@tanstack/ai 0.31.0 → 0.33.0

package/dist/esm/utilities/numbers.d.ts

@@ -0,0 +1,8 @@
+/**
+ * Return the first candidate that is a finite `number`, or `undefined`.
+ *
+ * Handy for picking a value from among several possible spellings/sources where
+ * only some are populated — e.g. the provider-native sampling option names read
+ * by the OTel middleware, or the optional numeric fields on `TokenUsage`.
+ */
+export declare function firstNumber(...candidates: Array<unknown>): number | undefined;

package/dist/esm/utilities/numbers.js

@@ -0,0 +1,12 @@
+function firstNumber(...candidates) {
+  for (const candidate of candidates) {
+    if (typeof candidate === "number" && Number.isFinite(candidate)) {
+      return candidate;
+    }
+  }
+  return void 0;
+}
+export {
+  firstNumber
+};
+//# sourceMappingURL=numbers.js.map

package/dist/esm/utilities/numbers.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"numbers.js","sources":["../../../src/utilities/numbers.ts"],"sourcesContent":["/**\n * Return the first candidate that is a finite `number`, or `undefined`.\n *\n * Handy for picking a value from among several possible spellings/sources where\n * only some are populated — e.g. the provider-native sampling option names read\n * by the OTel middleware, or the optional numeric fields on `TokenUsage`.\n */\nexport function firstNumber(...candidates: Array<unknown>): number | undefined {\n  for (const candidate of candidates) {\n    if (typeof candidate === 'number' && Number.isFinite(candidate)) {\n      return candidate\n    }\n  }\n  return undefined\n}\n"],"names":[],"mappings":"AAOO,SAAS,eAAe,YAAgD;AAC7E,aAAW,aAAa,YAAY;AAClC,QAAI,OAAO,cAAc,YAAY,OAAO,SAAS,SAAS,GAAG;AAC/D,aAAO;AAAA,IACT;AAAA,EACF;AACA,SAAO;AACT;"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tanstack/ai",
-  "version": "0.31.0",
+  "version": "0.33.0",
   "description": "Type-safe TypeScript AI SDK for streaming chat, tool calling, agents, structured outputs, and multimodal generation.",
   "author": "Tanner Linsley",
   "license": "MIT",
@@ -76,7 +76,7 @@
     "@ag-ui/core": "^0.0.52",
     "@standard-schema/spec": "^1.1.0",
     "partial-json": "^0.1.7",
-    "@tanstack/ai-event-client": "0.6.2"
+    "@tanstack/ai-event-client": "0.6.4"
   },
   "peerDependencies": {
     "@opentelemetry/api": ">=1.9.0"

package/skills/ai-core/media-generation/SKILL.md

@@ -3,8 +3,9 @@ name: ai-core/media-generation
 description: >
   Image, audio, video, speech (TTS), and transcription generation using
   activity-specific adapters: generateImage() with openaiImage/geminiImage,
-  generateAudio() with geminiAudio/falAudio, generateVideo() with async
-  polling, generateSpeech() with openaiSpeech, generateTranscription() with
+  generateAudio() with geminiAudio/falAudio, generateVideo() with
+  openaiVideo/geminiVideo (async polling, per-model typed durations),
+  generateSpeech() with openaiSpeech, generateTranscription() with
   openaiTranscription. React hooks: useGenerateImage, useGenerateAudio,
   useGenerateSpeech, useTranscription, useGenerateVideo.
   TanStack Start server function integration with toServerSentEventsResponse.
@@ -189,6 +190,103 @@ Result shape: `ImageGenerationResult` with `images` array where each entry
 has `b64Json?`, `url?`, and `revisedPrompt?`. OpenAI image URLs expire
 after 1 hour -- download or display immediately.
 
