dsh-opencode-go-usage

中文 | English

A DeepSeek Harness web-GUI plugin for users connected to OpenCode Go models:

• Quota windows: the subscription's three usage windows — 5-hour rolling / weekly / monthly — with percent used, the reference plan limit, and reset time;
• DSH session details: token usage of every DeepSeek Harness session that used an opencode-go model, drillable down to every model call (turn/step level);
• Cache stats: cache read (hit) / cache write shown at the totals, per-session, and per-call levels, plus a cache hit rate (cache read as a share of total input);
• Model filter: filter sessions by model (chips) to see each model's usage and cache behavior, plus a by-model comparison table;
• Bilingual UI: Chinese / English copy with a one-click language switch in the top-right of the panel (drives and persists the Settings → General → Language preference);
• Composer dock widget: a live one-line quota readout under the composer (30s polling) that changes color by threshold;
• Data freshness: both the quota page and the dock widget show the last successful update time; a failed poll keeps the last successful data and marks it stale instead of blanking it out;
• Diagnostics: the quota page can expand a Diagnostics block — HTTP status, parse version, credential source with a masked key, and the latest 3 request snapshots (raw response bodies), for cross-checking against the official page;
• Billing-caveat note: the DSH details page states that local stats are for reference only, since logged tokens can differ from the official bill.

Features

• Settings sidebar section "OpenCode Go" (a settings.section contribution) with two tabs: Quota and DSH sessions
• Host-side Typert Remote opencodeUsage with three methods: usage / dshUsage / dshSessionMessages
• The DSH session details come entirely from DeepSeek Harness session logs (every assistant/message event carries token accounting) — no local OpenCode client required, works on any DSH deployment
• Cache stats: cacheReadTokens (hit) / cacheWriteTokens (write) per session and per call; hit rate = cache read ÷ (uncached input + cache read)
• Model filter: sessions grouped by source.model, one-click chip filtering that narrows both the totals card and the list; the all-models view adds a By model table (sessions / input / cache read / cache write / output / reasoning / hit rate)
• Reasoning tokens: opencode-go is served through DSH's pi-ai adapter, which folds reasoning into output tokens and never records reasoningTokens — the reasoning column therefore shows — (not reported) instead of a misleading 0; adapters that do report a reasoning breakdown (e.g. dsh-llm-deepseek) display real numbers
• Per-call detail: each model call shows its time, model, and all five token counters; long sessions keep the latest 400 calls (totals always cover every call)
• Composer dock widget (a conversation.composer.dock contribution): 🟢 5h 22% · W 13% · M 13% · ↻ 2h13m (W = weekly, M = monthly, ↻ = reset countdown; hover shows the full labels), colored by the 5-hour rolling window thresholds (default: <60% green / 60–85% orange / ≥85% red)
• Data freshness: the quota page auto-refreshes every 30s (same cadence as the dock) and shows the last successful update time; any failed fetch keeps the previous good data marked stale, and only an explicit Refresh click shows the loading state
• Diagnostics: the quota page can expand a Diagnostics block — last successful fetch / HTTP status / parse version / credential source with masked key / the latest 3 request snapshots (each with time, status, error, and the raw response body; oversized bodies are truncated to a preview)
• Precondition check: if opencode-go is missing from Settings → Models, or no API key is found, it shows guidance instead of an error
• API key resolution: the credential reference the opencode-go provider profile declares (apiKeyEnv, discovered through the llm provider directory), then the conventional OPENCODE_GO_API_KEY from the DSH credentials seam, then OpenCode's auth.json

Install

dsh plugin --profile web add github:yascitom/dsh-opencode-go-box

Or from a local source checkout:

dsh plugin --profile web add file:/path/to/dsh-opencode-go-usage

The package declares dsh.bundle.patch, so dsh plugin add reconciles it into dsh.profile.bundles automatically — no manual cordis.patch.yml row needed. Restart dsh web so the host half and the served client bundle pick up the plugin. The plugin needs the standard web bundle composition (the api-gateway client Remote and the settings.section slot) — the default dsh web profile has both.

Configuration

Host-side tunables live on the plugin row (id: opencode-go-usage); override in $DSH_HOME/profiles/web/cordis.patch.yml:

- id: opencode-go-usage
  config:
    baseUrl: https://opencode.ai/zen/go/v1/usage   # default
    timeoutMs: 15000                                # default
    warnPercent: 60                                 # default: orange at 60%+
    dangerPercent: 85                               # default: red at 85%+
    maxSessions: 30                                 # default: max sessions scanned for DSH details

The usage endpoint

GET https://opencode.ai/zen/go/v1/usage
Authorization: Bearer <API_KEY>

<API_KEY> is the OpenCode Go key (sk-opencode-…) already stored when the model was connected. The endpoint returns:

{
  "usage": {
    "rolling": { "status": "ok", "percent": 9,  "resetsAt": "…" },
    "weekly":  { "status": "ok", "percent": 12, "resetsAt": "…" },
    "monthly": { "status": "ok", "percent": 6,  "resetsAt": "…" }
  }
}

percent is 0–100; resetsAt is ISO-8601. The endpoint is not yet in OpenCode's public docs.

Layout

Known limitations

• The usage endpoint is undocumented and may change; parsing is defensive, and non-200 responses surface as a friendly status rather than a crash.
• Quota limits ($12 / $30 / $60) are shown for context only and are not part of the response; they follow the OpenCode Go plan and can drift.
• DSH session details report token accounting from DSH session logs (no cost amounts) and only cover conversations inside DeepSeek Harness.
• Local token stats can differ from server-side billing (billing rules, rounding, retries, hidden tokens, etc.); the panel marks them for reference only — the official bill is authoritative.

License

MIT