provider.models hook never runs for config-declared providers (ordering) #44081

Summary

The provider.models plugin hook is never invoked for providers that are declared by a
plugin's own config hook (i.e. providers that exist only in config, not in the models.dev
catalog). The hook loop runs against database before config providers are merged into it,
so database[providerID] is still undefined and the provider is skipped.

Built-in plugins (modal, github-copilot, openai/codex, digitalocean) are unaffected
because their providers already exist in the models.dev catalog.

Where

packages/opencode/src/provider/provider.ts (observed on v1.18.18):

const database = mapValues(catalog, toPublicInfo)   // :1376  models.dev catalog only

const configProviders = Object.entries(cfg.provider ?? {})   // :1414  read, not merged yet

for (const hook of plugins) {                                // :1424
  const p = hook.provider
  const models = p?.models
  if (!p || !models) continue

  const providerID = ProviderV2.ID.make(p.id)
  if (disabled.has(providerID)) continue

  const provider = database[providerID]
  if (!provider) continue                                    // :1433  <-- config providers exit here
  ...
}

// extend database from config                               // :1451  <-- too late
for (const [providerID, provider] of configProviders) { ... }

Reproduce

A plugin that declares a provider in its config hook and also exports
provider: { id, models }:

export default async (ctx, options) => ({
  provider: {
    id: "my-provider",
    models: async () => { console.log("NEVER CALLED"); return {} },
  },
  config: async (cfg) => {
    cfg.provider = { ...(cfg.provider ?? {}), "my-provider": { /* ... */ } }
  },
})

Expected: models() is called and its catalogue is used.
Actual: models() is never called. The provider still appears in /config/providers, but
only with the snapshot published by the config hook.

Impact

For an OpenAI-compatible gateway (our case is LiteLLM), the config hook is a boot-time
snapshot: the catalogue is captured once when the process starts. provider.models is the
documented way to resolve a provider's catalogue on demand, so a gateway plugin cannot
refresh its model list without restarting the whole server — which, for opencode serve,
means dropping every in-flight session. We ended up with 16 models in the picker that the
gateway had stopped serving; every request against them returned HTTP 400 until a restart.

Suggested fix

Move the plugin hook loop after the "extend database from config" loop, or seed database
with the config providers before running the hooks. Happy to send a PR if you prefer a
particular ordering.

Note on hot reload

Separately, the provider list is memoised in InstanceState.make and there is no call to
InstanceState.invalidate for it, so even with the ordering fixed the hook would run once
per instance rather than on demand. Filing that here only as context — the ordering bug is
the concrete defect.

Version

v1.18.18