+#### Image-conditioned generation: multimodal `prompt` parts
+
+Both `generateImage()` and `generateVideo()` accept the `prompt` either as
+a plain string or as an ordered array of content parts (`TextPart` /
+`ImagePart` / `VideoPart` / `AudioPart` — the same shapes used elsewhere in
+TanStack AI). Part order is meaningful: natively multimodal providers
+(Gemini, OpenRouter) receive parts in order; named-field providers (OpenAI,
+fal, xAI) extract media parts and flatten the text. Prompt text is always
+sent verbatim — to reference inputs from the prompt, write the provider's
+own syntax (fal `@Image1`, OpenAI "image 1" prose); the SDK never injects
+or rewrites markers. Each media part may carry an optional
+`metadata.role` hint that adapters use to route the part to the
+provider-specific field. The accepted part types are narrowed per model at
+compile time via the adapter's input-modality map.
+
+```typescript
+import { generateImage } from '@tanstack/ai'
+import { openaiImage } from '@tanstack/ai-openai'
+
+// Image-to-image (OpenAI gpt-image-2 / gpt-image-1, dall-e-2)
+await generateImage({
+  adapter: openaiImage('gpt-image-2'),
+  prompt: [
+    { type: 'text', content: 'Turn this into a cinematic product photo' },
+    { type: 'image', source: { type: 'url', value: 'https://…/product.png' } },
+  ],
+})
+
+// Multi-reference (up to 16 for gpt-image models; up to ~14 for Gemini native
+// — a provider limit, not enforced by the SDK)
+await generateImage({
+  adapter: openaiImage('gpt-image-2'),
+  prompt: [
+    { type: 'text', content: 'Apply the second image as style to the first' },
+    { type: 'image', source: { type: 'url', value: 'https://…/product.png' } },
+    { type: 'image', source: { type: 'url', value: 'https://…/style.png' } },
+  ],
+})
+
+// Inpaint via metadata.role === 'mask' (OpenAI gpt-image models, dall-e-2; fal mask_url)
+await generateImage({
+  adapter: openaiImage('gpt-image-2'),
+  prompt: [
+    { type: 'text', content: 'Replace the masked region with a tree' },
+    { type: 'image', source: { type: 'url', value: photoUrl } },
+    {
+      type: 'image',
+      source: { type: 'url', value: maskUrl },
+      metadata: { role: 'mask' },
+    },
+  ],
+})
+
+// Image-to-video (OpenAI Sora: single input_reference; fal: image_url + optional end_image_url)
+import { generateVideo } from '@tanstack/ai'
+import { falVideo } from '@tanstack/ai-fal'
+
+await generateVideo({
+  adapter: falVideo('fal-ai/kling-video/v3/pro/image-to-video'),
+  prompt: [
+    { type: 'image', source: { type: 'url', value: firstFrameUrl } },
+    { type: 'text', content: 'Slow cinematic push-in' },
+    {
+      type: 'image',
+      source: { type: 'url', value: lastFrameUrl },
+      metadata: { role: 'end_frame' },
+    },
+  ],
+})
+```
+
+**Role hints** (`metadata.role`):
+
+| Role            | Maps to                                                                                               |
+| --------------- | ----------------------------------------------------------------------------------------------------- |
+| `'reference'`   | fal `reference_image_urls`; Gemini multimodal part; positional otherwise                              |
+| `'character'`   | Same as `'reference'`; Veo `referenceImages` slot (planned — no Veo adapter yet)                      |
+| `'mask'`        | OpenAI `mask` (gpt-image-2, gpt-image-1, dall-e-2); fal `mask_url`                                    |
+| `'control'`     | fal `control_image_url` (ControlNet / depth / pose)                                                   |
+| `'start_frame'` | fal `start_image_url` (or the endpoint's field, e.g. `image_url` on Kling i2v); Veo `image` (planned) |
+| `'end_frame'`   | fal `end_image_url` (or e.g. `tail_image_url` / `last_frame_url`); Veo `lastFrame` (planned)          |
+
+**Provider support matrix:**
+
+| Provider   | `generateImage` image parts                                                                                                                                                                              | `generateVideo` image parts                                                                                                                                                                        |
+| ---------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| OpenAI     | gpt-image-2 / gpt-image-1 / -mini → `images.edit()` (up to 16). dall-e-2 → edit (1). dall-e-3 throws.                                                                                                    | Sora-2 / -pro → `input_reference` (single). Throws if >1.                                                                                                                                          |
+| Gemini     | Native (gemini-\*-flash-image, "nano-banana") → multimodal `contents`. Imagen throws.                                                                                                                    | No native Veo adapter yet — deferred to a follow-up.                                                                                                                                               |
+| fal        | Per-endpoint field names from a generated map (`pnpm generate:fal-image-fields`). Defaults: 1 input → `image_url`; >1 → `image_urls`; roles → `mask_url` / `control_image_url` / `reference_image_urls`. | Per-endpoint map (e.g. Kling i2v start frame → `image_url`). Defaults: 1 input → `image_url`; `start_frame`/`end_frame` → `start_image_url`/`end_image_url`; `reference` → `reference_image_urls`. |
+| Grok       | grok-imagine models → `/v1/images/edits` JSON endpoint (≤3 sources, addressed by xAI in request order; prompt sent verbatim; mask/control throw). grok-2-image-1212 throws.                              | n/a                                                                                                                                                                                                |
+| OpenRouter | Prompt parts map 1:1 onto multimodal `text` / `image_url` content parts, preserving interleaved order.                                                                                                   | n/a                                                                                                                                                                                                |
+| Anthropic  | n/a (no image generation API).                                                                                                                                                                           | n/a                                                                                                                                                                                                |
+
+Video and audio prompt parts follow the same `metadata.role` convention
+for video-to-video and lipsync flows on fal; other providers throw when
+they're passed.
+
 ### 2. Audio Generation (Music, Sound Effects)
 
