Google, OpenRouter, xAI, Ollama) are ever blocked; Claude/Codex/Copilot/Kimi (subscription), DashScope (free) and rocco/Mac (local) are never blocked. A metered model with no known price adds $0 to the ledger, so it warns; if you set a perProviderDailyUsd cap for that provider, its unpriced requests are blocked instead (their spend cannot be counted). Fix by adding the price to ~/.dsh/model-hub/prices.json ({"provider/model": {"inPerM":..,"outPerM":..,"cacheReadPerM":..}}).
• Prices: LiteLLM table cached 24 h in prices.litellm.json; offline fallback is discovered from a pip-installed litellm (or DSH_HUB_PRICE_FALLBACK). Overrides in prices.json win. Unknown models are unpriced, never guessed.
• The cap check reads the ledger fresh for every metered request, so a second DSH process's spend counts at once. Steps already in flight when a cap trips can still finish, so a cap can be overshot by those.

macOS app window (Helixboard-style)

/Applications/DSH v2.app (pinned in the Dock) opens DeepSeek Harness in a chromeless Chrome --app window (same pattern as Helixcode Dashboard v1). It starts dsh web if needed (log: ~/Library/Logs/DSH/dsh-web.log, 0600) and opens it with a launch token the server accepts (dsh mints a new one each boot). Build: copy macos/Info.plist → Contents/, macos/launcher → Contents/MacOS/, DSH.icns → Contents/Resources/.

The composer dock shows Claude 5h/weekly %, a Full access switch (sets permission.defaultPreset and /permission danger-full-access on the live session). The Codex chip was removed on 2026-11-08: nothing fetches Codex usage, so it could only ever say on/off (sign in to Codex under Model hub → Accounts).

Host API + CLI

Host API on 127.0.0.1:8185 (Host header must be loopback:8185). Rules: any request carrying a disallowed Origin gets 403 (GETs too); a POST needs an allowed Origin (http://127.0.0.1:3080, http://localhost:3080) and header x-dsh-model-hub: 1, JSON body only (max 16 KB). CORS headers are sent only for an allowed Origin. There is no legacy header and no /login or /logout: sign-in and sign-out go only through DSH's native flow. Responses never contain secrets.

Route	What
GET /status, /healthz	features (imageCap, compactionSlim, spendGuard), plan usage, access
GET /accounts	provider health, expiry, watchdog state (reportOnly when no refresh seam). Each row has errorKind (revoked / rejected / unreachable / other / null) and terminal; lastError is a short URL-free string
GET /accounts/flows, POST /accounts/signin {provider}, GET /accounts/signin/<p>, POST .../answer {text}, POST .../cancel	DSH-native sign-in driven through ctx.authorization
POST /accounts/refresh	run the watchdog tick now, forced: also retries terminal (revoked) and backed-off grants
GET /rocco/gpus, /rocco/models, `POST /rocco/switch {model, gpus?	auto?, force?}, POST /rocco/unload`
`GET /spend?since=today	7d
GET /prices?models=provider/model,...	price rows (litellm / override / unpriced) + providerHealth
GET /cache[?session=<id>]	prompt-cache hit ratios

Every completed model step is normalised, priced, appended to ~/.dsh/model-hub/spend.jsonl, and fed to the cache tracker. A metered request over a cap (~/.dsh/model-hub/caps.json) is stopped before it is sent, with a SPEND_CAP error. Subscription, free and local routes are never blocked. DSH_HUB_REFRESH=0 makes the account watchdog report-only.

CLI (~/.local/bin/dsh-hub, every command takes --json): status | accounts | gpus | switch <model> [--gpus 2,3 | --auto] [--force] | unload | spend [--since] [--by] | prices <provider/model…> | cache [--session] | doctor. accounts, spend, prices read the data dir and work with DSH down; status and cache need the host API; gpus, switch, unload talk to the shim directly. Tests: npm test.

UI

The client bundle is lib/client.js (module id @local/dsh-model-hub, must equal the package name). It talks only to the host API on 127.0.0.1:8185 with the header x-dsh-model-hub: 1.

• Settings -> Models -> "Model hub" (slot settings.models.footer): Accounts (status dot, expiry, watchdog error, Sign in / Sign in again via DSH's native flow through POST /accounts/signin*, which opens the provider link, shows device codes, accepts a pasted code, and can be cancelled), rocco GPUs (per-GPU cards with owners and a stale-reservation tag, model select, GPU checkboxes, Load / Unload with 2 s progress polling), Costs (today / API-equivalent / 7 days, daily cap meter, price rows).
• Composer dock (slot conversation.composer.dock): existing Claude and Codex chips plus COST TODAY $X.XX · CACHE N%. Colour follows caps.action (warn = amber, block = red); a cache-drop alert from /cache raises a toast with the reason.
• Claude's provider card is a read-only plan-usage card. When a GPU pin is active the rocco panel shows a Return to auto button.
• Screenshots (headless Chromium, 1400 px and 900 px): docs/specs/screens/.
• Deploy: cd ~ && dsh plugin --profile web install (client.js changes load on page refresh; no DSH restart needed).

Account errors, external spend, provider health

• Terminal refresh failures. When a refresh fails with invalid_grant / unauthorized_client (errorKind: "revoked"), the watchdog stops retrying (terminal: true) until the stored credential's expiry changes (a new sign-in) or POST /accounts/refresh forces a tick. Other kinds (rejected 401/403, unreachable timeouts/5xx, other) keep the 15 min backoff. /status does not call the usage API for an expired Anthropic grant (plan: {mode:"expired", expiresAt}); codex.available is true only for an unexpired openai-codex grant.
• External spend is DISPLAY ONLY. GET /spend adds external from the agent-patcher meter ledger (~/.claude/abliteration/spend.jsonl, override with env DSH_HUB_EXTRA_LEDGERS, a single path; rows without provider count as abliteration). It is never read by the caps or the spend guard: caps still see only ~/.dsh/model-hub/spend.jsonl.
• providerHealth. The llm/stream listener watches each request's terminal finish chunk; a provider failure with HTTP 402, insufficient_quota or a "credit" message is recorded in ~/.dsh/model-hub/provider-health.json (0600) and shown in GET /prices as providerHealth[provider] = {code, message, at}. The next successful step from that provider clears it. The hub's own SPEND_CAP blocks are never recorded.
• /rocco/* error replies are always { error: <string> }.

License and attribution

MIT © 2026 Antonio Mastropaolo — see LICENSE. Source: https://github.com/antonio-mastropaolo/dsh-model-hub. If you reuse or index this code, keep the copyright notice and attribute it to Antonio Mastropaolo and antonio-mastropaolo/dsh-model-hub. Embedded provider icons come from LobeHub Icons (MIT); see THIRD_PARTY_NOTICES.md.