fix(models): stop inventing chat capabilities for specialty surfaces (#8016) (#8022)

Co-authored-by: RaviTharuma <ravitharuma@users.noreply.github.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>

src/lib/modelCapabilities.ts

@@ -16,7 +16,24 @@ import { getModelContextOverride } from "@/lib/db/modelContextOverrides";
 import { getModelCapabilityOverride } from "@/lib/db/modelCapabilityOverrides";
 import { isVisionModelId } from "@/shared/constants/visionModels";
 
-const TOOL_CALLING_UNSUPPORTED_PATTERNS: string[] = [];
+const TOOL_CALLING_UNSUPPORTED_PATTERNS: string[] = [
+  // Specialty / non-chat surfaces must never inherit optimistic tool defaults (#8016)
+  "whisper",
+  "tts-1",
+  "gpt-4o-mini-tts",
+  "omni-moderation",
+  "moderation",
+  "eleven_multilingual",
+  "eleven_turbo",
+  "seedance",
+  "/veo",
+  "veo-",
+  "rerank",
+  "embedding",
+  "dall-e",
+  "flux-",
+  "stable-diffusion",
+];
 const REASONING_UNSUPPORTED_PATTERNS = [
   "antigravity/claude-sonnet-4-6",
   "antigravity/claude-sonnet-4-5",
@@ -26,8 +43,39 @@ const REASONING_UNSUPPORTED_PATTERNS = [
   "antigravity/gpt-oss-",
   "antigravity/gemini-3",
   "antigravity/tab_",
+  // Specialty / non-chat surfaces (#8016)
+  "whisper",
+  "tts-1",
+  "gpt-4o-mini-tts",
+  "omni-moderation",
+  "moderation",
+  "eleven_multilingual",
+  "eleven_turbo",
+  "seedance",
+  "/veo",
+  "veo-",
+  "rerank",
+  "embedding",
+  "dall-e",
+  "flux-",
+  "stable-diffusion",
 ];
 
+/** Catalog/API surface types that are not chat completions. */
+const NON_CHAT_SURFACE_TYPES = new Set([
+  "audio",
+  "video",
+  "image",
+  "moderation",
+  "rerank",
+  "embedding",
+  "music",
+]);
+
+export function isNonChatCatalogSurface(type: unknown): boolean {
+  return typeof type === "string" && NON_CHAT_SURFACE_TYPES.has(type);
+}
+
 const MAX_TOKENS_UNSUPPORTED_PATTERNS = [
   "o1-preview",
   "o1-mini",

src/lib/modelMetadataRegistry.ts

@@ -2,7 +2,10 @@ import { randomUUID } from "node:crypto";
 import { parseModel } from "@omniroute/open-sse/services/model.ts";
 import { getModelInfo } from "@/sse/services/model";
 import { getModelAliases } from "@/lib/db/models";
-import { getResolvedModelCapabilities } from "@/lib/modelCapabilities";
+import {
+  getResolvedModelCapabilities,
+  isNonChatCatalogSurface,
+} from "@/lib/modelCapabilities";
 import {
   getAuthoritativeContextWindow,
   getAuthoritativeProviderContextWindow,
@@ -351,12 +354,24 @@ export function enrichCatalogModelEntry<T extends JsonRecord>(
     getAuthoritativeProviderContextWindow(provider, model) ??
     getAuthoritativeContextWindow(metadata.model) ??
     getAuthoritativeContextWindow(model);
+  const specialtySurface = isNonChatCatalogSurface(entry.type);
   const capabilityFields = {
     ...(typeof metadata.capabilities.vision === "boolean"
       ? { vision: metadata.capabilities.vision }
       : {}),
-    tool_calling: metadata.capabilities.toolCalling,
-    reasoning: metadata.capabilities.reasoning,
+    // #8016: never invent chat tool/reasoning defaults onto specialty surfaces.
+    // Only copy boolean true when authoritative metadata says so; specialty rows
+    // default to false instead of optimistic chat heuristics.
+    tool_calling: specialtySurface
+      ? metadata.capabilities.supportsTools === true || metadata.capabilities.toolCalling === true
+        ? true
+        : false
+      : metadata.capabilities.toolCalling,
+    reasoning: specialtySurface
+      ? metadata.capabilities.supportsThinking === true || metadata.capabilities.reasoning === true
+        ? true
+        : false
+      : metadata.capabilities.reasoning,
     // #6241: surface thinking support + the canonical effort tiers so the frontend can
     // render the effort/thinking toggles. `thinking` is kept for back-compat; `supportsThinking`
     // is the explicit flag and `effort_tiers` lists the selectable reasoning levels
@@ -403,10 +418,22 @@ export function enrichCatalogModelEntry<T extends JsonRecord>(
   }
 
   if (
+    !specialtySurface &&
     (typeof nextEntry.context_length !== "number" || authoritativeContextWindow !== null) &&
     typeof metadata.limits.contextWindow === "number"
   ) {
     nextEntry.context_length = metadata.limits.contextWindow;
+  } else if (
+    specialtySurface &&
+    authoritativeContextWindow !== null &&
+    typeof authoritativeContextWindow === "number"
+  ) {
+    // Only authoritative static windows may decorate specialty rows.
+    nextEntry.context_length = authoritativeContextWindow;
+  } else if (specialtySurface && typeof nextEntry.context_length === "number") {
+    // Keep an explicit source-provided context if the emitter already set one.
+  } else if (specialtySurface) {
+    delete nextEntry.context_length;
   }
 
   if (typeof metadata.limits.maxOutputTokens === "number" && metadata.limits.maxOutputTokens > 0) {

tests/unit/model-capabilities-specialty-8016.test.ts

@@ -0,0 +1,44 @@
+import assert from "node:assert/strict";
+import { describe, it } from "node:test";
+import {
+  getResolvedModelCapabilities,
+  isNonChatCatalogSurface,
+} from "../../src/lib/modelCapabilities.ts";
+import { enrichCatalogModelEntry } from "../../src/lib/modelMetadataRegistry.ts";
+
+describe("specialty catalog surfaces (#8016)", () => {
+  it("recognizes non-chat catalog types", () => {
+    assert.equal(isNonChatCatalogSurface("audio"), true);
+    assert.equal(isNonChatCatalogSurface("video"), true);
+    assert.equal(isNonChatCatalogSurface("moderation"), true);
+    assert.equal(isNonChatCatalogSurface("chat"), false);
+    assert.equal(isNonChatCatalogSurface(undefined), false);
+  });
+
+  it("does not optimistically enable tools/reasoning for specialty model ids", () => {
+    const whisper = getResolvedModelCapabilities({ provider: "openai", model: "whisper-1" });
+    assert.equal(whisper.toolCalling, false);
+    assert.equal(whisper.reasoning, false);
+
+    const tts = getResolvedModelCapabilities({ provider: "openai", model: "tts-1" });
+    assert.equal(tts.toolCalling, false);
+    assert.equal(tts.reasoning, false);
+
+    const veo = getResolvedModelCapabilities({ provider: "veo-free", model: "veo" });
+    assert.equal(veo.toolCalling, false);
+    assert.equal(veo.reasoning, false);
+  });
+
+  it("enrichment does not invent chat tool/reasoning on typed specialty rows", () => {
+    const enriched = enrichCatalogModelEntry({
+      id: "openai/whisper-1",
+      owned_by: "openai",
+      root: "whisper-1",
+      type: "audio",
+      subtype: "transcription",
+    });
+    const caps = (enriched.capabilities || {}) as Record<string, unknown>;
+    assert.equal(caps.tool_calling, false);
+    assert.equal(caps.reasoning, false);
+  });
+});