 Distinct from TTS — `generateAudio()` produces non-speech audio content.
@@ -331,6 +429,31 @@ const stream = generateVideo({
 return toServerSentEventsResponse(stream)
 ```
 
+Google Veo (`@tanstack/ai-gemini`) uses the same jobs/polling flow. Its
+`duration` option is typed per model (e.g. `4 | 6 | 8` for Veo 3.x,
+`5 | 6 | 8` for Veo 2); use `adapter.snapDuration(seconds)` to coerce raw
+seconds and `adapter.availableDurations()` to enumerate the valid set.
+Image prompt parts route by `metadata.role`: first un-roled /
+`'start_frame'` image → input image, `'end_frame'` → `lastFrame`,
+`'reference'` / `'character'` → `referenceImages`:
+
+```typescript
+import { geminiVideo } from '@tanstack/ai-gemini'
+
+const adapter = geminiVideo('veo-3.1-generate-preview')
+adapter.availableDurations() // { kind: 'discrete', values: [4, 6, 8] }
+
+const { jobId } = await generateVideo({
+  adapter,
+  prompt: 'A golden retriever playing in sunflowers',
+  size: '16:9', // Veo sizes are aspect ratios: '16:9' | '9:16'
+  duration: adapter.snapDuration(7), // 6
+  modelOptions: { resolution: '1080p', generateAudio: true },
+})
+// Note: Veo result URLs require the Google API key to download
+// (x-goog-api-key header or ?key= query parameter).
+```
+
 Client hook with job tracking:
 
 ```tsx
@@ -607,7 +730,54 @@ generateSpeech({
 
 > Source: Gemini TTS adapter validation; CodeRabbit review of PR #463.
 
-### h. LOW: Writing a logging middleware to see media chunks flow through
+### h. HIGH: Passing image prompt parts to a model that doesn't support image-conditioned generation
+
+Not every model accepts image-conditioned prompts. The `prompt` type is
+narrowed per model, so passing an image part to a text-only model
+(dall-e-3, Imagen, grok-2-image) is a **compile-time error**; adapters
+also throw a clear runtime error as a backstop, so users learn at call
+time rather than getting silently wrong output.
+
+```typescript
+// WRONG — dall-e-3 has no edit/inputs API; image parts are a type error
+generateImage({
+  adapter: openaiImage('dall-e-3'),
+  prompt: [
+    { type: 'text', content: 'Edit this' },
+    { type: 'image', source: { type: 'url', value: url } }, // ❌ type error
+  ],
+})
+
+// WRONG — Imagen is text-to-image only; same compile-time rejection
+generateImage({
+  adapter: geminiImage('imagen-4.0-generate-001'),
+  prompt: [
+    { type: 'text', content: 'Edit this' },
+    { type: 'image', source: { type: 'url', value: url } }, // ❌ type error
+  ],
+})
+
+// CORRECT — use a model that supports image-conditioned generation
+generateImage({
+  adapter: openaiImage('gpt-image-2'), // edits up to 16 images
+  prompt: [
+    { type: 'text', content: 'Edit this' },
+    { type: 'image', source: { type: 'url', value: url } },
+  ],
+})
+
+generateImage({
+  adapter: geminiImage('gemini-3.1-flash-image-preview'), // native multimodal
+  prompt: [
+    { type: 'text', content: 'Edit this' },
+    { type: 'image', source: { type: 'url', value: url } },
+  ],
+})
+```
+
+> Source: docs/media/image-generation.md, docs/media/video-generation.md.
+
+### i. LOW: Writing a logging middleware to see media chunks flow through
 
