@axlsdk/axl 0.23.2 → 0.23.3

package/dist/index.cjs

@@ -219,6 +219,7 @@ __export(index_exports, {
   resolveRuntime: () => resolveRuntime,
   resolveThinkingOptions: () => resolveThinkingOptions,
   stringStreamFromEvents: () => stringStreamFromEvents,
+  summarizeModelInput: () => summarizeModelInput,
   tool: () => tool,
   workflow: () => workflow,
   zodToJsonSchema: () => zodToJsonSchema
@@ -1416,38 +1417,112 @@ function unsafeTransportError(surface, remediation, exposedData) {
   );
 }
 
+// src/providers/rich-input.ts
+function richInputParts(input) {
+  return typeof input === "string" ? [] : input.filter((part) => part.type !== "text");
+}
+function firstRichPart(input, history, modality) {
+  const find = (value) => richInputParts(value).find((part) => modality === void 0 || part.type === modality);
+  return find(input) ?? history.map((message) => find(message.content)).find((part) => part !== void 0);
+}
+
+// src/providers/audio-format.ts
+var OPENAI_CHAT_AUDIO_FORMATS = Object.freeze({
+  "audio/wav": "wav",
+  "audio/x-wav": "wav",
+  "audio/wave": "wav",
+  "audio/mpeg": "mp3",
+  "audio/mp3": "mp3"
+});
+var OPENROUTER_AUDIO_FORMATS = Object.freeze({
+  ...OPENAI_CHAT_AUDIO_FORMATS,
+  "audio/aiff": "aiff",
+  "audio/x-aiff": "aiff",
+  "audio/aac": "aac",
+  "audio/ogg": "ogg",
+  "audio/flac": "flac",
+  "audio/x-flac": "flac",
+  "audio/m4a": "m4a",
+  "audio/x-m4a": "m4a",
+  "audio/mp4": "m4a",
+  "audio/l16": "pcm16"
+});
+function resolveAudioFormat(formats, mediaType2) {
+  if (typeof mediaType2 !== "string") return void 0;
+  const essence = mediaType2.split(";", 1)[0].trim().toLowerCase();
+  return Object.prototype.hasOwnProperty.call(formats, essence) ? formats[essence] : void 0;
+}
+
 // src/providers/openai-compatible.ts
-function compatibleImageBase64(source) {
+function unmappableMediaTypeFeature(mediaType2) {
+  return mediaType2 === void 0 ? "audio media type (missing)" : `audio media type '${mediaType2}'`;
+}
+function compatibleInlineBase64(source) {
   return source.type === "base64" ? source.data : Buffer.from(source.data.buffer, source.data.byteOffset, source.data.byteLength).toString(
     "base64"
   );
 }
-function openRouterImageParts(parts, model) {
+function compatibleRichParts(parts, model, modalities, provider) {
+  const reject = (modality, source, feature) => new UnsupportedModelInputError({
+    provider,
+    model,
+    modality,
+    source,
+    ...feature ? { feature } : {}
+  });
   const content = [];
   for (const part of parts) {
     if (part.type === "text") {
       content.push({ type: "text", text: part.text });
       continue;
     }
-    const { source } = part;
-    if (source.type === "provider-file") {
+    if (part.type === "audio") {
+      const { source: source2 } = part;
+      const audio = modalities.audio;
+      if (!audio) throw reject("audio", source2.type, "audio input");
+      if (source2.type === "provider-file" || !audio.sources.includes(source2.type)) {
+        throw reject("audio", source2.type);
+      }
+      const format = resolveAudioFormat(audio.formats, source2.mediaType);
+      if (format === void 0) {
+        throw reject("audio", source2.type, unmappableMediaTypeFeature(source2.mediaType));
+      }
+      content.push({
+        type: "input_audio",
+        // Caller base64 passes through verbatim — never decoded and re-encoded.
+        input_audio: { data: compatibleInlineBase64(source2), format }
+      });
+      if (part.label) content.push({ type: "text", text: `[Audio: ${part.label}]` });
+      continue;
+    }
+    if (part.type !== "image") {
+      const unmapped = part;
       throw new UnsupportedModelInputError({
-        provider: "openrouter",
+        provider,
         model,
-        modality: "image",
-        source: "provider-file"
+        modality: unmapped.type,
+        feature: "this input modality"
       });
     }
+    const { source } = part;
+    const image = modalities.image;
+    if (!image) throw reject("image", source.type, "image input for this model");
+    if (source.type === "provider-file" || !image.sources.includes(source.type)) {
+      throw reject("image", source.type);
+    }
     content.push({
       type: "image_url",
       image_url: {
-        url: source.type === "url" ? source.url : `data:${source.mediaType};base64,${compatibleImageBase64(source)}`
+        url: source.type === "url" ? source.url : `data:${source.mediaType};base64,${compatibleInlineBase64(source)}`
       }
     });
     if (part.label) content.push({ type: "text", text: `[Image: ${part.label}]` });
   }
   return content;
 }
+function requestAudioParts(input, history) {
+  return [input, ...history.map((message) => message.content)].flatMap((content) => richInputParts(content)).filter((part) => part.type === "audio");
+}
 function resolvePerModel(value, model, fallback) {
   if (value === void 0) return fallback;
   return typeof value === "function" ? value(model) : value;
@@ -1461,6 +1536,9 @@ function sortedKeys(table) {
   }
   return keys;
 }
+function reportedTokenCount(value) {
+  return typeof value === "number" && Number.isSafeInteger(value) && value >= 0 ? value : void 0;
+}
 function priceFromTable(table, model, promptTokens, completionTokens, cachedTokens, match = "prefix") {
   const cached = cachedTokens ?? 0;
   if (!Number.isSafeInteger(promptTokens) || !Number.isSafeInteger(completionTokens) || !Number.isSafeInteger(cached) || promptTokens < 0 || completionTokens < 0 || cached < 0 || cached > promptTokens) {
@@ -1564,26 +1642,47 @@ function normalizeMessageContent(content) {
 var OpenAICompatibleProvider = class {
   name;
   reportsRequestLifecycle = true;
+  /**
+   * Modalities the effective profile declares for this model. A blank model
+   * declares nothing, so a rich request against one fails validation exactly
+   * as it did before any modality existed.
