visual-ai-assertions 0.25.0 → 0.26.0

package/dist/index.js

@@ -1,132 +1,3 @@
-// src/constants.ts
-var ReasoningEffort = {
-  LOW: "low",
-  MEDIUM: "medium",
-  HIGH: "high",
-  XHIGH: "xhigh"
-};
-var ImageDetail = {
-  AUTO: "auto",
-  LOW: "low",
-  HIGH: "high"
-};
-var DEFAULT_IMAGE_DETAIL = ImageDetail.AUTO;
-var DEFAULT_MAX_IMAGE_DIMENSION = 1568;
-var Provider = {
-  ANTHROPIC: "anthropic",
-  OPENAI: "openai",
-  GOOGLE: "google",
-  OPENROUTER: "openrouter"
-};
-var Model = {
-  Anthropic: {
-    FABLE_5_1: "claude-fable-5-1",
-    FABLE_5: "claude-fable-5",
-    OPUS_5: "claude-opus-5",
-    OPUS_4_8: "claude-opus-4-8",
-    OPUS_4_7: "claude-opus-4-7",
-    OPUS_4_6: "claude-opus-4-6",
-    SONNET_5: "claude-sonnet-5",
-    SONNET_4_6: "claude-sonnet-4-6",
-    HAIKU_4_5: "claude-haiku-4-5"
-  },
-  OpenAI: {
-    GPT_6_ASTRA: "gpt-6-astra",
-    GPT_5_6_SOL: "gpt-5.6-sol",
-    GPT_5_6_TERRA: "gpt-5.6-terra",
-    GPT_5_6_LUNA: "gpt-5.6-luna",
-    GPT_5_5: "gpt-5.5",
-    GPT_5_4: "gpt-5.4",
-    GPT_5_4_PRO: "gpt-5.4-pro",
-    GPT_5_4_MINI: "gpt-5.4-mini",
-    GPT_5_4_NANO: "gpt-5.4-nano",
-    GPT_5_2: "gpt-5.2",
-    GPT_5_MINI: "gpt-5-mini"
-  },
-  Google: {
-    GEMINI_3_8_FLASH: "gemini-3.8-flash",
-    GEMINI_3_7_FLASH: "gemini-3.7-flash",
-    GEMINI_3_6_FLASH: "gemini-3.6-flash",
-    GEMINI_3_5_FLASH: "gemini-3.5-flash",
-    GEMINI_3_5_FLASH_LITE: "gemini-3.5-flash-lite",
-    GEMINI_3_1_PRO_PREVIEW: "gemini-3.1-pro-preview",
-    GEMINI_3_1_FLASH_LITE: "gemini-3.1-flash-lite",
-    GEMINI_3_FLASH_PREVIEW: "gemini-3-flash-preview"
-  },
-  /**
-   * Models routed through OpenRouter (https://openrouter.ai). Slugs always
-   * carry a vendor prefix (`vendor/model`), which is how provider inference
-   * recognizes them. All listed models accept image input.
-   */
-  OpenRouter: {
-    MUSE_SPARK_1_3: "meta/muse-spark-1.3",
-    GROK_4_6: "x-ai/grok-4.6",
-    GROK_4_5: "x-ai/grok-4.5",
-    KIMI_K3: "moonshotai/kimi-k3",
-    KIMI_K2_7_CODE: "moonshotai/kimi-k2.7-code",
-    QWEN_3_8_MAX: "qwen/qwen3.8-max",
-    QWEN_3_7_PLUS: "qwen/qwen3.7-plus",
-    QWEN_3_6_FLASH: "qwen/qwen3.6-flash",
-    GLM_5_3_FLASH: "z-ai/glm-5.3-flash"
-  }
-};
-var DEFAULT_MODELS = {
-  [Provider.ANTHROPIC]: Model.Anthropic.SONNET_4_6,
-  [Provider.OPENAI]: Model.OpenAI.GPT_5_6_LUNA,
-  [Provider.GOOGLE]: Model.Google.GEMINI_3_FLASH_PREVIEW,
-  [Provider.OPENROUTER]: Model.OpenRouter.QWEN_3_6_FLASH
-};
-var DEFAULT_MAX_TOKENS = 4096;
-var OPENAI_REASONING_MAX_TOKENS = 16384;
-var OPENAI_HEAVY_REASONING_MAX_TOKENS = 32768;
-var MODELS_REQUIRING_LARGE_OUTPUT_BUDGET = /* @__PURE__ */ new Set([
-  Model.OpenAI.GPT_6_ASTRA
-]);
-var MODEL_TO_PROVIDER = new Map([
-  ...Object.values(Model.Anthropic).map((m) => [m, Provider.ANTHROPIC]),
-  ...Object.values(Model.OpenAI).map((m) => [m, Provider.OPENAI]),
-  ...Object.values(Model.Google).map((m) => [m, Provider.GOOGLE]),
-  ...Object.values(Model.OpenRouter).map((m) => [m, Provider.OPENROUTER])
-]);
-var VALID_PROVIDERS = Object.values(Provider);
-var PROVIDER_DEFAULT_REASONING = {
-  openai: "medium",
-  anthropic: "off",
-  google: "off",
-  // Varies by upstream model; the driver sends no reasoning field unless configured.