 Every media activity — `generateAudio`, `generateSpeech`,
 `generateTranscription`, `generateImage`, `generateVideo` — accepts the

package/src/activities/chat/index.ts

@@ -33,7 +33,11 @@ import type {
   ClientToolRequest,
   ToolResult,
 } from './tools/tool-calls'
-import type { AnyTextAdapter, StructuredOutputOptions } from './adapter'
+import type {
+  AnyTextAdapter,
+  StructuredOutputOptions,
+  StructuredOutputResult,
+} from './adapter'
 import type {
   AgentLoopStrategy,
   AnyTool,
@@ -646,6 +650,7 @@ class TextEngine<
         this.deferredPromises.push(promise)
       },
       // Provider / adapter info
+      activity: 'chat',
       provider: config.adapter.name,
       model: config.params.model,
       source: 'server',
@@ -1197,6 +1202,23 @@ class TextEngine<
     }
   }
 
+  /**
+   * Tools available for execution this turn. The discovery tool is dropped
+   * from the advertised set (`this.tools`) once every lazy tool is discovered,
+   * but a model may still re-request discovery; this widens execution lookup
+   * to include it so such calls don't fail with "Unknown tool". Centralised so
+   * both execution sites (`processToolCalls` and `checkForPendingToolCalls`)
+   * stay in sync.
+   */
+  private resolveExecutableTools(
+    toolCalls: ReadonlyArray<ToolCall>,
+  ): ReadonlyArray<AnyTool> {
+    return this.lazyToolManager.getExecutableTools(
+      this.tools,
+      toolCalls.map((tc) => tc.function.name),
+    )
+  }
+
   private async *checkForPendingToolCalls(): AsyncGenerator<
     StreamChunk,
     ToolPhaseResult,
@@ -1245,7 +1267,7 @@ class TextEngine<
 
     const generator = executeToolCalls(
       executablePendingCalls,
-      this.tools,
+      this.resolveExecutableTools(executablePendingCalls),
       approvals,
       clientToolResults,
       (eventName, data) => this.createCustomEventChunk(eventName, data),
@@ -1407,7 +1429,7 @@ class TextEngine<
 
     const generator = executeToolCalls(
       executableToolCalls,
-      this.tools,
+      this.resolveExecutableTools(executableToolCalls),
       approvals,
       clientToolResults,
       (eventName, data) => this.createCustomEventChunk(eventName, data),
@@ -2861,7 +2883,7 @@ async function* fallbackStructuredOutputStream(
     timestamp,
   }
 
-  let result: { data: unknown; rawText: string }
+  let result: StructuredOutputResult<unknown>
   try {
     result = await adapter.structuredOutput(options)
   } catch (error) {
@@ -2917,6 +2939,12 @@ async function* fallbackStructuredOutputStream(
     model,
     timestamp,
     finishReason: 'stop',
+    // Forward adapter-reported token usage so consumers reading
+    // `RUN_FINISHED.usage` (and the engine's `runOnUsage` middleware hook) see
+    // it on the fallback path, mirroring the native streaming path. The
+    // conditional spread avoids emitting `usage: undefined` for adapters that
+    // don't report it. See #758.
+    ...(result.usage ? { usage: result.usage } : {}),
   }
 }
 

package/src/activities/chat/middleware/types.ts

@@ -80,6 +80,13 @@ export interface ChatMiddlewareContext<TContext = unknown> {
 
   // --- Provider / adapter info (immutable for the lifetime of the request) ---
 
+  /**
+   * Which activity this context describes — always `'chat'`. Present so the
+   * chat context structurally satisfies the base `GenerationMiddlewareContext`,
+   * letting an observe-only middleware authored against the base (e.g.
+   * `otelMiddleware`) run on both chat and media activities.
+   */
+  activity: 'chat'
   /** Provider name (e.g., 'openai', 'anthropic') */
   provider: string
   /** Model identifier (e.g., 'gpt-4o') */

package/src/activities/chat/tools/lazy-tool-manager.ts

@@ -1,7 +1,14 @@
 import { convertSchemaToJsonSchema } from './schema-converter'
-import type { Tool } from '../../../types'
+import type { AnyTool, Tool } from '../../../types'
 