+   */
+  inputModalities(model) {
+    return model.trim().length > 0 ? this.profile.capabilities?.inputModalities ?? {} : {};
+  }
   inputCapabilities(model) {
-    return this.name === "openrouter" && model.trim().length > 0 ? { image: { sources: ["url", "bytes", "base64"] } } : {};
+    const { image, audio } = this.inputModalities(model);
+    return {
+      ...image ? { image: { sources: image.sources } } : {},
+      // `formats` is engine-internal wire detail; the runtime contract is the
+      // accepted source kinds only.
+      ...audio ? { audio: { sources: audio.sources } } : {}
+    };
   }
   validateInput(request) {
     const modelOverride = request.providerOptions?.model;
     const effectiveModel = typeof modelOverride === "string" ? modelOverride : request.model;
-    const fail = (source, feature) => {
+    const failWith = (modality, source, feature) => {
       throw new UnsupportedModelInputError({
         provider: this.name,
         model: effectiveModel || request.model,
-        modality: "image",
+        modality,
         ...source ? { source } : {},
         ...feature ? { feature } : {}
       });
     };
+    const fail = (source, feature) => failWith(firstRichPart(request.input, request.history)?.type ?? "image", source, feature);
+    const modalities = this.inputModalities(effectiveModel);
+    const audioPart = firstRichPart(request.input, request.history, "audio");
+    if (audioPart && !modalities.audio) {
+      failWith("audio", audioPart.source.type, "audio input");
+    }
     if (request.providerOptions && "model" in request.providerOptions && (typeof modelOverride !== "string" || modelOverride.trim().length === 0)) {
       fail(void 0, "invalid model providerOptions");
     }
-    if (this.name !== "openrouter" || effectiveModel.trim().length === 0) {
-      fail(void 0, "image input for this model");
+    const imagePart = firstRichPart(request.input, request.history, "image");
+    if (!modalities.image && (!audioPart || imagePart)) {
+      failWith("image", void 0, "image input for this model");
     }
     if (request.providerOptions && "messages" in request.providerOptions) {
       fail(void 0, "raw messages providerOptions");
@@ -1594,11 +1693,21 @@ var OpenAICompatibleProvider = class {
       if (message.content.some(
         (part) => part.type === "image" && part.source.type === "provider-file"
       )) {
-        fail("provider-file");
+        failWith("image", "provider-file");
       }
     }
     if (Array.isArray(request.input) && request.input.some((part) => part.type === "image" && part.source.type === "provider-file")) {
-      fail("provider-file");
+      failWith("image", "provider-file");
+    }
+    const audio = modalities.audio;
+    if (audio) {
+      for (const part of requestAudioParts(request.input, request.history)) {
+        const source = part.source;
+        if (!audio.sources.includes(source.type)) failWith("audio", source.type);
+        if (resolveAudioFormat(audio.formats, source.mediaType) === void 0) {
+          failWith("audio", source.type, unmappableMediaTypeFeature(source.mediaType));
+        }
+      }
     }
     return { effectiveModel };
   }
@@ -1783,9 +1892,15 @@ var OpenAICompatibleProvider = class {
     return body;
   }
   formatMessage(msg, model) {
+    const modalities = this.inputModalities(model);
     const out = {
       role: this.profile.roleFor ? this.profile.roleFor(msg.role, model) : msg.role,
-      content: Array.isArray(msg.content) && this.name === "openrouter" ? openRouterImageParts(msg.content, model) : msg.content
+      // History user turns go through the same builder as the caller's input,
+      // so a rich part is reproduced identically in every continuation request.
+      // Array content is routed UNCONDITIONALLY: a profile that declares no
+      // modality has the builder reject every rich part (and pass text through)
+      // rather than emitting raw `InputContentPart` objects on the wire.