-  openrouter: "off"
-};
-var Content = {
-  /** Detects Lorem ipsum, TODO, TBD, and similar placeholder text */
-  PLACEHOLDER_TEXT: "placeholder-text",
-  /** Detects error messages, banners, stack traces, or error codes */
-  ERROR_MESSAGES: "error-messages",
-  /** Detects broken image icons or failed-to-load image indicators */
-  BROKEN_IMAGES: "broken-images",
-  /** Detects UI elements that unintentionally overlap and obscure content */
-  OVERLAPPING_ELEMENTS: "overlapping-elements"
-};
-var Layout = {
-  /** Detects elements that unintentionally overlap each other */
-  OVERLAP: "overlap",
-  /** Detects content cut off or extending beyond container boundaries */
-  OVERFLOW: "overflow",
-  /** Detects inconsistent alignment of text, images, and UI components */
-  ALIGNMENT: "alignment"
-};
-var Accessibility = {
-  /** Detects insufficient color contrast between text and backgrounds */
-  CONTRAST: "contrast",
-  /** Detects text that is cut off, overlapping, too small, or obscured */
-  READABILITY: "readability",
-  /** Detects interactive elements that are not visually distinct */
-  INTERACTIVE_VISIBILITY: "interactive-visibility",
-  /** Detects color choices likely to be indistinguishable to viewers with common color vision deficiencies */
-  COLOR_BLINDNESS: "color-blindness",
-  /** Detects information conveyed by color alone, without a non-color cue (icon, text, pattern, position) */
-  COLOR_ALONE: "color-alone"
-};
-
 // src/errors.ts
 var VisualAIError = class extends Error {
   code;
@@ -493,7 +364,7 @@ function visibleRole(finalState, requireCorrectRendering) {
 var ELEMENTS_VISIBLE_CLIPPING_RULES = [
   "When an element is partly rendered but cut off at an edge, decide whether ordinary scrolling would bring it fully into view. For example, a card peeking past the end of a horizontal carousel, a filter chip in a row that continues past the screen edge, or a list item partly below the bottom of a scrolling feed is reachable that way, so the check for that element PASSES. Say in your reasoning that it is reached by scrolling.",
   "An element that scrolling cannot bring into view is NOT properly visible: one sliced by the screen edge itself, or cut off or overlapped by fixed chrome such as the status bar, a notch, a home indicator, a sticky header, or a fixed bottom navigation bar. That is a layout fault, so the check for that element FAILS. Describe the clipping in your reasoning.",
-  "An element you cannot see at all is not visible, even if the page might reveal it after scrolling. Judge only what this screenshot actually shows."
+  "The scrolling allowance above applies only to elements that are at least partly rendered. If no part of an element is on screen, the check for that element FAILS: do not infer that it exists below the fold. Judge only what this screenshot actually shows."
