Fix | OpenRouter Video Vision for Supported Models

Author: Bredrumb

src/commands/config/model/text.ts

@@ -54,7 +54,7 @@ const MODEL_SELECT_ID = "model_select";
  * Helper function to get localized LLM description based on user's locale
  * @param model - LLM model row from database
  * @param locale - User's preferred locale (e.g., "ja", "en-US")
- * @returns Localized description with flags prepended (e.g., "(FREE+TOOLS) Description")
+ * @returns Localized description with flags prepended (e.g., "(FREE+TOOLS+IMG+VID) Description")
  */
 function getLocalizedDescription(model: LlmRow, locale: string): string {
 	// Normalize locale to handle variations (e.g., "ja-JP" -> "ja")
@@ -72,7 +72,7 @@ function getLocalizedDescription(model: LlmRow, locale: string): string {
 	const baseDescription =
 		description || model.llm_description || `${model.llm_provider} model`;
 
-	// Skip flags for account-setting (don't show TOOLS+IMAGES+etc. for this special model)
+	// Skip flags for account-setting (don't show TOOLS+IMG+VID+etc. for this special model)
 	if (model.llm_codename === "account-setting") {
 		return baseDescription;
 	}
@@ -81,7 +81,8 @@ function getLocalizedDescription(model: LlmRow, locale: string): string {
 	const flags: string[] = [];
 	if (model.is_free) flags.push("FREE");
 	if (model.has_tools) flags.push("TOOLS");
-	if (model.sees_images) flags.push("IMAGES");
+	if (model.sees_images) flags.push("IMG");
+	if (model.sees_videos) flags.push("VID");
 	if (model.supports_structoutput) flags.push("STRUCT");
 	//if (model.is_uncensored) flags.push("UNCENSORED");
 

src/commands/tool/estimate/cost.ts

@@ -1054,6 +1054,7 @@ async function measureOpenRouterInputTokens(
 	const messages = await adapter.buildProbeMessages(
 		contextItems,
 		providerConfig.seesImages ?? true,
+		providerConfig.seesVideos ?? false,
 	);
 	const requestBody = buildOpenRouterProbeRequest(providerConfig, messages);
 