-const DISCOVERY_TOOL_NAME = '__lazy__tool__discovery__'
+/**
+ * Name of the synthetic tool the LLM calls to discover lazy tools.
+ *
+ * Exported so callers building custom message-compaction / history-trimming
+ * logic can reference the discovery tool by constant instead of hard-coding
+ * the string (which is an internal contract that could change).
+ */
+export const DISCOVERY_TOOL_NAME = '__lazy__tool__discovery__'
 
 /**
  * Manages lazy tool discovery for the chat agent loop.
@@ -87,6 +94,35 @@ export class LazyToolManager {
     return active
   }
 
+  /**
+   * Returns the tools that should be available for *execution* this turn.
+   *
+   * This is the advertised set (`getActiveTools()`, passed in as `activeTools`)
+   * plus the discovery tool when a pending call references it but it is no
+   * longer advertised. Once every lazy tool has been discovered the discovery
+   * tool is dropped from the advertised set, but a model may still re-request
+   * discovery (long context / hallucination); keeping it executable lets that
+   * call return the schemas again instead of failing with "Unknown tool".
+   *
+   * The advertised set is intentionally left unchanged — only execution lookup
+   * is widened. Operates on the already-built `activeTools`: it must NOT call
+   * `getActiveTools()`, which would reset `hasNewDiscoveries` before the
+   * post-execution refresh check in the agent loop.
+   */
+  getExecutableTools(
+    activeTools: ReadonlyArray<AnyTool>,
+    pendingToolCallNames: ReadonlyArray<string>,
+  ): ReadonlyArray<AnyTool> {
+    if (
+      this.discoveryTool &&
+      pendingToolCallNames.includes(DISCOVERY_TOOL_NAME) &&
+      !activeTools.some((t) => t.name === DISCOVERY_TOOL_NAME)
+    ) {
+      return [...activeTools, this.discoveryTool]
+    }
+    return activeTools
+  }
+
   /**
    * Returns whether new tools have been discovered since the last getActiveTools() call.
    */
@@ -221,8 +257,14 @@ export class LazyToolManager {
         for (const name of args.toolNames) {
           const tool = lazyToolMap.get(name)
           if (tool) {
-            manager.discoveredTools.add(name)
-            manager.hasNewDiscoveries = true
+            // Only flag a refresh for genuinely new discoveries. Re-requesting
+            // an already-discovered tool still returns its schema below (the
+            // model asked for it), but must not trigger a redundant tool-list
+            // refresh + continue in the agent loop.
+            if (!manager.discoveredTools.has(name)) {
+              manager.discoveredTools.add(name)
+              manager.hasNewDiscoveries = true
+            }
             const jsonSchema = tool.inputSchema
               ? convertSchemaToJsonSchema(tool.inputSchema)
               : undefined

package/src/activities/generateAudio/index.ts

@@ -8,8 +8,16 @@
 import { aiEventClient } from '@tanstack/ai-event-client'
 import { streamGenerationResult } from '../stream-generation-result.js'
 import { resolveDebugOption } from '../../logger/resolve'
+import {
+  createGenerationContext,
+  runGenerationError,
+  runGenerationFinish,
+  runGenerationStart,
+  runGenerationUsage,
+} from '../middleware'
 import type { InternalLogger } from '../../logger/internal-logger'
 import type { DebugOption } from '../../logger/types'
+import type { GenerationMiddleware } from '../middleware'
 import type { AudioAdapter } from './adapter'
 import type { AudioGenerationResult, StreamChunk } from '../../types'
 
@@ -70,6 +78,12 @@ export interface AudioActivityOptions<
    * control and/or a custom `Logger`.
    */
   debug?: DebugOption
+  /**
+   * Observe-only middleware notified on start, usage, success, and error. Pass
+   * `otelMiddleware()` to emit OpenTelemetry spans, or implement the
+   * `GenerationMiddleware` contract for a custom backend.
+   */
+  middleware?: Array<GenerationMiddleware>
 }
 