 ];
 var ELEMENTS_VISIBLE_FINAL_STATE_RULE = "Judge each element in its finished, presented state. Things a design draws on top of an element \u2014 a badge, a favourite icon, a duration or price pill, a gradient scrim \u2014 coexist with finished content and leave it visible. An overlay that says the element is NOT ready \u2014 a loading spinner, a skeleton placeholder, a shimmer, a progress bar, an error or retry overlay \u2014 means the element is not properly visible even when you can still make out what sits underneath, so the check for that element FAILS. Name which of the two you are seeing in your reasoning.";
 var ELEMENTS_VISIBLE_CORRECT_RENDERING_RULE = "An element that is present but clearly defective in how it is rendered is NOT properly visible: text at contrast too low to read, elements overlapping or colliding with one another, an element visibly out of alignment with the siblings it should line up with, or text cut off mid-word inside its own container. The check for that element FAILS. In your reasoning, say that the element is present and then name the defect. Only clear, unambiguous defects count: do not fail an element for tight spacing, stylistic choices, or anything you would have to argue for.";
@@ -524,6 +395,144 @@ function buildElementsVisibilityPrompt(elements, visible, options) {
   return buildCheckPrompt(statements, { role, instructions });
 }
 
+// src/constants.ts
+var ReasoningEffort = {
+  MINIMAL: "minimal",
+  LOW: "low",
+  MEDIUM: "medium",
+  HIGH: "high",
+  XHIGH: "xhigh"
+};
+var ImageDetail = {
+  AUTO: "auto",
+  LOW: "low",
+  HIGH: "high"
+};
+var DEFAULT_IMAGE_DETAIL = ImageDetail.AUTO;
+var DEFAULT_MAX_IMAGE_DIMENSION = 1568;
+var Provider = {
+  ANTHROPIC: "anthropic",
+  OPENAI: "openai",
+  GOOGLE: "google",
+  OPENROUTER: "openrouter"
+};
+var Model = {
+  Anthropic: {
+    FABLE_5_1: "claude-fable-5-1",
+    FABLE_5: "claude-fable-5",
+    OPUS_5_5: "claude-opus-5-5",
+    OPUS_5: "claude-opus-5",
+    OPUS_4_8: "claude-opus-4-8",
+    OPUS_4_7: "claude-opus-4-7",
+    OPUS_4_6: "claude-opus-4-6",
+    SONNET_5_5: "claude-sonnet-5-5",
+    SONNET_5: "claude-sonnet-5",
+    SONNET_4_6: "claude-sonnet-4-6",
+    HAIKU_4_5: "claude-haiku-4-5"
+  },
+  OpenAI: {
+    GPT_6_ASTRA: "gpt-6-astra",
+    GPT_6_1_SOL: "gpt-6.1-sol",
+    GPT_6_SOL: "gpt-6-sol",
+    GPT_6_LUNA: "gpt-6-luna",
+    GPT_5_6_SOL: "gpt-5.6-sol",
+    GPT_5_6_TERRA: "gpt-5.6-terra",
+    GPT_5_6_LUNA: "gpt-5.6-luna",
+    GPT_5_5: "gpt-5.5",
+    GPT_5_4: "gpt-5.4",
+    GPT_5_4_PRO: "gpt-5.4-pro",
+    GPT_5_4_MINI: "gpt-5.4-mini",
+    GPT_5_4_NANO: "gpt-5.4-nano",
+    GPT_5_2: "gpt-5.2",
+    GPT_5_MINI: "gpt-5-mini"
+  },
+  Google: {
+    GEMINI_3_8_FLASH: "gemini-3.8-flash",
+    GEMINI_3_7_FLASH: "gemini-3.7-flash",
+    GEMINI_3_6_FLASH: "gemini-3.6-flash",
+    GEMINI_3_5_FLASH: "gemini-3.5-flash",
+    GEMINI_3_5_FLASH_LITE: "gemini-3.5-flash-lite",
+    GEMINI_3_1_PRO_PREVIEW: "gemini-3.1-pro-preview",
+    GEMINI_3_1_FLASH_LITE: "gemini-3.1-flash-lite",
+    GEMINI_3_FLASH_PREVIEW: "gemini-3-flash-preview"
+  },
+  /**
+   * Models routed through OpenRouter (https://openrouter.ai). Slugs always
+   * carry a vendor prefix (`vendor/model`), which is how provider inference
+   * recognizes them. All listed models accept image input.
