feat(gemini/vertex): surface Veo video models in dynamic discovery (#3839)

Gemini / Vertex / Vertex AI Express already discover their catalog dynamically from
v1beta/models, but video (Veo) models use predictLongRunning, which was not mapped —
so they never surfaced. parseGeminiModelsList now recognizes predictLongRunning and
exposes Veo video models alongside chat/image/embedding/audio.

Integrated into release/v3.8.25.

Co-authored-by: diegosouzapw <diegosouza.pw@gmail.com>

file-size-baseline.json

@@ -101,7 +101,7 @@
     "src/shared/constants/providers.ts": 3147,
     "src/shared/constants/sidebarVisibility.ts": 1006,
     "src/shared/services/cliRuntime.ts": 1090,
-    "src/shared/validation/schemas.ts": 2522,
+    "src/shared/validation/schemas.ts": 2523,
     "src/sse/handlers/chat.ts": 1425,
     "src/sse/services/auth.ts": 2207
   },
@@ -126,5 +126,6 @@
   "_rebaseline_2026_06_14_3825_combo_stickiness": "Re-baseline #3825 (sessionless combo stickiness + reasoning-aware readiness): combo.ts 5164→5198 (+34, pós-prettier). Crescimento = deriveComboSessionKey() + effectiveSessionId threading nos sites de read/write do pin server-side. streamReadinessPolicy.ts não-frozen (sob cap). Lógica coesa no handler de combo; não-extraível.",
   "_rebaseline_2026_06_14_r3_3835_kiro_pricing": "PR #3835 own growth: pricing.ts 1470→1508 (+38 = missing Kiro pricing rows, claude-sonnet-4.6 etc., pure data). Also carries inherited release/v3.8.25 drift not yet frozen by prior r2 merges: RequestLoggerV2.tsx 1276→1282 (#3820 resizable log table) and combo.ts 5198→5203 (#3811 round-robin replay-response fix). This PR does not touch RequestLoggerV2/combo.ts source; updating the frozen values restores Fast Quality Gates on the current release branch.",
   "_rebaseline_2026_06_14_r3_3849_transient_hide": "PR #3849 own growth: providerPageHelpers.ts 939→955 (+16 = expanded JSDoc on the auto-hide policy + transient-failure guard in evaluateTestAllEntry). Cohesive helper logic; not extractable.",
-  "_rebaseline_2026_06_14_r3_3838_opencode_quota": "PR #3838 own growth: usage.ts 3394→3408 (+14 = clearer OpenCode Go missing-quota-API messages with OMNIROUTE_OPENCODE_GO_QUOTA_URL override hint + upstream issue links). Message text only; no logic extractable."
+  "_rebaseline_2026_06_14_r3_3838_opencode_quota": "PR #3838 own growth: usage.ts 3394→3408 (+14 = clearer OpenCode Go missing-quota-API messages with OMNIROUTE_OPENCODE_GO_QUOTA_URL override hint + upstream issue links). Message text only; no logic extractable.",
+  "_rebaseline_2026_06_14_r3_3839_veo_video": "PR #3839 own growth: schemas.ts 2522→2523 (+1 = Veo video model (predictLongRunning) validation for Gemini/Vertex dynamic discovery)."
 }

