dsh-opencodego-multikey

中文 | English

<p align="center"> <img src="https://img.shields.io/badge/DeepSeek%20Harness-plugin-1f6feb" alt="DSH plugin"> <img src="https://img.shields.io/github/v/release/zhuchundashuaige/dsh-opencodego-multikey?sort=semver" alt="Release"> <img src="https://img.shields.io/badge/license-MIT-2da44e" alt="License"> </p>

An OpenCode Go multi-API-key gateway plugin for the DeepSeek Harness Web GUI.

DSH ships an opencode-go provider that only allows one API key per route. This plugin runs a local reverse proxy that folds many keys into one pool:

• Dedicated provider "OpenCode Go Multikey": on load the plugin adds a new provider route opencode-go-multikey (display name OpenCode Go Multikey) whose models are identical to OpenCode Go and whose api key can be anything, with baseURL pointed at the local proxy; picking any of its models load-balances the key pool automatically;
• Multi-key management: add / remove / enable / disable any number of API keys on Settings → OpenCodeGo Multi-Key;
• Quota-aware scheduling: every request automatically uses the key with the most remaining quota; ties are round-robined; nearly-exhausted / invalid / rate-limited keys are skipped and quarantined, then recover when the quarantine lapses;
• Usage shown by dsh-usage-stats: the plugin draws no usage panel. Because "OpenCode Go Multikey" is a dedicated provider, dsh-usage-stats can aggregate the pool's usage cleanly under that provider.

---

Features

---

Architecture

                        ┌──────────────────────── DSH Web process ──────────────────────┐
 DSH pi-ai request ────►│ opencode-go provider baseURL → local proxy                  │
 (baseURL=              │      │                                                      │
  http://127.0.0.1:19781)     ▼                                                       │
                        │  ┌────────────────────────────────────────────┐              │
                        │  │ lib/proxy.js reverse proxy (loopback only)  │              │
                        │  │  1. pick key with the most remaining quota  │              │
                        │  │  2. optional model downgrade (fallbacks)    │              │
                        │  │  3. forward to https://opencode.ai/zen/go   │              │
                        │  │  4. pass through usage (DSH token meter)    │              │
                        │  └──────────────────┬─────────────────────────┘              │
                        │                     │ usage / quarantine                      │
                        │  ┌──────────────────▼─────────────────────────┐              │
                        │  │ KeyPool (memory + state file)               │              │
                        │  └──────────────────┬─────────────────────────┘              │
                        │                     │ pull quota every 60 s                  │
                        │  ┌──────────────────▼─────────────────────────┐              │
                        │  │ lib/quota.js  GET /v1/usage (Bearer)        │              │
                        │  └────────────────────────────────────────────┘              │
                        │  management API (same-origin, loopback): /api/opencodego-multikey/* │
                        └──▲───────────────────────────────────────────────────────────┘
                           │
                    Browser: Settings → "OpenCodeGo Multi-Key" manages the pool;
                    usage is shown per model in the sidebar "Usage & Balance" (dsh-usage-stats)

---

Install

dsh plugin --profile web add github:zhuchundashuaige/dsh-opencodego-multikey

or run the installer from this repository:

node scripts/install.mjs

After installing, restart dsh web once (the host-side "OpenCode Go Multikey" provider injection needs the new code), then hard-refresh the browser (Ctrl+Shift+R).

Hot reload during development

Depend on a local link to this checkout to live-edit the panel:

dsh plugin --profile web remove dsh-opencodego-multikey
dsh plugin --profile web add link:C:\path\to\dsh-opencodego-multikey

The client bundle is served with no-cache and re-hashed: after editing lib/client.js, hard-refresh the browser (host-side lib/*.js changes still require a dsh restart).

> dsh plugin add adds the plugin to dsh.profile.bundles automatically because the package declares dsh.bundle.

---

Configure the DSH provider

On load the plugin auto-writes (via the settings service, idempotently):

• Adds a dedicated provider opencode-go-multikey (display name OpenCode Go Multikey): models identical to opencode-go, api and apiKeyEnv inherited (any api key), baseURL = http://127.0.0.1:{listenPort} (the local proxy);
• Cleans any legacy （Multikey） model variants from the opencode-go provider (its models and baseURL are left as-is).

