feat(modality-bridge): resolve audio input capability

open-sse/config/providers/shared.ts

@@ -50,6 +50,7 @@ export interface RegistryModel {
   supportsReasoning?: boolean;
   supportedThinkingEfforts?: readonly string[];
   supportsVision?: boolean;
+  supportsAudio?: boolean;
   supportsXHighEffort?: boolean;
   maxOutputTokens?: number;
   targetFormat?: string;

src/lib/modelCapabilities.ts

@@ -121,6 +121,7 @@ export interface ResolvedModelCapabilities {
   supportsThinking: boolean | null;
   supportsTools: boolean | null;
   supportsVision: boolean | null;
+  supportsAudio: boolean | null;
   supportsMaxTokens: boolean;
   attachment: boolean | null;
   structuredOutput: boolean | null;
@@ -539,6 +540,25 @@ function resolveVisionCapability(
   return null;
 }
 
+/**
+ * Resolve whether a chat model accepts audio input.
+ *
+ * Explicit catalog metadata wins. Synced input modalities are authoritative
+ * only when they contain at least one declared modality; an empty list means
+ * that no source knows the answer and remains `null` so Audio Bridge can act
+ * conservatively.
+ */
+export function resolveAudioCapability(
+  spec: Pick<ModelSpec, "supportsAudio"> | undefined,
+  registryModel: { supportsAudio?: boolean } | null,
+  modalitiesInput: readonly string[]
+): boolean | null {
+  if (typeof registryModel?.supportsAudio === "boolean") return registryModel.supportsAudio;
+  if (typeof spec?.supportsAudio === "boolean") return spec.supportsAudio;
+  if (modalitiesInput.length === 0) return null;
+  return modalitiesInput.some((entry) => String(entry).toLowerCase().includes("audio"));
+}
+
 /**
  * Issue #6524: an operator-set `max_output_tokens` capability override (see
  * `src/lib/db/modelCapabilityOverrides.ts`) is the manual escape hatch for a
@@ -719,6 +739,7 @@ export function getResolvedModelCapabilities(
     lookupKey,
     customVisionOverride
   );
+  const supportsAudio = resolveAudioCapability(spec, registryModel, modalitiesInput);
 
   // #8250: when resolve promoted vision over a contradictory attachment=false,
   // expose attachment=true so catalog / Vision Bridge / clients see one verdict.
@@ -736,6 +757,7 @@ export function getResolvedModelCapabilities(
     supportsThinking,
     supportsTools,
     supportsVision,
+    supportsAudio,
     supportsMaxTokens: heuristicMaxTokens(lookupKey),
     attachment,
     structuredOutput: synced?.structured_output ?? null,

src/shared/constants/modelSpecs.ts

@@ -15,6 +15,7 @@ export interface ModelSpec {
   supportsThinking?: boolean;
   supportsTools?: boolean;
   supportsVision?: boolean;
+  supportsAudio?: boolean;
   // Model defaults to adaptive thinking and REJECTS an explicit `thinking.type:"disabled"`
   // (upstream returns 400). Used to normalize the request when a combo/route substitutes
   // this model after the client already chose `disabled`. See issue #3554.

tests/unit/model-capabilities-audio.test.ts

@@ -0,0 +1,21 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+
+import { resolveAudioCapability } from "../../src/lib/modelCapabilities.ts";
+
+test("explicit registry audio capability wins", () => {
+  assert.equal(resolveAudioCapability(undefined, { supportsAudio: true }, []), true);
+});
+
+test("explicit static audio capability is used when registry is silent", () => {
+  assert.equal(resolveAudioCapability({ supportsAudio: false }, null, []), false);
+});
+
+test("synced input modalities identify audio-capable and text-only models", () => {
+  assert.equal(resolveAudioCapability(undefined, null, ["text", "audio"]), true);
+  assert.equal(resolveAudioCapability(undefined, null, ["text"]), false);
+});
+
+test("missing audio capability evidence stays unknown", () => {
+  assert.equal(resolveAudioCapability(undefined, null, []), null);
+});