feat(acp): advertise vision-bridge image capability in initialize response (#6269)

* feat(acp): advertise vision-bridge image capability in initialize response

Adds `_meta.imageCapability` to both stdio and HTTP ACP initialize
responses so external hosts like sudowork can feature-detect native
image handling instead of maintaining a hardcoded allowlist.

Resolves #6086

* test(acp): cover image capability advertisement

* docs(acp): clarify image capability threshold

* fix(acp): align image capability contract

packages/cli/src/acp-integration/acpAgent.test.ts

@@ -148,6 +148,11 @@ vi.mock('@qwen-code/qwen-code-core', () => ({
     AGENT: 'agent',
   },
   FORK_SUBAGENT_TYPE: 'fork',
+  IMAGE_CAPABILITY: Object.freeze({
+    autoHandlesWrongModel: true,
+    maxBytes: 10380902,
+    maxImagesPerTurn: 4,
+  }),
   ALL_PROVIDERS: [
     {
       id: 'deepseek',
@@ -1298,6 +1303,13 @@ describe('QwenAgent MCP SSE/HTTP support', () => {
           sse: true,
           http: true,
         },
+        _meta: {
+          imageCapability: {
+            autoHandlesWrongModel: true,
+            maxBytes: 10380902,
+            maxImagesPerTurn: 4,
+          },
+        },
       },
     });
 

packages/cli/src/acp-integration/acpAgent.ts

@@ -64,6 +64,7 @@ import {
   runManagedAutoMemoryDream,
   runManagedRememberByAgent,
   matchesAnyServerPattern,
+  IMAGE_CAPABILITY,
   registerAcpEventLoopLagGauge,
   startEventLoopLagMonitor,
 } from '@qwen-code/qwen-code-core';
@@ -2832,6 +2833,9 @@ class QwenAgent implements Agent {
           sse: true,
           http: true,
         },
+        _meta: {
+          imageCapability: IMAGE_CAPABILITY,
+        },
       },
     };
   }

packages/cli/src/serve/acp-http/dispatch.ts

@@ -18,6 +18,7 @@ import {
   writeWorkspaceContextFile,
   type SessionArchiveState,
   type SubagentLevel,
+  IMAGE_CAPABILITY,
 } from '@qwen-code/qwen-code-core';
 // Import the permission error classes from the same module REST's
 // `sendPermissionVoteError` uses, so `instanceof` matches the class the bridge
@@ -843,6 +844,7 @@ export class AcpDispatcher {
               this.sessionShellCommandEnabled,
             ),
           },
+          imageCapability: IMAGE_CAPABILITY,
         },
       },
     };

packages/cli/src/serve/acp-http/transport.test.ts

@@ -979,6 +979,20 @@ describe('ACP Streamable HTTP transport (over the wire)', () => {
     );
   });
 
+  it('initialize advertises image capability at _meta.imageCapability', async () => {
+    const { body } = await initializeRaw();
+    const result = body['result'] as {
+      agentCapabilities: {
+        _meta: { imageCapability: Record<string, unknown> };
+      };
+    };
+    expect(result.agentCapabilities._meta.imageCapability).toEqual({
+      autoHandlesWrongModel: true,
+      maxBytes: 10380902,
+      maxImagesPerTurn: 4,
+    });
+  });
+
   it('initialize advertises _qwen/workspace/permissions methods', async () => {
     const { body } = await initializeRaw();
     const result = body['result'] as {

packages/core/src/index.ts

@@ -216,6 +216,7 @@ export * from './services/gitWorktreeService.js';
 export { DEFAULT_MAX_TOOL_CALLS_PER_TURN } from './services/loopDetectionService.js';
 export * from './services/visionBridge/vision-bridge-service.js';
 export * from './services/visionBridge/image-part-utils.js';
+export * from './services/visionBridge/image-capability.js';
 export * from './services/sessionRecap.js';
 export * from './services/sessionService.js';
 export * from './services/sessionTitle.js';

packages/core/src/services/visionBridge/image-capability.ts

@@ -0,0 +1,28 @@
+/**
+ * @license
+ * Copyright 2025 Qwen Team
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+/**
+ * Vision-bridge capability advertisement for ACP initialize responses.
+ *
+ * Boolean flags describe backend-handled behavior for ACP hosts; numeric
+ * thresholds track the same constants enforced by the vision bridge.
+ *
+ * Placed in `agentCapabilities._meta.imageCapability` of the ACP
+ * `initialize` response on both stdio and HTTP paths.
+ */
+import {
+  VISION_BRIDGE_MAX_IMAGE_BASE64_BYTES,
+  VISION_BRIDGE_MAX_IMAGES,
+} from './vision-bridge-constants.js';
+
+export const IMAGE_CAPABILITY = Object.freeze({
+  /** Text-only active models are handled by the vision bridge when configured. */
+  autoHandlesWrongModel: true,
+  /** Current per-image inline base64 payload cap, in bytes. */
+  maxBytes: VISION_BRIDGE_MAX_IMAGE_BASE64_BYTES,
+  /** Current max images processed per turn. */
+  maxImagesPerTurn: VISION_BRIDGE_MAX_IMAGES,
+});

packages/core/src/services/visionBridge/image-part-utils.ts

@@ -5,15 +5,7 @@
  */
 
 import type { Part, PartListUnion } from '@google/genai';
-
-/**
- * Conservative cap on a single image part's base64 length (in MB) before the
- * vision bridge refuses it, so the bridge never makes a side call a provider
- * would reject for size. Measured on the base64 string, which overstates the
- * decoded bytes by ~33%, keeping this comfortably under the repo's ~10MB
- * decoded inline-media ceiling.
- */
-const MAX_IMAGE_BASE64_MB = 9.9;
+import { VISION_BRIDGE_MAX_IMAGE_BASE64_BYTES } from './vision-bridge-constants.js';
 
 /**
  * Normalize a {@link PartListUnion} into a flat array of {@link Part} objects.
@@ -159,7 +151,7 @@ export function isUsableImagePart(part: Part): boolean {
   if (typeof data !== 'string' || data.length === 0) {
     return false;
   }
-  return data.length / (1024 * 1024) <= MAX_IMAGE_BASE64_MB;
+  return data.length <= VISION_BRIDGE_MAX_IMAGE_BASE64_BYTES;
 }
 
 /**

packages/core/src/services/visionBridge/vision-bridge-constants.ts

@@ -0,0 +1,16 @@
+/**
+ * @license
+ * Copyright 2025 Qwen Team
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+export const VISION_BRIDGE_MAX_IMAGES = 4;
+
+/**
+ * Conservative cap on a single image part's base64 payload before the vision
+ * bridge refuses it. Measured on the inline base64 string, which is what ACP
+ * transports over the wire.
+ */
+export const VISION_BRIDGE_MAX_IMAGE_BASE64_BYTES = Math.floor(
+  9.9 * 1024 * 1024,
+);

packages/core/src/services/visionBridge/vision-bridge-service.ts

@@ -16,10 +16,10 @@ import {
   replaceImagesWithText,
   splitImageParts,
 } from './image-part-utils.js';
+import { VISION_BRIDGE_MAX_IMAGES } from './vision-bridge-constants.js';
 
 const debugLogger = createDebugLogger('VISION_BRIDGE');
 const BRIDGE_MAX_OUTPUT_TOKENS = 2048;
-const VISION_BRIDGE_MAX_IMAGES = 4;
 const VISION_BRIDGE_TIMEOUT_MS = 30_000;
 // Cap intent so @-file contents in nonImageParts aren't dumped to the bridge model.
 const BRIDGE_INTENT_MAX_CHARS = 2000;