+   */
+  OpenRouter: {
+    MUSE_SPARK_1_3: "meta/muse-spark-1.3",
+    GROK_4_7: "x-ai/grok-4.7",
+    GROK_4_6: "x-ai/grok-4.6",
+    GROK_4_5: "x-ai/grok-4.5",
+    KIMI_K3: "moonshotai/kimi-k3",
+    KIMI_K2_7_CODE: "moonshotai/kimi-k2.7-code",
+    QWEN_3_8_MAX: "qwen/qwen3.8-max",
+    QWEN_3_7_PLUS: "qwen/qwen3.7-plus",
+    QWEN_3_6_FLASH: "qwen/qwen3.6-flash",
+    GLM_5_3_FLASH: "z-ai/glm-5.3-flash",
+    MIMO_V2_6_PRO: "xiaomi/mimo-v2.6-pro"
+  }
+};
+var DEFAULT_MODELS = {
+  [Provider.ANTHROPIC]: Model.Anthropic.SONNET_5_5,
+  [Provider.OPENAI]: Model.OpenAI.GPT_6_1_SOL,
+  [Provider.GOOGLE]: Model.Google.GEMINI_3_8_FLASH,
+  [Provider.OPENROUTER]: Model.OpenRouter.MUSE_SPARK_1_3
+};
+var DEFAULT_MAX_TOKENS = 4096;
+var OPENAI_REASONING_MAX_TOKENS = 16384;
+var OPENAI_HEAVY_REASONING_MAX_TOKENS = 32768;
+var MODELS_REQUIRING_LARGE_OUTPUT_BUDGET = /* @__PURE__ */ new Set([
+  Model.OpenRouter.QWEN_3_8_MAX,
+  Model.OpenRouter.QWEN_3_7_PLUS
+]);
+var MODEL_TO_PROVIDER = new Map([
+  ...Object.values(Model.Anthropic).map((m) => [m, Provider.ANTHROPIC]),
+  ...Object.values(Model.OpenAI).map((m) => [m, Provider.OPENAI]),
+  ...Object.values(Model.Google).map((m) => [m, Provider.GOOGLE]),
+  ...Object.values(Model.OpenRouter).map((m) => [m, Provider.OPENROUTER])
+]);
+var VALID_PROVIDERS = Object.values(Provider);
+var PROVIDER_DEFAULT_REASONING = {
+  openai: "medium",
+  anthropic: "off",
+  google: "off",
+  // Varies by upstream model; the driver sends no reasoning field unless configured.
+  openrouter: "off"
+};
+var Content = {
+  /** Detects Lorem ipsum, TODO, TBD, and similar placeholder text */
+  PLACEHOLDER_TEXT: "placeholder-text",
+  /** Detects error messages, banners, stack traces, or error codes */
+  ERROR_MESSAGES: "error-messages",
+  /** Detects broken image icons or failed-to-load image indicators */
+  BROKEN_IMAGES: "broken-images",
+  /** Detects UI elements that unintentionally overlap and obscure content */
+  OVERLAPPING_ELEMENTS: "overlapping-elements"
+};
+var Layout = {
+  /** Detects elements that unintentionally overlap each other */
+  OVERLAP: "overlap",
+  /** Detects content cut off or extending beyond container boundaries */
+  OVERFLOW: "overflow",
+  /** Detects inconsistent alignment of text, images, and UI components */
+  ALIGNMENT: "alignment"
+};
+var Accessibility = {
+  /** Detects insufficient color contrast between text and backgrounds */
+  CONTRAST: "contrast",
+  /** Detects text that is cut off, overlapping, too small, or obscured */
+  READABILITY: "readability",
+  /** Detects interactive elements that are not visually distinct */
+  INTERACTIVE_VISIBILITY: "interactive-visibility",
+  /** Detects color choices likely to be indistinguishable to viewers with common color vision deficiencies */
+  COLOR_BLINDNESS: "color-blindness",
+  /** Detects information conveyed by color alone, without a non-color cue (icon, text, pattern, position) */
+  COLOR_ALONE: "color-alone"
+};
+
 // src/templates/accessibility.ts
 var ALL_CHECKS = Object.values(Accessibility);
 var ACCESSIBILITY_ROLE = "Evaluate this screenshot for visual accessibility. Focus on what you can actually perceive \u2014 apparent contrast levels, text legibility, and visual distinctiveness of interactive elements.";
@@ -633,17 +642,22 @@ function parseRetryAfter(value) {
 var XHIGH_CAPABLE_MODELS = /* @__PURE__ */ new Set([
   Model.Anthropic.FABLE_5_1,
   Model.Anthropic.FABLE_5,
+  Model.Anthropic.OPUS_5_5,
   Model.Anthropic.OPUS_5,
   Model.Anthropic.OPUS_4_8,
   Model.Anthropic.OPUS_4_7,
+  Model.Anthropic.SONNET_5_5,
   Model.Anthropic.SONNET_5
 ]);
 function mapEffort(level, model) {
+  if (level === "minimal") return "low";
   if (level !== "xhigh") return level;
   return XHIGH_CAPABLE_MODELS.has(model) ? "xhigh" : "max";
 }
 var BUDGET_THINKING_MODELS = /* @__PURE__ */ new Set([Model.Anthropic.HAIKU_4_5]);
 var EFFORT_TO_BUDGET_TOKENS = {
+  // 1024 is Anthropic's minimum thinking budget, so minimal and low coincide.