src/lib/providerModels/geminiModelsParser.ts

@@ -3,20 +3,30 @@
  *
  * Each model's `supportedGenerationMethods` is mapped to OmniRoute endpoints:
  *   - generateContent / generateAnswer → "chat"
- *   - predict / predictLongRunning      → "images"   (Imagen models)
- *   - embedContent                      → "embeddings"
- *   - bidiGenerateContent               → "audio"
+ *   - predict                          → "images"  (Imagen image generation)
+ *   - predictLongRunning               → "video"   (Veo video generation)
+ *   - embedContent                     → "embeddings"
+ *   - bidiGenerateContent              → "audio"   (Live real-time audio)
  *
- * This is shared by the `gemini` discovery config and the `vertex` discovery branch: Vertex AI
- * Express keys (and Service Account JSON via a minted OAuth token) list models from the same
- * endpoint, so image models (imagen-*, gemini-*-image) surface dynamically instead of being
- * limited to the small static registry list.
+ * Model-id heuristics refine the long-running bucket because Google exposes both
+ * Imagen and Veo via long-running methods on the same endpoint:
+ *   - id contains "veo"    → ensure "video"
+ *   - id contains "imagen" → force "images" (never "video")
+ *
+ * Note: `gemini-*-image` models (e.g. gemini-3-pro-image) generate images via the
+ * regular `generateContent` path, so they stay "chat" (image output is a chat
+ * modality) and are intentionally NOT reclassified as "images".
+ *
+ * This is shared by the `gemini` discovery config and the `vertex` /
+ * `vertex-partner` (incl. Vertex AI Express key) discovery branches, so every
+ * model the account can access — chat, image, video, audio and embeddings —
+ * surfaces dynamically instead of being limited to the small static registry.
  */
 const METHOD_TO_ENDPOINT: Record<string, string> = {
   generateContent: "chat",
   embedContent: "embeddings",
   predict: "images",
-  predictLongRunning: "images",
+  predictLongRunning: "video",
   bidiGenerateContent: "audio",
   generateAnswer: "chat",
 };
@@ -45,20 +55,33 @@ export function parseGeminiModelsList(data: any): GeminiDiscoveryModel[] {
     const methods: string[] = Array.isArray(m.supportedGenerationMethods)
       ? (m.supportedGenerationMethods as string[])
       : [];
-    const endpoints = [
-      ...new Set(
-        methods
-          .filter((method) => !IGNORED_METHODS.has(method))
-          .map((method) => METHOD_TO_ENDPOINT[method] || "chat")
-      ),
-    ];
-    if (endpoints.length === 0) endpoints.push("chat");
+
+    const endpoints = new Set<string>(
+      methods
+        .filter((method) => !IGNORED_METHODS.has(method))
+        .map((method) => METHOD_TO_ENDPOINT[method] || "chat")
+    );
+
+    const id = ((m.name as string) || (m.id as string) || "").replace(/^models\//, "");
+    const lowerId = id.toLowerCase();
+
+    // Google exposes Imagen (image) and Veo (video) via long-running methods; the
+    // method alone can't always distinguish them, so refine by model id.
+    if (lowerId.includes("veo")) {
+      endpoints.add("video");
+    }
+    if (lowerId.includes("imagen")) {
+      endpoints.delete("video");
+      endpoints.add("images");
+    }
+
+    if (endpoints.size === 0) endpoints.add("chat");
 
     return {
       ...m,
-      id: ((m.name as string) || (m.id as string) || "").replace(/^models\//, ""),
-      name: (m.displayName as string) || ((m.name as string) || "").replace(/^models\//, ""),
-      supportedEndpoints: endpoints,
+      id,
+      name: (m.displayName as string) || id,
+      supportedEndpoints: [...endpoints],
       ...(typeof m.inputTokenLimit === "number" ? { inputTokenLimit: m.inputTokenLimit } : {}),
       ...(typeof m.outputTokenLimit === "number" ? { outputTokenLimit: m.outputTokenLimit } : {}),
       ...(typeof m.description === "string" ? { description: m.description } : {}),

src/shared/validation/schemas.ts

@@ -1056,6 +1056,7 @@ export const providerModelMutationSchema = z.object({
         "rerank",
         "images",
         "audio",
+        "video",
         "audio-transcriptions",
         "audio-speech",
         "images-generations",

tests/unit/gemini-models-parser.test.ts

@@ -36,6 +36,18 @@ const SAMPLE = {
       displayName: "Gemini Live",
       supportedGenerationMethods: ["bidiGenerateContent"],
     },
+    {
+      name: "models/veo-3.0-generate-001",
+      displayName: "Veo 3.0",
+      supportedGenerationMethods: ["predictLongRunning"],
+    },
+    {
+      // Defensive: an Imagen model exposed via a long-running method must stay
+      // "images", never "video".
+      name: "models/imagen-future-preview",
+      displayName: "Imagen Future",
+      supportedGenerationMethods: ["predictLongRunning"],
+    },
   ],
 };
 
@@ -77,6 +89,20 @@ test("parseGeminiModelsList maps embedContent and bidiGenerateContent", () => {
   );
 });
 
+test("parseGeminiModelsList maps Veo predictLongRunning models to the video endpoint", () => {
+  const models = parseGeminiModelsList(SAMPLE);
+  const veo = models.find((m) => m.id === "veo-3.0-generate-001");
+  assert.ok(veo, "veo-3.0-generate-001 should be present");
+  assert.deepEqual(veo!.supportedEndpoints, ["video"]);
+});
+
+test("parseGeminiModelsList keeps Imagen as images even via a long-running method", () => {
+  const models = parseGeminiModelsList(SAMPLE);
+  const imagen = models.find((m) => m.id === "imagen-future-preview");
+  assert.ok(imagen, "imagen-future-preview should be present");
+  assert.deepEqual(imagen!.supportedEndpoints, ["images"]);
+});
+
 test("parseGeminiModelsList defaults to chat and tolerates empty/missing input", () => {
   assert.deepEqual(parseGeminiModelsList({}), []);
   assert.deepEqual(parseGeminiModelsList(null), []);