src/db/seed.sql

@@ -43,8 +43,8 @@ VALUES
   ('openrouter', 'tngtech/deepseek-r1t2-chimera', false, false, true, false, false, true, false, false, false, true, true, 'Advanced Chimera DeepSeek model that is great at role-playing', 'ロールプレイに優れた高度なChimera DeepSeekモデル'),
   ('openrouter', 'x-ai/grok-4-fast', false, false, true, true, false, true, true, false, false, false, true, 'Fast and efficient general-purpose model', '高速かつ効率的な汎用モデル'),
   ('openrouter', 'x-ai/grok-4.1-fast', false, false, true, false, false, true, true, false, false, false, true, 'Latest fast and efficient general-purpose model', '高速かつ効率的な汎用モデル'),
-  ('openrouter', 'google/gemini-3-flash-preview', false, false, false, false, false, true, true, false, false, false, true, 'Latest Gemini 3 Flash preview via OpenRouter with tool use and image understanding', 'OpenRouter経由でツール利用と画像理解に対応した最新のGemini 3 Flashプレビュー'),
-  ('openrouter', 'google/gemini-3-pro-preview', false, false, false, false, false, true, true, false, false, false, true, 'Latest Gemini 3 Pro preview via OpenRouter (same capabilities as Gemini 3 Flash)', 'OpenRouter経由の最新Gemini 3 Proプレビュー（Gemini 3 Flashと同等の機能）'),
+  ('openrouter', 'google/gemini-3-flash-preview', false, false, false, false, false, true, true, true, false, false, true, 'Latest Gemini 3 Flash preview via OpenRouter with tool use and image understanding', 'OpenRouter経由でツール利用と画像理解に対応した最新のGemini 3 Flashプレビュー'),
+  ('openrouter', 'google/gemini-3-pro-preview', false, false, false, false, false, true, true, true, false, false, true, 'Latest Gemini 3 Pro preview via OpenRouter (same capabilities as Gemini 3 Flash)', 'OpenRouter経由の最新Gemini 3 Proプレビュー（Gemini 3 Flashと同等の機能）'),
   ('openrouter', 'anthropic/claude-sonnet-4.5', false, false, false, false, false, true, true, false, false, false, true, 'State-of-the-art performance in complex tasks and problems, also great in role-playing and creative writing', '複雑なタスクや問題に優れた最先端性能を持ち、ロールプレイや創作にも秀でたモデル'),
   ('openrouter', 'anthropic/claude-haiku-4.5', false, false, false, false, false, true, true, false, false, false, true, 'Lightweight version of claude-sonnet-4.5', 'claude-sonnet-4.5の軽量版'),
   ('openrouter', 'openai/gpt-5.1', true, false, true, true, false, true, true, false, false, false, true, 'State-of-the-art performance in complex tasks and problems', '複雑なタスクや問題に優れた最先端性能'),
@@ -58,9 +58,9 @@ VALUES
   ('openrouter', 'mistralai/mistral-small-3.1-24b-instruct:free', false, false, false, true, true, true, true, false, false, false, false, 'Free multimodal model with enhanced reasoning and vision capabilities', '強化された推論とビジョン機能を備えた無料のマルチモーダルモデル'),
   ('openrouter', 'z-ai/glm-4.5-air:free', false, true, false, false, true, true, false, false, false, false, false, 'Free lightweight model with thinking mode for reasoning and agent tasks', '推論とエージェントタスク向けのシンキングモードを備えた無料軽量モデル'),
   ('openrouter', 'tngtech/tng-r1t-chimera:free', false, false, false, true, true, true, false, false, false, false, false, 'Free experimental model for creative storytelling and character interaction', '創作とキャラクター対話に特化した無料の実験モデル'),
-  ('openrouter', 'qwen/qwen3.5-35b-a3b', false, false, false, false, false, true, true, false, false, false, true, 'Qwen 3.5 35B A3B model with tool use, vision, and structured output support', 'ツール利用・画像理解・構造化出力に対応したQwen 3.5 35B A3Bモデル'),
-  ('openrouter', 'qwen/qwen3.5-27b', false, false, false, false, false, true, true, false, false, false, true, 'Qwen 3.5 27B model with tool use, vision, and structured output support', 'ツール利用・画像理解・構造化出力に対応したQwen 3.5 27Bモデル'),
-  ('openrouter', 'qwen/qwen3.5-flash-02-23', false, false, false, false, false, true, true, false, false, false, true, 'Fast Qwen 3.5 Flash (02-23) model with tool use, vision, and structured output support', 'ツール利用・画像理解・構造化出力に対応した高速Qwen 3.5 Flash（02-23）モデル'),
+  ('openrouter', 'qwen/qwen3.5-35b-a3b', false, false, false, false, false, true, true, true, false, false, true, 'Qwen 3.5 35B A3B model with tool use, vision, and structured output support', 'ツール利用・画像理解・構造化出力に対応したQwen 3.5 35B A3Bモデル'),
+  ('openrouter', 'qwen/qwen3.5-27b', false, false, false, false, false, true, true, true, false, false, true, 'Qwen 3.5 27B model with tool use, vision, and structured output support', 'ツール利用・画像理解・構造化出力に対応したQwen 3.5 27Bモデル'),
+  ('openrouter', 'qwen/qwen3.5-flash-02-23', false, false, false, false, false, true, true, true, false, false, true, 'Fast Qwen 3.5 Flash (02-23) model with tool use, vision, and structured output support', 'ツール利用・画像理解・構造化出力に対応した高速Qwen 3.5 Flash（02-23）モデル'),
   ('openrouter', 'aion-labs/aion-2.0', false, false, false, false, false, false, false, false, false, false, false, 'Cheap role-play fine-tune of DeepSeek with no tools, vision, or structured output support', 'ツール・画像理解・構造化出力に対応しない、DeepSeekベースの低コストなロールプレイ特化ファインチューニングモデル'),
   ('openrouter', 'account-setting', false, false, false, false, false, true, true, true, true, false, true, 'Advanced: Uses your OpenRouter account default model', '上級者向け：OpenRouterアカウントのデフォルトモデルを使用'),
   -- Custom Provider Bootstrap Entry (allows "custom" to appear in provider dropdown)

src/locales/en-US.ts

@@ -1386,7 +1386,8 @@ You may opt out of my Memory features by using the {personalPrivacy} command, as
 				openrouter_model_selection_description: `OpenRouter offers access to many different AI models.
 	- Currently available models are based on popularity and performance, with tags for distinction:
 	  - (TOOLS) = Supports tool usage (web search, self-learning, stickers, etc.)
-	  - (IMAGES) = Sees images
+	  - (IMG) = Sees images
+	  - (VID) = Sees videos
 	  - (STRUCT) = Supports structured output (needed for persona generation and expression initialization)
 	  - (FREE) = No cost, but may have rate limits
 	- If you can't find what you want, try using the \`account-setting\` provider option

src/locales/ja.ts

@@ -1387,7 +1387,8 @@ TomoriBotをセットアップすることで、あなたとサーバーメン
 				openrouter_model_selection_description: `OpenRouterは多くの異なるAIモデルへのアクセスを提供します。
 	- 現在利用可能なモデルは人気と性能に基づいており、区別のためのタグが付いています：
 	  - (TOOLS) = ツール使用をサポート（ウェブ検索、自己学習、ステッカーなど）
-	  - (IMAGES) = 画像を認識
+	  - (IMG) = 画像を認識
+	  - (VID) = 動画を認識
 	  - (STRUCT) = 構造化出力をサポート（ペルソナ生成や表情の初期化に必要）
 	  - (FREE) = 無料ですが、レート制限がある場合があります
 	- 希望のモデルが見つからない場合は、\`account-setting\`プロバイダーオプションを試してみてください

