@stonehao/dsh-account-pool

Provider pool plugin for DeepSeek Harness (dsh): the harness provider registry with 429/quota cooldown failover, a per-provider usage ledger, and live quota probes.

Model

The pool IS the dsh provider registry. Every route llm.listProviders() reports is a member; the pool re-derives lazily (default every 30s, on request and status hits), so provider changes in dsh land on the next turn. There is no separate account layer — no tiers, no model overrides, no custom routes: providers dsh does not configure are not the pool's business.

• Routing (agent/request): a request naming a member provider passes through untouched (your session's model choice always wins). When that member is cooling, the request drifts to the next healthy member (round-robin over registry order, serving that member's first configured model). Requests for non-members are never rewritten.
• Failover (agent/request-error): 429/quota-class failures cool the serving member and answer {kind:'retry'} while another healthy member exists.
• Ledger: per-provider requests/failures/tokens persisted under dataDir.
• Status route: POST /api/account-pool/status returns {ok, value:{rev, generatedAt, accounts[]}} (or {ok, value:null} when the registry is empty) under the authenticated /api fence.
• UI (browser half): a right-sidebar tab with member cards (live cooldown countdowns, token usage, balance capsules), a main-panel dashboard (conversation.view), and a page in Settings → Built-in plugins.

Configuration

Schemastery Config (all fields have defaults):

| field | default | meaning | |---|---|---| | dataDir | ~/.dsh/account-pool | holds providers.json, ledger.json, cooldown.json | | cooldownMs | 60000 | base 429/quota cooldown | | honorRetryAfter | true | extend cooldown to provider Retry-After when larger | | quotaPollMs | 300000 | quota probe interval; 0 disables the poller | | discoverRefreshMs | 30000 | pool re-derivation age; request/status hits trigger it |

providers.json (metadata only)

Keyed by provider route; unknown fields and unknown providers are inert — this file cannot add, remove, or reroute members.

{
  "zai-coding-cn": { "label": "主力度", "quotaProbe": "deepseek-balance" },
  "another-route": { "quotaProbe": "deepseek-balance", "quotaBaseURL": "https://api.deepseek.com", "key": "sk-..." }
}

• label — display name override (defaults to the registry's provider name)
• quotaProbe / quotaBaseURL — balance probe selection and endpoint
• key — probe credential, synced into the harness credential store; the only place a key is ever written down

Install

Two channels; both mount the same cordis.patch.yml.

From npm:

dsh plugin --profile web add @stonehao/dsh-account-pool

From this repository's source (no publish step; the version tracks the repo commit and is pinned in the profile lockfile):

dsh plugin --profile web add "git+https://gitee.com/haostone/dsh-plugins.git#path:packages/dsh-account-pool"

The source channel builds lib/client.js through the package's prepare script, and pnpm only runs a git package's build scripts when the profile allowlists it — add this to $DSH_HOME/profiles/<name>/pnpm-workspace.yaml first, or use npm run plugin install account-pool --profile web --apply from the repository root, which handles it.

dsh plugin writes the package into the profile's dependencies and dsh.profile.bundles, and mounts it through this package's cordis.patch.yml — no hand-edited patch rows, no junction to maintain. Restart the profile afterwards. See the repository AGENTS.md for the link-vs-hardlink behaviour and for isolated-home testing.