dsh-opencode-routing

English · 简体中文 · 繁體中文 · 日本語 · 한국어 · Русский

A DeepSeek Harness (dsh) plugin that merges two earlier plugins into one package: an OpenCode account pool with automatic A→B failover and quota routing, plus a per-account / per-model usage dashboard. Both live on a single Settings page, "OpenCode 账户与用量".

• Client: one settings.section, refreshed every 60 s; accounts can be added, edited, reordered and probed; model cards carry capability tags; both tables pan by dragging; the UI speaks 10 languages (browser language wins).
• Host: two endpoints — GET /opencode-usage.json (usage, ?refresh=1 forces a refresh) and GET|POST /opencode-quota-router.json (pool and routing; actions below).

Install

npm i dsh-opencode-routing

The package ships its own cordis.patch.yml: it declares the opencode-go / opencode-free providers and a single insert (id: opencode-router). The usage: and router: sub-objects hold each half's settings (both halves own a path, so they cannot share one flat namespace).

- insert:
    - id: opencode-router
      name: 'dsh-opencode-routing'
      config:
        usageUrl: https://opencode.ai/inference/go/v1/usage
        apiKeyRef: OPENCODE_API_KEY
        usage:
          path: /opencode-usage
          plan: go
        router:
          path: /opencode-quota-router
          routes:
            - primary: { provider: opencode-go, model: deepseek-v4.1-flash }
              fallback: { provider: deepseek-account, model: deepseek-flash }

Top-level keys apply to both halves (usageUrl, apiKeyRef, probeIntervalMs, …); keys inside a sub-object only override that half.

Account pool and failover

Pool order is usage order. Exactly one account is "current" at any time, and its key is mirrored into the credential reference the provider reads (default OPENCODE_API_KEY). Switching accounts therefore needs no provider config change: resolveApiKey runs after the agent/request hook, on every streaming call.

• A runs dry, B takes over: on QUOTA / ACCOUNT_QUOTA / 429 once the official allowance is confirmed spent, or on an auth-class error (AUTH, INVALID_CREDENTIAL, MISSING_CREDENTIAL, ACCOUNT_SIGN_IN_REQUIRED, CLIENT_UPGRADE_REQUIRED), the current account is marked unusable (until the window resets, or for 30 minutes on auth failure), routing moves to the next enabled account and retries the same step — the turn is never downgraded.
• Single-model quarantine: when only one model is out of allowance and the pool still has budget, only that model is quarantined (30 minutes by default); every other model keeps using the plan.
• Whole pool dry: only then does the route fall back to its fallback (DeepSeek account by default), with a zero-token GET /inference/go/v1/usage probe every 30 minutes (or at the rolling window reset) that returns to the plan automatically.
• Reset: "重置封禁标记" in Settings clears every exhausted/auth mark and returns to the pool immediately. Accounts and keys are untouched.

Usage dashboard

The official /inference/go/v1/usage endpoint only returns rounded rolling / weekly / monthly percentages and no per-model breakdown, while the plan meters each model separately (5 h = 20 % of the monthly allowance, week = 50 %, month = 100 %). So the page shows two sets of numbers:

• Official percentages (one line per account): straight from the API — real account usage, including other clients and other machines.
• Per-model money (model table): a local estimate. It rescans $DSH_HOME/sessions for the last 40 days and only counts assistant/message events whose provider starts with opencode, priced at the public rates (including DeepSeek's 2× peak window). Usage from other tools has no local record.

"查看用量" opens a per-account dialog with the three windows, their reset times and the probe result (it probes that account only and never switches models).

The peak card reports both calendars side by side — the official API and OpenCode — with the current period and how long (in hours, one decimal) until it flips. Capability chips on the model cards carry a tooltip explaining what the tag means, the header links to the OpenCode Console subscription page, and a floating arrow returns to the top once you have scrolled down. Every price cell shows the cached-input, uncached-input and output line, so a cached hit (the green row) can be compared directly with what the official API would charge for the same tokens.