+      content: Array.isArray(msg.content) ? compatibleRichParts(msg.content, model, modalities, this.name) : msg.content
     };
     if (msg.name && (this.profile.capabilities?.emitsMessageName ?? true)) out.name = msg.name;
     if (msg.tool_calls) out.tool_calls = msg.tool_calls;
@@ -1816,6 +1931,8 @@ var OpenAICompatibleProvider = class {
   }
   toUsage(raw) {
     if (!raw) return void 0;
+    const audioInput = reportedTokenCount(raw.prompt_tokens_details?.audio_tokens);
+    const audioOutput = reportedTokenCount(raw.completion_tokens_details?.audio_tokens);
     return {
       prompt_tokens: raw.prompt_tokens,
       completion_tokens: raw.completion_tokens,
@@ -1825,7 +1942,9 @@ var OpenAICompatibleProvider = class {
       // prompt_tokens_details. Preserve the normalized hit count for the
       // generic table estimator; its split is validated before pricing.
       cached_tokens: raw.prompt_cache_hit_tokens ?? raw.prompt_tokens_details?.cached_tokens,
-      cache_write_tokens: raw.prompt_tokens_details?.cache_write_tokens
+      cache_write_tokens: raw.prompt_tokens_details?.cache_write_tokens,
+      ...audioInput !== void 0 ? { audio_input_tokens: audioInput } : {},
+      ...audioOutput !== void 0 ? { audio_output_tokens: audioOutput } : {}
     };
   }
   reportedCost(raw) {
@@ -1851,6 +1970,9 @@ var OpenAICompatibleProvider = class {
       case "from-response":
         return typeof reportedCost === "number" && Number.isFinite(reportedCost) && reportedCost >= 0 ? reportedCost : void 0;
       case "table": {
+        if (typeof usage.audio_input_tokens === "number" && usage.audio_input_tokens > 0 || typeof usage.audio_output_tokens === "number" && usage.audio_output_tokens > 0) {
+          return void 0;
+        }
         if (!isBuiltinTablePricingEligible(this.profile, {
           baseUrl: this.baseUrl,
           request: context?.request,
@@ -2198,6 +2320,26 @@ var DIRECT_OPENAI_CATALOG = [
     }),
     contextBoundary: LONG_CONTEXT_BOUNDARY
   },
+  // Audio-capable models. Audio rates reviewed 2026-09-08 against
+  // https://developers.openai.com/api/docs/pricing (no version stamp on the
+  // page): `gpt-audio-1.5` and `gpt-audio` both list Text $2.50 in / $10.00
+  // out and Audio $32.00 in / $64.00 out per 1M tokens, with no cached-input
+  // row — hence no `cachedInput` here, which makes a cache-hit report unpriced
+  // rather than silently billed at the full input rate. Only these exact ids
+  // are listed: no snapshot id is published for either, and inventing one
+  // would price an unknown model.
+  {
+    aliases: ["gpt-audio-1.5"],
+    snapshotBase: "gpt-audio-1.5",
+    short: { input: 2.5 / M, output: 10 / M },
+    audio: { input: 32 / M, output: 64 / M }
+  },
+  {
+    aliases: ["gpt-audio"],
+    snapshotBase: "gpt-audio",
+    short: { input: 2.5 / M, output: 10 / M },
+    audio: { input: 32 / M, output: 64 / M }
+  },
   // Existing direct models retain literal current Standard rows. These flat
   // rows keep native pricing aligned with the compatibility wrapper while the
   // private estimator still validates categories and billing mode strictly.
@@ -2348,32 +2490,45 @@ var CANONICAL_OPENAI_BASE_URL = "https://api.openai.com/v1";
 function isTextContentPart(value, type) {
   return value !== null && typeof value === "object" && value.type === type && typeof value.text === "string";
 }
-function hasUnmodeledDirectOpenAIContent(request) {
+function isChatAudioContentPart(value) {
+  if (value === null || typeof value !== "object") return false;
+  const part = value;
+  return part.type === "input_audio" && part.input_audio !== null && typeof part.input_audio === "object";
+}
+function classifyDirectOpenAIContent(request) {
+  if (!request) return "text";
   if ("messages" in request) {
-    if (!Array.isArray(request.messages)) return true;
+    if (!Array.isArray(request.messages)) return "unmodeled";
+    let carriesAudio = false;
     for (const message of request.messages) {
-      if (message === null || typeof message !== "object") return true;
+      if (message === null || typeof message !== "object") return "unmodeled";
       const content = message.content;
       if (content === void 0 || content === null || typeof content === "string") continue;
-      if (!Array.isArray(content) || content.some((part) => !isTextContentPart(part, "text"))) {
-        return true;
+      if (!Array.isArray(content)) return "unmodeled";
+      for (const part of content) {
+        if (isTextContentPart(part, "text")) continue;
+        if (isChatAudioContentPart(part)) {
+          carriesAudio = true;
+          continue;
+        }
+        return "unmodeled";
       }
     }
-    return false;
+    return carriesAudio ? "audio" : "text";
   }
   if ("input" in request) {
     const input = request.input;
-    if (typeof input === "string") return false;