So the Model dropdown gains a full "OpenCode Go Multikey" set of models — picking any routes the request through the pool load-balancer. The apiKeyEnv value only satisfies DSH's config validation; the proxy ignores the inbound credential and uses the Key pool, which also lets dsh-usage-stats aggregate usage neatly under this dedicated provider.

To manage by hand, set injectProvider: false, create the provider yourself with its baseURL at the proxy, and manage the models manually.

---

Usage

1. In the Model dropdown pick a model under the "OpenCode Go Multikey" provider (e.g. minimax-m3) — calls automatically load-balance the key pool. 2. Open Settings → OpenCodeGo Multi-Key, paste an OpenCode Go API key (optional label), hit Add. The plugin probes its quota immediately and auto-refreshes every 60 s. Each key card shows status (active / quarantined / disabled) and the rolling / weekly / monthly quota bars (used % + reset time). 3. Disable / delete a key, or un-quarantine a key that was auto-isolated. 4. View aggregate usage with dsh-usage-stats: open the sidebar "Usage & Balance" and see each "OpenCode Go Multikey" model's pooled usage.

Scheduling rules

• Each request starts by picking from enabled, non-quarantined, non-exhausted keys.
• Sorted by remaining quota: monthly > weekly > rolling.
• Keys with unknown quota compete with a neutral score (50).
• Quota ≤ exhaustThresholdPct (default 2%) counts as exhausted and is skipped.
• Equal score → least-recently-picked wins (round-robin).
• fallbacks match the model's real id (the supplier's model ids equal opencode-go's).

Model downgrade (optional)

- id: opencodego-multikey
  config:
    fallbackThresholdPct: 10
    fallbacks:
      qwen3.7-max: qwen3.7-plus
      qwen3.7-plus: minimax-m3

When the selected key's remaining quota ≤ fallbackThresholdPct (default 10%) the request model is rewritten to the cheaper one. Downgrade never happens while the quota is unknown.

---

Configuration

All passed via the config of the opencodego-multikey row (all optional; defaults below).

Example:

# <DSH_HOME>/profiles/web/cordis.patch.yml
- id: opencodego-multikey
  config:
    listenPort: 19781
    upstreamBaseURL: https://opencode.ai/zen/go
    fallbackThresholdPct: 15
    fallbacks:
      qwen3.7-max: qwen3.7-plus

---

Management API (loopback, panel use only)

All endpoints accept only loopback requests (127.0.0.1 / ::1, Host header checked).

---

Usage accounting

The plugin no longer renders usage itself. The proxy forwards the upstream usage unmodified, so DSH's own token meter records each call and dsh-usage-stats displays the pooled usage per "OpenCode Go Multikey" model. To do so the proxy understands every usage shape while streaming: OpenAI prompt_tokens / completion_tokens / total_tokens, OpenAI Responses input_tokens / output_tokens / input_tokens_details.cached_tokens, Anthropic input_tokens / output_tokens / cache_read_input_tokens / cache_creation_input_tokens.

• SSE responses parse data: lines and merge by per-field maximum (OpenAI returns the cumulative total at the end; Anthropic reports input at message_start and output at message_delta).
• Models are forwarded with their real ids, so dsh-usage-stats groups usage under the "OpenCode Go Multikey" model id that DSH actually called.

---

Development

npm run check   # syntax-check every file
npm test        # unit + integration tests (node:test, no DSH needed)

Coverage: quota-window parsing and scoring, key-pool selection / round-robin / quarantine, （Multikey） injection idempotency and suffix normalization, plus an end-to-end proxy test against a local fake upstream.

---

i18n

The Settings page reads a single locale namespace opencodegoMultiKey (zh / en dictionaries, registered via ctx.locale); the nav label and hints follow the active UI language. This README ships 中文 and English.

---

Security

See [SECURITY.md](./SECURITY.md). Key points: proxy & API loopback-only; keys live only in memory and the state file; every view shows masked keys; upstream uses the upstreamBaseURL scheme (HTTPS by default).

---

License

[MIT](./LICENSE)