+  minimal: 1024,
   low: 1024,
   medium: 4096,
   high: 8192,
@@ -760,6 +774,9 @@ function sleep(ms) {
   return new Promise((resolve2) => setTimeout(resolve2, ms));
 }
 var GOOGLE_THINKING_LEVEL = {
+  // Gemini does define a "minimal" thinking level, but some models reject it
+  // (e.g. Gemini 3.1 Pro), so "minimal" clamps to "low" here as well.
+  minimal: "low",
   low: "low",
   medium: "medium",
   high: "high",
@@ -1077,6 +1094,9 @@ var OpenAIDriver = class {
 // src/providers/openrouter.ts
 var OPENROUTER_BASE_URL = "https://openrouter.ai/api/v1";
 var OPENROUTER_REASONING_EFFORT = {
+  // OpenRouter normalizes upstream vendors to low/medium/high only, so
+  // "minimal" has no native equivalent and clamps to the floor.
+  minimal: "low",
   low: "low",
   medium: "medium",
   high: "high",
@@ -1236,10 +1256,20 @@ function parseBooleanEnv(envName, value) {
     `Invalid ${envName} value: "${value}". Use "true", "1", "false", or "0".`
   );
 }
+function parseReasoningEffortEnv(envName, value) {
+  if (value === void 0 || value === "") return void 0;
+  const levels = Object.values(ReasoningEffort);
+  const lower = value.toLowerCase();
+  if (levels.includes(lower)) return lower;
+  throw new VisualAIConfigError(
+    `Invalid ${envName} value: "${value}". Use one of: ${levels.join(", ")}.`
+  );
+}
 var debugDeprecationWarned = false;
 function resolveConfig(config) {
   const provider = resolveProvider(config);
   const model = config.model ?? process.env.VISUAL_AI_MODEL ?? DEFAULT_MODELS[provider];
+  const reasoningEffort = config.reasoningEffort ?? parseReasoningEffortEnv("VISUAL_AI_REASONING_EFFORT", process.env.VISUAL_AI_REASONING_EFFORT);
   const debug = config.debug ?? parseBooleanEnv("VISUAL_AI_DEBUG", process.env.VISUAL_AI_DEBUG) ?? false;
   const debugPrompt = config.debugPrompt ?? parseBooleanEnv("VISUAL_AI_DEBUG_PROMPT", process.env.VISUAL_AI_DEBUG_PROMPT) ?? false;
   const debugResponse = config.debugResponse ?? parseBooleanEnv("VISUAL_AI_DEBUG_RESPONSE", process.env.VISUAL_AI_DEBUG_RESPONSE) ?? false;
@@ -1257,12 +1287,12 @@ function resolveConfig(config) {
   }
   const userSetMaxTokens = config.maxTokens !== void 0;
   let maxTokens = config.maxTokens ?? DEFAULT_MAX_TOKENS;
-  const effortNeedsLargeBudget = config.reasoningEffort === "high" || config.reasoningEffort === "xhigh";
+  const effortNeedsLargeBudget = reasoningEffort === "high" || reasoningEffort === "xhigh";
   const modelNeedsLargeBudget = MODELS_REQUIRING_LARGE_OUTPUT_BUDGET.has(model);
   if (!userSetMaxTokens && (provider === "openai" || provider === "openrouter") && (effortNeedsLargeBudget || modelNeedsLargeBudget)) {
     maxTokens = modelNeedsLargeBudget ? OPENAI_HEAVY_REASONING_MAX_TOKENS : OPENAI_REASONING_MAX_TOKENS;
     if (debug) {
-      const reason = modelNeedsLargeBudget ? `model "${model}", which exhausts smaller budgets on reasoning at any effort` : `provider "${provider}" with reasoningEffort "${config.reasoningEffort}"`;
+      const reason = modelNeedsLargeBudget ? `model "${model}", which exhausts smaller budgets on reasoning at any effort` : `provider "${provider}" with reasoningEffort "${reasoningEffort}"`;
       process.stderr.write(
         `[visual-ai-assertions] Auto-increased maxTokens from ${DEFAULT_MAX_TOKENS} to ${maxTokens} for ${reason}.