-    if (!Array.isArray(input)) return true;
+    if (typeof input === "string") return "text";
+    if (!Array.isArray(input)) return "unmodeled";
     for (const item of input) {
-      if (item === null || typeof item !== "object") return true;
+      if (item === null || typeof item !== "object") return "unmodeled";
       const typed = item;
       switch (typed.type) {
         case "message": {
           const content = typed.content;
           if (typeof content === "string") break;
           if (!Array.isArray(content) || content.some((part) => !isTextContentPart(part, "input_text"))) {
-            return true;
+            return "unmodeled";
           }
           break;
         }
@@ -2383,31 +2538,30 @@ function hasUnmodeledDirectOpenAIContent(request) {
           const output = typed.output;
           if (typeof output === "string") break;
           if (!Array.isArray(output) || output.some((part) => !isTextContentPart(part, "input_text"))) {
-            return true;
+            return "unmodeled";
           }
           break;
         }
         case "custom_tool_call":
-          if (typeof typed.input !== "string") return true;
+          if (typeof typed.input !== "string") return "unmodeled";
           break;
         case "custom_tool_call_output":
-          if (typeof typed.output !== "string") return true;
+          if (typeof typed.output !== "string") return "unmodeled";
           break;
         case "reasoning":
           break;
         default:
-          return true;
+          return "unmodeled";
       }
     }
-    return false;
+    return "text";
   }
-  return false;
+  return "text";
 }
-function isEligibleDirectOpenAIContext(context) {
+function isEligibleDirectOpenAIContext(context, entry) {
   const request = context?.request;
   const response = context?.response;
   if (context?.baseUrl !== void 0 && context.baseUrl !== CANONICAL_OPENAI_BASE_URL) return false;
-  if (request && hasUnmodeledDirectOpenAIContent(request)) return false;
   const tier = response?.service_tier ?? response?.serviceTier ?? request?.service_tier ?? request?.serviceTier;
   if (tier !== void 0 && (typeof tier !== "string" || !["standard", "default"].includes(tier.toLowerCase()))) {
     return false;
@@ -2422,7 +2576,14 @@ function isEligibleDirectOpenAIContext(context) {
   if (reasoning !== null && typeof reasoning === "object" && "mode" in reasoning && reasoning.mode !== void 0) {
     return false;
   }
-  if (request?.modalities !== void 0 && (!Array.isArray(request.modalities) || request.modalities.length === 0 || request.modalities.some((modality) => modality !== "text")) || ["audio", "image_generation", "web_search_options"].some((key) => request?.[key] !== void 0)) {
+  const audioOutputRated = entry.audio?.output !== void 0;
+  if (request?.modalities !== void 0 && (!Array.isArray(request.modalities) || request.modalities.length === 0 || request.modalities.some(
+    (modality) => modality !== "text" && !(modality === "audio" && audioOutputRated)
+  ))) {
+    return false;
+  }
+  if (request?.audio !== void 0 && !audioOutputRated) return false;
+  if (["image_generation", "web_search_options"].some((key) => request?.[key] !== void 0)) {
     return false;
   }
   if (Array.isArray(request?.tools) && request.tools.some(
@@ -2433,23 +2594,42 @@ function isEligibleDirectOpenAIContext(context) {
   return true;
 }
 function estimateDirectOpenAICost(model, usage, context) {
-  if (!isEligibleDirectOpenAIContext(context)) return void 0;
   const entry = directOpenAIModel(model);
   if (!entry) return void 0;
-  const { prompt_tokens, completion_tokens, cached_tokens, cache_write_tokens } = usage;
+  const content = classifyDirectOpenAIContent(context?.request);
+  if (content === "unmodeled") return void 0;
+  if (!isEligibleDirectOpenAIContext(context, entry)) return void 0;
+  const {
+    prompt_tokens,
+    completion_tokens,
+    cached_tokens,
+    cache_write_tokens,
+    audio_input_tokens,
+    audio_output_tokens
+  } = usage;
+  if (content === "audio" && audio_input_tokens === void 0) return void 0;
   const cached = cached_tokens ?? 0;
   const cacheWrite = cache_write_tokens ?? 0;
-  if (![prompt_tokens, completion_tokens, cached, cacheWrite].every(
+  const audioInput = audio_input_tokens ?? 0;
+  const audioOutput = audio_output_tokens ?? 0;
+  if (![prompt_tokens, completion_tokens, cached, cacheWrite, audioInput, audioOutput].every(
     (count) => Number.isSafeInteger(count) && count >= 0
-  ) || cached + cacheWrite > prompt_tokens) {
+  ) || cached + cacheWrite + audioInput > prompt_tokens || audioOutput > completion_tokens) {
     return void 0;
   }
-  const rates = entry.long !== void 0 && entry.contextBoundary !== void 0 && prompt_tokens > entry.contextBoundary ? entry.long : entry.short;
+  if (audioInput > 0 && entry.audio === void 0 || audioOutput > 0 && entry.audio?.output === void 0) {