"编辑密钥" fetches the account's current key from GET /opencode-quota-router.json?key=<account id> and pre-fills the input (a local-only endpoint; the list snapshot only ever carries a mask). Model names follow vendor casing (deepseek-v4.1-flash → DeepSeek V4.1 Flash); vendors with a published brand mark render as an inline SVG, the rest keep a letter badge. "下次探测" is a real timer (30 minutes by default) and the page re-renders every second, so it counts down; opening the page only reads the snapshot and never triggers a probe.

"此刻用哪个更划算" is derived from the subscription leverage: Go charges $10 for $100 of list-priced allowance (≈ ¥710), so OpenCode always costs 10 % of the official price — 90 % cheaper. Real data never favours the official API; the button in the card's corner switches to sample data so you can see what that case would look like.

The capability column comes from the same catalog endpoint (/inference/go/v1/models): modalities.input and capabilities are rendered as text / image / video / audio / PDF, tool calls, reasoning, plus context length. Zero-priced models show "无限" (unlimited).

Peak / off-peak and holidays

DeepSeek's off-peak window is everything outside Beijing time Mon–Fri 09:00–12:00 and 14:00–18:00 (lunch break, nights, weekends and statutory holidays included) at exactly half the peak price. OpenCode only treats UTC Mon–Fri 01–04 and 06–10 as peak and ignores Chinese statutory holidays (anomalyco/opencode#53071, maintainer: "Go pricing follows our own schedule, not DeepSeek's holiday calendar.").

The plugin ships the 2026–2031 Chinese statutory holiday table. It is only applied to the official side when holidayAware: true is set (usage.holidayAware); it defaults to off because OpenCode keeps charging peak on those days. The table lives in usage.js as HOLIDAYS, and isChinaHoliday() can be called on its own.

The subscription price and its allowance can be overridden with usage.planPriceUsd / usage.planAllowanceUsd; the discount on the page is derived from those two numbers, so a plan change from OpenCode is a config edit, not a patch.

Settings API

GET /opencode-quota-router.json returns {mode, reason, active, accounts[], quarantine[], routes[], lastFailure}; account keys are always masked (hint). POST to the same path with {action, ...}:

action	arguments	effect
add	label?, key	create an account (key ≥ 8 chars), store it and probe
update	id, label? / enabled? / key?	rename, disable/enable, rotate the key
remove	id	delete an account and its key (one is always kept)
activate	id	switch to it now
move	id, direction: up|down	reorder
probe	id?	probe every account, or just one
reset	—	clear the exhausted/auth marks and return to the pool

GET /opencode-usage.json returns {plan, configuredPlan, detectedPlan, holidayAware, peak, usage, accounts[], activeAccount, models[], totals}; ?refresh=1 forces a fresh official snapshot first.

Languages

UI strings live in DICT in client.js: zh (source), zh-tw, en, plus es / fr / de / pt / ja / ko / ru (inlined from translation files by scripts/merge-i18n.mjs; a missing key falls back to en, then to the key itself). localeOf() matches the browser language prefix and separates Simplified from Traditional Chinese.

Tests

node test.mjs

No model quota is required: fake context, fake credential store, fake webServer, and a local usage endpoint that answers per API key. Coverage: composite mounting, account adoption, single-model quarantine, A→B failover with key mirroring, per-account probing, whole-pool fallback, reset, settings-action validation, the usage view (capability tags, unlimited models, peak/off-peak maths, billing lines) and the single-account degradation.

Relation to the old packages

dsh-opencode-quota-router and dsh-opencode-usage-settings are its predecessors; everything they did is merged here. Both are marked deprecated on npm and point at this package. Repositories: 2869949486/dsh-opencode-routing (current), dsh-opencode-quota-router, dsh-opencode-usage.

Contributing

Issues and pull requests are welcome — see CONTRIBUTING.md. Run node test.mjs before opening a PR.

License

MIT — see LICENSE.