feat(models): show OpenCode Go provider usage quota

[TiAmo-chen]

Summary

Adds an OpenCode Go usage adapter so the provider usage card works for opencode-go (the provider added by the pi model catalog; it is already listed in the app, with icon, and authenticated through auth.json).

Today the request is rejected before any network call:

POST /api/provider-usage/query  {"providerId":"opencode-go"}
HTTP 400  {"error":"Unsupported provider"}

With this change the card shows the plan's rolling / weekly / monthly usage with their reset times.

Changes

• lib/provider-usage-ids.ts — add opencode-go to PROVIDER_USAGE_IDS. ProviderUsageSummary gates rendering on this list, so no UI change is needed.
• lib/provider-usage.ts
  • endpoint https://opencode.ai/zen/go/v1/usage and display name OpenCode Go;
  • normalizeOpenCodeGo() maps usage.{rolling,weekly,monthly}.percent → remaining and the ISO resetsAt → epoch seconds (the client renders new Date(1000 * resetsAt));
  • queryProviderUsage() falls back to the first model's baseUrl when the provider itself has none (see below).
• lib/provider-usage.test.mjs — coverage for the new adapter, the id predicate and the origin guard.

Why the baseUrl fallback is required

opencode-go is a catalog provider whose base URL is defined per model, so provider.baseUrl is undefined and the existing guard fails even after whitelisting:

provider.baseUrl                  = undefined
provider.getModels()[0].baseUrl   = https://opencode.ai/zen/go

isOfficialProviderUsageOrigin("opencode-go", provider.baseUrl)               → false
isOfficialProviderUsageOrigin("opencode-go", models[0].baseUrl)              → true
isOfficialProviderUsageOrigin("opencode-go", "https://proxy.example.com/…")  → false

The auth.baseUrl check inside fetchJson() is untouched, so pointing the provider at a custom origin is still rejected for official queries.

Response shape

GET https://opencode.ai/zen/go/v1/usage        (Authorization: Bearer <key from auth.json>)

{ "usage": {
    "rolling": { "status": "ok", "percent": 3,  "resetsAt": "2026-09-14T05:53:50.774Z" },
    "weekly":  { "status": "ok", "percent": 1,  "resetsAt": "2026-09-21T00:00:00.774Z" },
    "monthly": { "status": "ok", "percent": 98, "resetsAt": "2026-09-19T02:38:27.774Z" } } }

Only the used percentage and a reset timestamp are reported (no dollar amounts), so this produces three unit: "percent" buckets and no metrics.

Verification

• node --experimental-strip-types --test lib/provider-usage.test.mjs → 5/5 pass, including the new
  normalizes OpenCode Go windows into remaining percentages and reset times.
• Live call through queryProviderUsage("opencode-go") on this branch: status: "ready" with
  Rolling 95% / Weekly 98% / Monthly 2% remaining and reset timestamps.
• The same logic was also exercised against the shipped 0.9.1 build artifact and through a real HTTP
  round-trip of /api/provider-usage/query, plus regression checks that deepseek still returns its
  balances and that unsupported ids still answer 400 Unsupported provider.

Notes

• The query uses the credential pi has in auth.json. Setups that rotate a key pool will only see that
  single key; per-key rows could be a follow-up if wanted.
• OpenCode Go reports no window length, only resetsAt, so the buckets carry no windowMinutes.
• The payload's per-window status (ok / rate-limited) is currently unused; it could later map onto the
  existing period: "Unlimited" / rate-limit affordances.