+    return void 0;
+  }
+  if (audioInput > 0 && cached + cacheWrite > 0) return void 0;
+  const crossesLongContext = entry.long !== void 0 && entry.contextBoundary !== void 0 && prompt_tokens > entry.contextBoundary;
+  if (crossesLongContext && (audioInput > 0 || audioOutput > 0)) return void 0;
+  const rates = crossesLongContext ? entry.long : entry.short;
   if (cached > 0 && rates.cachedInput === void 0 || cacheWrite > 0 && rates.cacheWrite === void 0) {
     return void 0;
   }
-  const ordinary = prompt_tokens - cached - cacheWrite;
-  return ordinary * rates.input + cached * (rates.cachedInput ?? 0) + cacheWrite * (rates.cacheWrite ?? 0) + completion_tokens * rates.output;
+  const ordinary = prompt_tokens - cached - cacheWrite - audioInput;
+  const textOutput = completion_tokens - audioOutput;
+  return ordinary * rates.input + cached * (rates.cachedInput ?? 0) + cacheWrite * (rates.cacheWrite ?? 0) + audioInput * (entry.audio?.input ?? 0) + textOutput * rates.output + audioOutput * (entry.audio?.output ?? 0);
 }
 function estimateOpenAICost(model, promptTokens, completionTokens, cachedTokens) {
   return estimateDirectOpenAICost(model, {
@@ -2539,6 +2719,14 @@ var OPENAI_PROFILE = {
   envBaseUrl: "OPENAI_BASE_URL",
   pricing: { kind: "table", table: OPENAI_PRICING, match: "exact" },
   reasoning: { emit: openaiReasoningEmit, capture: "none" },
+  capabilities: {
+    // Chat Completions carries recorded audio via `input_audio` (wav/mp3 only).
+    // Images are deliberately NOT declared: the Responses API serves images, and
+    // Chat-Completions image pricing is unmodeled here (fork F4).
+    inputModalities: {
+      audio: { sources: ["bytes", "base64"], formats: OPENAI_CHAT_AUDIO_FORMATS }
+    }
+  },
   roleFor: (role, model) => role === "system" && isOSeriesModel(model) ? "developer" : role,
   maxTokensField: "max_completion_tokens",
   parallelToolCalls: (model) => !isOSeriesModel(model)
@@ -2567,19 +2755,37 @@ function base64FromSource(source) {
     "base64"
   );
 }
-function responseImageParts(parts) {
+function responseImageParts(parts, model) {
   const mapped = [];
   for (const part of parts) {
     if (part.type === "text") {
       mapped.push({ type: "input_text", text: part.text });
       continue;
     }
+    if (part.type === "audio") {
+      throw new UnsupportedModelInputError({
+        provider: "openai-responses",
+        model,
+        modality: "audio",
+        source: part.source.type,
+        feature: "audio input"
+      });
+    }
+    if (part.type !== "image") {
+      const unmapped = part;
+      throw new UnsupportedModelInputError({
+        provider: "openai-responses",
+        model,
+        modality: unmapped.type,
+        feature: "this input modality"
+      });
+    }
     const { source } = part;
     if (source.type === "provider-file") {
       if (source.provider !== "openai-responses") {
         throw new UnsupportedModelInputError({
           provider: "openai-responses",
-          model: "unknown",
+          model,
           modality: "image",
           source: "provider-file"
         });
@@ -2597,6 +2803,21 @@ function responseImageParts(parts) {
   }
   return mapped;
 }
+function toResponsesUsage(raw) {
+  if (!raw) return void 0;
+  const audioInput = reportedTokenCount(raw.input_tokens_details?.audio_tokens);
+  const audioOutput = reportedTokenCount(raw.output_tokens_details?.audio_tokens);
+  return {
+    prompt_tokens: raw.input_tokens,
+    completion_tokens: raw.output_tokens,
+    total_tokens: raw.total_tokens,
+    reasoning_tokens: raw.output_tokens_details?.reasoning_tokens,
+    cached_tokens: raw.input_tokens_details?.cached_tokens,
+    cache_write_tokens: raw.input_tokens_details?.cache_write_tokens,
+    ...audioInput !== void 0 ? { audio_input_tokens: audioInput } : {},
+    ...audioOutput !== void 0 ? { audio_output_tokens: audioOutput } : {}
+  };
+}
 var OpenAIResponsesProvider = class {
   name = "openai-responses";
   reportsRequestLifecycle = true;
@@ -2612,15 +2833,18 @@ var OpenAIResponsesProvider = class {
   validateInput(request) {
     const modelOverride = request.providerOptions?.model;
     const effectiveModel = typeof modelOverride === "string" ? modelOverride : request.model;
-    const fail = (source, feature) => {
+    const failWith = (modality, source, feature) => {
       throw new UnsupportedModelInputError({
         provider: this.name,
         model: effectiveModel || request.model,
-        modality: "image",
+        modality,
         ...source ? { source } : {},
         ...feature ? { feature } : {}
       });
     };
+    const fail = (source, feature) => failWith("image", source, feature);
+    const audioPart = firstRichPart(request.input, request.history, "audio");
+    if (audioPart) failWith("audio", audioPart.source.type, "audio input");