src/providers/openrouter/openrouterProvider.ts

@@ -119,6 +119,7 @@ async function getDefaultOpenrouterModel(): Promise<string> {
 export interface OpenrouterProviderConfig extends ProviderConfig {
 	// OpenRouter uses OpenAI-compatible API, simple configuration
 	seesImages?: boolean; // Whether the model supports image inputs
+	seesVideos?: boolean; // Whether the model supports video inputs
 	// Sampling parameters to control output quality
 	topP?: number; // Nucleus sampling (0.0-1.0)
 	topK?: number; // Top-k sampling
@@ -145,7 +146,7 @@ export class OpenrouterProvider extends BaseLLMProvider implements LLMProvider {
 			supportsStreaming: true,
 			supportsFunctionCalling: true,
 			supportsImages: true, // Depends on specific models
-			supportsVideos: false, // Most models don't support video yet
+			supportsVideos: true, // Supported on video-capable models (e.g. Gemini via OpenRouter)
 		};
 	}
 
@@ -489,6 +490,7 @@ export class OpenrouterProvider extends BaseLLMProvider implements LLMProvider {
 			temperature: adjustedTemperature,
 			maxOutputTokens: resolvedMaxOutputTokens,
 			seesImages: effectiveSeesImages, // Use effective value (may be overridden)
+			seesVideos: effectiveSeesVideos, // Wire through video capability flag
 			// repetitionPenalty is hardcoded as a general token repetition dampener
 			repetitionPenalty: 1.1,
 			// Conditionally include user-configured sampling params (neutral = omit entirely)

src/providers/openrouter/openrouterStreamAdapter.ts

@@ -45,6 +45,7 @@ import type {
 export interface OpenrouterStreamConfig extends StreamConfig {
 	// OpenRouter uses OpenAI-compatible config, simple structure
 	seesImages?: boolean; // Whether the model supports image inputs
+	seesVideos?: boolean; // Whether the model supports video inputs
 	// Sampling parameters to control output quality
 	topP?: number; // Nucleus sampling (0.0-1.0)
 	topK?: number; // Top-k sampling
@@ -195,12 +196,15 @@ export class OpenrouterStreamAdapter implements StreamProvider {
 	public async buildProbeMessages(
 		contextItems: StructuredContextItem[],
 		seesImages = true,
+		seesVideos = false,
 	): Promise<Array<Record<string, unknown>>> {
 		return this.assembleOpenrouterContext(
 			contextItems,
 			[],
 			undefined,
 			seesImages,
+			"Assistant",
+			seesVideos,
 		);
 	}
 
@@ -423,6 +427,7 @@ export class OpenrouterStreamAdapter implements StreamProvider {
 			context.functionInteractionHistory,
 			openrouterConfig.seesImages ?? true, // Default to true for backward compatibility
 			context.tomoriState.tomori_nickname ?? "Assistant",
+			openrouterConfig.seesVideos ?? false, // Default false — videos are strictly opt-in per model
 		);
 
 		// Ensure model is provided
@@ -2338,6 +2343,7 @@ export class OpenrouterStreamAdapter implements StreamProvider {
 		}>,
 		seesImages: boolean = true,
 		botName: string = "Assistant",
+		seesVideos: boolean = false,
 	): Promise<Array<Record<string, unknown>>> {
 		const messages: Array<Record<string, unknown>> = [];
 		const systemInstructionParts: string[] = [];
@@ -2567,8 +2573,54 @@ export class OpenrouterStreamAdapter implements StreamProvider {
 								});
 							}
 						}
+					} else if (part.type === "video") {
+						// Videos follow the same role restriction as images - only user-role messages.
+						// For assistant turns, stage in pendingBotImageParts for a synthetic user turn.
+						const videoTargetParts =
+							role === "assistant" ? pendingBotImageParts : contentParts;
+
+						if (!seesVideos) {
+							log.info(
+								`Skipping video (model doesn't support videos): ${part.uri}`,
+							);
+							continue;
+						}
+
+						try {
+							const isHttpUrl =
+								part.uri.startsWith("http://") ||
+								part.uri.startsWith("https://");
+							const isDataUrl = part.uri.startsWith("data:");
+							if (!isHttpUrl && !isDataUrl) {
+								log.warn(
+									`Skipping unsupported video URI format for OpenRouter: ${part.uri}`,
+								);
+								continue;
+							}
+
+							// OpenRouter accepts direct public URLs and data URLs for video_url.
+							// Prefer direct URLs to avoid unnecessary fetch/encode overhead.
+							videoTargetParts.push({
+								type: "video_url",
+								video_url: { url: part.uri },
+							});
+
+							if (part.isYouTubeLink) {
+								log.success(`Added YouTube video to message: ${part.uri}`);
+							} else if (isHttpUrl) {
+								log.success(`Added direct video URL to message: ${part.uri}`);
+							} else {
+								log.success(`Added video data URL to message`);
+							}
+						} catch (videoErr) {
+							log.warn(`Error processing video: ${part.uri}`, {
+								error:
+									videoErr instanceof Error
+										? videoErr.message
+										: String(videoErr),
+							});
+						}
 					}
-					// Note: OpenRouter doesn't widely support video yet, skip video parts
 				}
 
 				// Add message