 // ===========================
@@ -135,7 +149,13 @@ async function runGenerateAudio<
 >(
   options: AudioActivityOptions<TAdapter, boolean>,
 ): Promise<AudioGenerationResult> {
-  const { adapter, stream: _stream, debug: _debug, ...rest } = options
+  const {
+    adapter,
+    stream: _stream,
+    debug: _debug,
+    middleware,
+    ...rest
+  } = options
   const model = adapter.model
   const requestId = createId('audio')
   const startTime = Date.now()
@@ -145,6 +165,17 @@ async function runGenerateAudio<
     (adapter as { name?: string }).name ??
     'unknown'
 
+  const mwCtx = createGenerationContext({
+    requestId,
+    activity: 'audio',
+    provider: adapter.name,
+    model,
+    modelOptions: rest.modelOptions,
+    createId,
+  })
+
+  await runGenerationStart(middleware, mwCtx)
+
   aiEventClient.emit('audio:request:started', {
     requestId,
     provider: adapter.name,
@@ -189,6 +220,12 @@ async function runGenerateAudio<
       audioDuration: result.audio.duration,
     })
 
+    if (result.usage) await runGenerationUsage(middleware, mwCtx, result.usage)
+    await runGenerationFinish(middleware, mwCtx, {
+      duration: elapsedMs,
+      usage: result.usage,
+    })
+
     return result
   } catch (error) {
     const elapsedMs = Date.now() - startTime
@@ -202,6 +239,10 @@ async function runGenerateAudio<
       modelOptions: rest.modelOptions as Record<string, unknown> | undefined,
       timestamp: Date.now(),
     })
+    await runGenerationError(middleware, mwCtx, {
+      error,
+      duration: elapsedMs,
+    })
     logger.errors('generateAudio activity failed', {
       error,
       source: 'generateAudio',

package/src/activities/generateImage/adapter.ts

@@ -1,4 +1,8 @@
-import type { ImageGenerationOptions, ImageGenerationResult } from '../../types'
+import type {
+  ImageGenerationOptions,
+  ImageGenerationResult,
+  ModelInputModalitiesByName,
+} from '../../types'
 
 /**
  * Resolve the size type for a model from the model-size map.
@@ -29,6 +33,8 @@ export interface ImageAdapterConfig {
  * - TProviderOptions: Base provider-specific options (already resolved)
  * - TModelProviderOptionsByName: Map from model name to its specific provider options
  * - TModelSizeByName: Map from model name to its supported sizes
+ * - TModelInputModalitiesByName: Map from model name to the non-text prompt
+ *   modalities it accepts (constrains the `prompt` part types at compile time)
  */
 export interface ImageAdapter<
   TModel extends string = string,
@@ -38,6 +44,8 @@ export interface ImageAdapter<
     string,
     string
   >,
+  TModelInputModalitiesByName extends ModelInputModalitiesByName =
+    ModelInputModalitiesByName,
 > {
   /** Discriminator for adapter kind - used by generate() to determine API shape */
   readonly kind: 'image'
@@ -53,6 +61,7 @@ export interface ImageAdapter<
     providerOptions: TProviderOptions
     modelProviderOptionsByName: TModelProviderOptionsByName
     modelSizeByName: TModelSizeByName
+    modelInputModalitiesByName: TModelInputModalitiesByName
   }
 
   /**
@@ -67,7 +76,7 @@ export interface ImageAdapter<
  * An ImageAdapter with any/unknown type parameters.
  * Useful as a constraint in generic functions and interfaces.
  */
-export type AnyImageAdapter = ImageAdapter<any, any, any, any>
+export type AnyImageAdapter = ImageAdapter<any, any, any, any, any>
 
 /**
  * Abstract base class for image generation adapters.
@@ -83,11 +92,14 @@ export abstract class BaseImageAdapter<
     string,
     string
   >,
+  TModelInputModalitiesByName extends ModelInputModalitiesByName =
+    ModelInputModalitiesByName,
 > implements ImageAdapter<
   TModel,
   TProviderOptions,
   TModelProviderOptionsByName,
-  TModelSizeByName
+  TModelSizeByName,
+  TModelInputModalitiesByName
 > {
   readonly kind = 'image' as const
   abstract readonly name: string
@@ -98,6 +110,7 @@ export abstract class BaseImageAdapter<
     providerOptions: TProviderOptions
     modelProviderOptionsByName: TModelProviderOptionsByName
     modelSizeByName: TModelSizeByName
+    modelInputModalitiesByName: TModelInputModalitiesByName
   }
 
   protected config: ImageAdapterConfig