     if (request.providerOptions && "model" in request.providerOptions && (typeof modelOverride !== "string" || modelOverride.trim().length === 0)) {
       fail(void 0, "invalid model providerOptions");
     }
@@ -2767,7 +2991,7 @@ var OpenAIResponsesProvider = class {
     const nonSystemMessages = messages.filter((m) => m.role !== "system");
     const body = {
       model: effectiveModel,
-      input: this.buildInput(nonSystemMessages),
+      input: this.buildInput(nonSystemMessages, effectiveModel),
       store: false,
       stream
     };
@@ -2819,7 +3043,7 @@ var OpenAIResponsesProvider = class {
   // ---------------------------------------------------------------------------
   // Internal: message → input mapping
   // ---------------------------------------------------------------------------
-  buildInput(messages) {
+  buildInput(messages, model) {
     const input = [];
     for (const msg of messages) {
       const text = typeof msg.content === "string" ? msg.content : "";
@@ -2857,7 +3081,7 @@ var OpenAIResponsesProvider = class {
         input.push({
           type: "message",
           role: msg.role,
-          content: typeof msg.content === "string" ? msg.content : responseImageParts(msg.content)
+          content: typeof msg.content === "string" ? msg.content : responseImageParts(msg.content, model)
         });
       }
     }
@@ -2916,14 +3140,7 @@ var OpenAIResponsesProvider = class {
         }
       }
     }
-    const usage = json.usage ? {
-      prompt_tokens: json.usage.input_tokens,
-      completion_tokens: json.usage.output_tokens,
-      total_tokens: json.usage.total_tokens,
-      reasoning_tokens: json.usage.output_tokens_details?.reasoning_tokens,
-      cached_tokens: json.usage.input_tokens_details?.cached_tokens,
-      cache_write_tokens: json.usage.input_tokens_details?.cache_write_tokens
-    } : void 0;
+    const usage = toResponsesUsage(json.usage);
     const cost = usage && !this.requestContainsImages(request) ? estimateDirectOpenAICost(json.model ?? model, usage, {
       baseUrl: this.baseUrl,
       request,
@@ -3030,14 +3247,7 @@ var OpenAIResponsesProvider = class {
       }
       case "response.completed": {
         const response = data.response;
-        const usage = response?.usage ? {
-          prompt_tokens: response.usage.input_tokens,
-          completion_tokens: response.usage.output_tokens,
-          total_tokens: response.usage.total_tokens,
-          reasoning_tokens: response.usage.output_tokens_details?.reasoning_tokens,
-          cached_tokens: response.usage.input_tokens_details?.cached_tokens,
-          cache_write_tokens: response.usage.input_tokens_details?.cache_write_tokens
-        } : void 0;
+        const usage = toResponsesUsage(response?.usage);
         const reasoningItems = response?.output?.filter((item) => item.type === "reasoning") ?? [];
         const providerMetadata = reasoningItems.length > 0 ? { openaiReasoningItems: reasoningItems } : void 0;
         return {
@@ -3086,19 +3296,37 @@ function anthropicBase64(source) {
     "base64"
   );
 }
-function anthropicImageBlocks(parts) {
+function anthropicImageBlocks(parts, model) {
   const blocks = [];
   for (const part of parts) {
     if (part.type === "text") {
       blocks.push({ type: "text", text: part.text });
       continue;
     }
+    if (part.type === "audio") {
+      throw new UnsupportedModelInputError({
+        provider: "anthropic",
+        model,
+        modality: "audio",
+        source: part.source.type,
+        feature: "audio input"
+      });
+    }
+    if (part.type !== "image") {
+      const unmapped = part;
+      throw new UnsupportedModelInputError({
+        provider: "anthropic",
+        model,
+        modality: unmapped.type,
+        feature: "this input modality"
+      });
+    }
     const { source } = part;
     if (source.type === "provider-file") {
       if (source.provider !== "anthropic") {
         throw new UnsupportedModelInputError({
           provider: "anthropic",
-          model: "unknown",
+          model,
           modality: "image",
           source: "provider-file"
         });
@@ -3560,15 +3788,18 @@ var AnthropicProvider = class {
   validateInput(request) {
     const modelOverride = request.providerOptions?.model;
     const effectiveModel = typeof modelOverride === "string" ? modelOverride : request.model;
-    const fail = (source, feature) => {
+    const failWith = (modality, source, feature) => {
       throw new UnsupportedModelInputError({
         provider: this.name,
         model: effectiveModel || request.model,
-        modality: "image",
+        modality,
         ...source ? { source } : {},
         ...feature ? { feature } : {}
       });
     };
+    const fail = (source, feature) => failWith("image", source, feature);
+    const audioPart = firstRichPart(request.input, request.history, "audio");