 `
@@ -1274,7 +1304,7 @@ function resolveConfig(config) {
     apiKey: config.apiKey,
     model,
     maxTokens,
-    reasoningEffort: config.reasoningEffort,
+    reasoningEffort,
     maxImageDimension: config.maxImageDimension ?? DEFAULT_MAX_IMAGE_DIMENSION,
     imageDetail: config.imageDetail ?? DEFAULT_IMAGE_DETAIL,
     timeout: config.timeout,
@@ -1297,6 +1327,10 @@ var PRICING_TABLE = {
     inputPricePerToken: 10 / PER_MILLION,
     outputPricePerToken: 50 / PER_MILLION
   },
+  [`${Provider.ANTHROPIC}:${Model.Anthropic.OPUS_5_5}`]: {
+    inputPricePerToken: 4 / PER_MILLION,
+    outputPricePerToken: 20 / PER_MILLION
+  },
   [`${Provider.ANTHROPIC}:${Model.Anthropic.OPUS_5}`]: {
     inputPricePerToken: 5 / PER_MILLION,
     outputPricePerToken: 25 / PER_MILLION
@@ -1305,6 +1339,10 @@ var PRICING_TABLE = {
     inputPricePerToken: 5 / PER_MILLION,
     outputPricePerToken: 25 / PER_MILLION
   },
+  [`${Provider.ANTHROPIC}:${Model.Anthropic.SONNET_5_5}`]: {
+    inputPricePerToken: 2 / PER_MILLION,
+    outputPricePerToken: 10 / PER_MILLION
+  },
   [`${Provider.ANTHROPIC}:${Model.Anthropic.SONNET_5}`]: {
     inputPricePerToken: 3 / PER_MILLION,
     outputPricePerToken: 15 / PER_MILLION
@@ -1331,6 +1369,22 @@ var PRICING_TABLE = {
     inputPricePerToken: 10 / PER_MILLION,
     outputPricePerToken: 50 / PER_MILLION
   },
+  // Cached input is $0.10/MTok (not modelled), half GPT-6 Sol's cached rate.
+  [`${Provider.OPENAI}:${Model.OpenAI.GPT_6_1_SOL}`]: {
+    inputPricePerToken: 2 / PER_MILLION,
+    outputPricePerToken: 10 / PER_MILLION
+  },
+  // Cached input is $0.20/MTok (not modelled). Prompts above 272K input tokens
+  // bill at 2x input / 1.5x output, which is far beyond screenshot-sized calls.
+  [`${Provider.OPENAI}:${Model.OpenAI.GPT_6_SOL}`]: {
+    inputPricePerToken: 2 / PER_MILLION,
+    outputPricePerToken: 10 / PER_MILLION
+  },
+  // Cached input is $0.01/MTok and cache writes $0.125/MTok; neither is modelled.
