fix: let server decide model when no user selection exists (#37)

Registry defaults unreliable for deployed models; return null so server uses its configured model. Fixes #35.

src/lib/model-selection.test.ts

@@ -5,7 +5,7 @@ import { chooseModelSelection } from "./model-selection.ts"
 const providers = [
   {
     id: "azure",
-    models: [{ id: "gpt-5.4" }, { id: "gpt-5.2" }],
+    models: [{ id: "gpt-5.4" }, { id: "gpt-5.2" }, { id: "claude-sonnet-4-6" }],
   },
 ]
 
@@ -14,29 +14,31 @@ test("keeps existing selection when still available", () => {
     providers,
     defaults: { azure: "gpt-5.4" },
     existing: { providerID: "azure", modelID: "gpt-5.2" },
-    agentModel: { providerID: "azure", modelID: "claude-sonnet-4-6" },
+    agentModel: null,
   })
   assert.deepEqual(selected, { providerID: "azure", modelID: "gpt-5.2" })
 })
 
-test("prefers provider default over unavailable agent model", () => {
+test("returns null when no existing selection (let server decide)", () => {
+  // Even though defaults and providers exist, we return null to let
+  // the server use its own configured model from opencode.json
   const selected = chooseModelSelection({
     providers,
-    defaults: { azure: "gpt-5.4" },
+    defaults: { azure: "claude-sonnet-4-6" },
     existing: null,
     agentModel: { providerID: "azure", modelID: "claude-sonnet-4-6" },
   })
-  assert.deepEqual(selected, { providerID: "azure", modelID: "gpt-5.4" })
+  assert.equal(selected, null)
 })
 
-test("falls back to first connected provider model when default missing", () => {
+test("returns null when existing selection no longer available", () => {
   const selected = chooseModelSelection({
     providers,
-    defaults: {},
-    existing: null,
+    defaults: { azure: "gpt-5.4" },
+    existing: { providerID: "azure", modelID: "nonexistent-model" },
     agentModel: null,
   })
-  assert.deepEqual(selected, { providerID: "azure", modelID: "gpt-5.4" })
+  assert.equal(selected, null)
 })
 
 test("returns null when no connected providers", () => {

src/lib/model-selection.ts

@@ -22,29 +22,31 @@ export function isModelAvailable(
   return provider.models.some((m) => m.id === selection.modelID)
 }
 
+/**
+ * Chooses a model selection for the session.
+ *
+ * Returns `null` when there is no prior user-explicit choice. A null model
+ * means the server's configured default (opencode.json `"model"` field) will
+ * be used — which is the correct behavior for first-launch and CI where the
+ * provider registry's "default" model may not be deployed on the user's
+ * resource.
+ *
+ * Only returns a non-null value when the user previously made an explicit
+ * selection that is still valid (model available on a connected provider).
+ */
 export function chooseModelSelection(params: {
   providers: ProviderRef[]
   defaults: Record<string, string>
   existing: ModelSelection | null
   agentModel: ModelSelection | null
 }): ModelSelection | null {
-  const { providers, defaults, existing, agentModel } = params
+  const { providers, existing } = params
 
+  // Keep existing user-chosen selection if it's still reachable
   if (isModelAvailable(providers, existing)) return existing
 
-  for (const provider of providers) {
-    const defaultModelID = defaults[provider.id]
-    if (!defaultModelID) continue
-    if (provider.models.some((m) => m.id === defaultModelID)) {
-      return { providerID: provider.id, modelID: defaultModelID }
-    }
-  }
-
-  if (providers.length > 0 && providers[0].models.length > 0) {
-    return { providerID: providers[0].id, modelID: providers[0].models[0].id }
-  }
-
-  if (isModelAvailable(providers, agentModel)) return agentModel
-
+  // Otherwise return null — let the server decide via its own config.
+  // The provider registry's "defaults" map is unreliable (it lists the
+  // upstream registry default, not what's deployed on the user's resource).
   return null
 }