+    if (audioPart) failWith("audio", audioPart.source.type, "audio input");
     if (request.providerOptions && "model" in request.providerOptions && (typeof modelOverride !== "string" || modelOverride.trim().length === 0)) {
       fail(void 0, "invalid model providerOptions");
     }
@@ -3739,7 +3970,7 @@ var AnthropicProvider = class {
     const systemText = systemMessages.map((m) => m.content).join("\n\n");
     const body = {
       model: effectiveModel,
-      messages: this.mapMessages(nonSystemMessages),
+      messages: this.mapMessages(nonSystemMessages, effectiveModel),
       max_tokens: options.maxTokens ?? 4096,
       stream
     };
@@ -3806,7 +4037,7 @@ ${jsonInstruction}` : jsonInstruction;
    * - assistant messages with tool_calls -> assistant with tool_use content blocks
    * - tool messages (tool results) -> user messages with tool_result content blocks
    */
-  mapMessages(messages) {
+  mapMessages(messages, model) {
     const result = [];
     for (const msg of messages) {
       const text = typeof msg.content === "string" ? msg.content : "";
@@ -3853,7 +4084,7 @@ ${jsonInstruction}` : jsonInstruction;
       } else if (msg.role === "user") {
         result.push({
           role: "user",
-          content: typeof msg.content === "string" ? msg.content : anthropicImageBlocks(msg.content)
+          content: typeof msg.content === "string" ? msg.content : anthropicImageBlocks(msg.content, model)
         });
       }
     }
@@ -4157,7 +4388,7 @@ function lastFallbackIndex(blocks) {
 function hasRichGeminiMessages(messages) {
   return messages.some((message) => Array.isArray(message.content));
 }
-function geminiImageBase64(source) {
+function geminiInlineBase64(source) {
   return source.type === "base64" ? source.data : Buffer.from(source.data.buffer, source.data.byteOffset, source.data.byteLength).toString(
     "base64"
   );
@@ -4169,6 +4400,50 @@ function geminiInteractionContent(parts, model) {
       content.push({ type: "text", text: part.text });
       continue;
     }
+    if (part.type === "audio") {
+      const { source: audioSource } = part;
+      if (audioSource.type === "provider-file") {
+        if (audioSource.provider !== "google") {
+          throw new UnsupportedModelInputError({
+            provider: "google",
+            model,
+            modality: "audio",
+            source: "provider-file"
+          });
+        }
+        if (!audioSource.mediaType) {
+          throw new UnsupportedModelInputError({
+            provider: "google",
+            model,
+            modality: "audio",
+            source: "provider-file",
+            feature: "Interactions URI audio mediaType"
+          });
+        }
+        content.push({
+          type: "audio",
+          uri: audioSource.reference,
+          mime_type: audioSource.mediaType
+        });
+      } else {
+        content.push({
+          type: "audio",
+          data: geminiInlineBase64(audioSource),
+          mime_type: audioSource.mediaType
+        });
+      }
+      if (part.label) content.push({ type: "text", text: `[Audio: ${part.label}]` });
+      continue;
+    }
+    if (part.type !== "image") {
+      const unmapped = part;
+      throw new UnsupportedModelInputError({
+        provider: "google",
+        model,
+        modality: unmapped.type,
+        feature: "this input modality"
+      });
+    }
     const { source } = part;
     if (source.type === "provider-file") {
       if (source.provider !== "google") {
@@ -4202,7 +4477,11 @@ function geminiInteractionContent(parts, model) {
         feature: "direct URL image input; pass bytes/base64 or a Gemini provider-file"
       });
     } else {
-      content.push({ type: "image", data: geminiImageBase64(source), mime_type: source.mediaType });
+      content.push({
+        type: "image",
+        data: geminiInlineBase64(source),
+        mime_type: source.mediaType
+      });
     }
     if (part.label) content.push({ type: "text", text: `[Image: ${part.label}]` });
   }
@@ -4259,7 +4538,7 @@ var GEMINI_PRICING = {
     output: 1e-5,
     longContext: { input: 25e-7, cached: 25e-8, output: 15e-6 }
   },
-  "gemini-2.5-flash": { input: 3e-7, cached: 3e-8, output: 25e-7 },
+  "gemini-2.5-flash": { input: 3e-7, cached: 3e-8, output: 25e-7, audioInput: 1e-6 },
   "gemini-2.5-flash-lite": { input: 1e-7, cached: 1e-8, output: 4e-7 },
   "gemini-3.1-pro-preview": {
     input: 2e-6,
@@ -4277,18 +4556,43 @@ var GEMINI_PRICING = {
   "gemini-3.5-flash": { input: 15e-7, cached: 15e-8, output: 9e-6 },
   "gemini-3.5-flash-lite": { input: 3e-7, cached: 3e-8, output: 25e-7 },
   "gemini-3.6-flash": { input: 75e-8, cached: 75e-9, output: 375e-8 },
-  "gemini-3.7-flash": { input: 75e-8, cached: 75e-9, output: 375e-8 },
+  "gemini-3.7-flash": {
+    input: 75e-8,
+    cached: 75e-9,
+    output: 375e-8,
+    audioInput: 75e-8
+  },