+  [`${Provider.OPENAI}:${Model.OpenAI.GPT_6_LUNA}`]: {
+    inputPricePerToken: 0.1 / PER_MILLION,
+    outputPricePerToken: 0.5 / PER_MILLION
+  },
   [`${Provider.OPENAI}:${Model.OpenAI.GPT_5_6_SOL}`]: {
     inputPricePerToken: 5 / PER_MILLION,
     outputPricePerToken: 30 / PER_MILLION
@@ -1427,6 +1481,10 @@ var PRICING_TABLE = {
     inputPricePerToken: 0.1 / PER_MILLION,
     outputPricePerToken: 0.2 / PER_MILLION
   },
+  [`${Provider.OPENROUTER}:${Model.OpenRouter.GROK_4_7}`]: {
+    inputPricePerToken: 1.6 / PER_MILLION,
+    outputPricePerToken: 4.8 / PER_MILLION
+  },
   [`${Provider.OPENROUTER}:${Model.OpenRouter.GROK_4_6}`]: {
     inputPricePerToken: 2 / PER_MILLION,
     outputPricePerToken: 6 / PER_MILLION
@@ -1460,6 +1518,13 @@ var PRICING_TABLE = {
   [`${Provider.OPENROUTER}:${Model.OpenRouter.GLM_5_3_FLASH}`]: {
     inputPricePerToken: 0.15 / PER_MILLION,
     outputPricePerToken: 0.5 / PER_MILLION
+  },
+  // Verified 2026-09-23 against https://openrouter.ai/api/v1/models; both
+  // upstream endpoints (Xiaomi, DeepInfra) charge the same rate. Cached input
+  // is $0.0036/MTok, not modelled (no provider gets a cache discount here).
+  [`${Provider.OPENROUTER}:${Model.OpenRouter.MIMO_V2_6_PRO}`]: {
+    inputPricePerToken: 0.435 / PER_MILLION,
+    outputPricePerToken: 0.87 / PER_MILLION
   }
 };
 function calculateCost(provider, model, inputTokens, outputTokens) {
@@ -2481,6 +2546,12 @@ function stripCodeFences(text) {
 }
 var CheckResponseSchema = CheckResultSchema.omit({ usage: true });
 var AskResponseSchema = AskResultSchema.omit({ usage: true });
+var AskImageResponseSchema = AskResponseSchema.omit({
+  frameReferences: true,
+  timestampReferences: true
+});
+var AskFramesResponseSchema = AskResponseSchema.omit({ timestampReferences: true });
+var AskNativeVideoResponseSchema = AskResponseSchema.omit({ frameReferences: true });
 var CompareResponseSchema = CompareResultSchema.omit({ usage: true });
 var STRAY_CONTROL_CHARS = /[\u0000-\u0008\u000b\u000c\u000e-\u001f\u007f]/g;
 function parseJson(text) {
@@ -2573,7 +2644,11 @@ function createDriver(provider, config) {
   return PROVIDER_REGISTRY[provider](config);
 }
 var checkSchemaOptions = toSchemaOptions(CheckResponseSchema);
-var askSchemaOptions = toSchemaOptions(AskResponseSchema);
+var askSchemaOptionsByMedia = {
+  image: toSchemaOptions(AskImageResponseSchema),
+  video: toSchemaOptions(AskFramesResponseSchema),
+  "native-video": toSchemaOptions(AskNativeVideoResponseSchema)
+};
 var compareSchemaOptions = toSchemaOptions(CompareResponseSchema);
 function mediaToProviderInputs(media) {
   if (media.kind === "image") {
@@ -2718,7 +2793,12 @@ function visualAI(config = {}) {
           media: dispatch.mediaContext
         });
         debugLog(resolvedConfig, "ask prompt", prompt, "prompt");
-        const { response, metadata } = await sendMedia(driver, dispatch, prompt, askSchemaOptions);
+        const { response, metadata } = await sendMedia(
+          driver,
+          dispatch,
+          prompt,
+          askSchemaOptionsByMedia[dispatch.mediaContext.kind]
+        );
         debugLog(resolvedConfig, "ask response", response.text, "response");
         const result = parseAskResponse(response.text);
         return {
@@ -2741,7 +2821,7 @@ function visualAI(config = {}) {
         debugLog(resolvedConfig, "compare prompt", prompt, "prompt");
         const response = await timedSendMessage(driver, [imgA, imgB], prompt, compareSchemaOptions);
         debugLog(resolvedConfig, "compare response", response.text, "response");
-        const supportsAnnotatedDiff = resolvedConfig.provider === "google" && resolvedConfig.model === Model.Google.GEMINI_3_FLASH_PREVIEW;
+        const supportsAnnotatedDiff = resolvedConfig.provider === "google" && DIFF_ALLOWED_MODELS.has(resolvedConfig.model);
         const effectiveDiffImage = options?.diffImage ?? (supportsAnnotatedDiff ? true : false);
         let diffImage;
         if (effectiveDiffImage) {