   "gemini-3.8-flash": { input: 75e-8, cached: 75e-9, output: 375e-8 }
 };
+var CANONICAL_GEMINI_BASE_URL = "https://generativelanguage.googleapis.com/v1beta";
 function isValidTokenCount2(value) {
   return typeof value === "number" && Number.isSafeInteger(value) && value >= 0;
 }
 function isStandardGeminiRequestTier(value) {
-  return value === void 0 || value === "SERVICE_TIER_UNSPECIFIED" || value === "SERVICE_TIER_STANDARD";
+  return value === void 0 || value === null || isDefinitiveStandardGeminiResponseTier(value);
 }
 function isDefinitiveStandardGeminiResponseTier(value) {
   return value === "SERVICE_TIER_UNSPECIFIED" || value === "SERVICE_TIER_STANDARD" || value === "standard";
 }
+function recordGeminiResponseTier(evidence, value) {
+  if (value === void 0) return;
+  evidence.reportsTier = true;
+  if (isDefinitiveStandardGeminiResponseTier(value)) {
+    evidence.hasDefinitiveStandardTier = true;
+  } else {
+    evidence.hasInvalidTier = true;
+  }
+}
+function geminiResponseTierEvidence(responseHeaders, ...values) {
+  const evidence = {
+    reportsTier: false,
+    hasDefinitiveStandardTier: false,
+    hasInvalidTier: false
+  };
+  recordGeminiResponseTier(evidence, responseHeaders?.get("x-gemini-service-tier") ?? void 0);
+  for (const value of values) recordGeminiResponseTier(evidence, value);
+  return evidence;
+}
 function geminiInteractionErrorStatus(code) {
   if (typeof code === "number" && Number.isInteger(code) && code >= 100 && code <= 599) {
     return code;
@@ -4347,8 +4651,19 @@ function estimateGeminiCost(model, usage) {
   if (!pricing || !isValidTokenCount2(usage.inputTokens) || !isValidTokenCount2(usage.outputTokens) || !isValidTokenCount2(cached) || cached > usage.inputTokens) {
     return void 0;
   }
-  const rate = usage.inputTokens > 2e5 && pricing.longContext ? pricing.longContext : pricing;
-  return (usage.inputTokens - cached) * rate.input + cached * rate.cached + usage.outputTokens * rate.output;
+  const split = usage.modalities;
+  const audio = split?.audioTokens ?? 0;
+  if (split !== void 0) {
+    if (audio > 0 && pricing.audioInput === void 0) return void 0;
+    if (!isValidTokenCount2(audio) || !isValidTokenCount2(split.nonAudioTokens) || audio + split.nonAudioTokens !== usage.inputTokens || cached + audio > usage.inputTokens) {
+      return void 0;
+    }
+    if (audio > 0 && cached > 0) return void 0;
+  }
+  const crossesLongContext = usage.inputTokens > 2e5 && pricing.longContext !== void 0;
+  if (crossesLongContext && audio > 0) return void 0;
+  const rate = crossesLongContext ? pricing.longContext : pricing;
+  return (usage.inputTokens - cached - audio) * rate.input + cached * rate.cached + audio * (pricing.audioInput ?? 0) + usage.outputTokens * rate.output;
 }
 function safeTokenSum2(...values) {
   const total = values.reduce((sum, value) => sum + value, 0);
@@ -4508,20 +4823,44 @@ var GeminiProvider = class {
   name = "google";
   reportsRequestLifecycle = true;
   inputCapabilities(model) {
-    return model.trim().length > 0 ? { image: { sources: ["bytes", "base64", "provider-file"] } } : {};
+    return model.trim().length > 0 ? {
+      image: { sources: ["bytes", "base64", "provider-file"] },
+      audio: { sources: ["bytes", "base64", "provider-file"] }
+    } : {};
   }
   validateInput(request) {
     const modelOverride = request.providerOptions?.model;
     const effectiveModel = typeof modelOverride === "string" ? modelOverride : request.model;
-    const fail = (source, feature) => {
+    const failWith = (modality, source, feature) => {
       throw new UnsupportedModelInputError({
         provider: this.name,
         model: effectiveModel || request.model,
-        modality: "image",
+        modality,
         ...source ? { source } : {},
         ...feature ? { feature } : {}
       });
     };
+    const fail = (source, feature) => failWith(firstRichPart(request.input, request.history)?.type ?? "image", source, feature);
+    const checkPart = (part) => {
+      if (part.type === "audio") {
+        if (part.source.type === "provider-file" && part.source.provider !== this.name)
+          failWith("audio", "provider-file");
+        if (part.source.type === "provider-file" && !part.source.mediaType)
+          failWith("audio", "provider-file", "Interactions URI audio mediaType");
+        return;
+      }
+      if (part.type !== "image") return;
+      if (part.source.type === "url")
+        failWith(
+          "image",
+          "url",
+          "direct URL image input; pass bytes/base64 or a Gemini provider-file"
+        );
+      if (part.source.type === "provider-file" && part.source.provider !== this.name)
+        failWith("image", "provider-file");
+      if (part.source.type === "provider-file" && !part.source.mediaType)
+        failWith("image", "provider-file", "Interactions URI image mediaType");
+    };
     if (request.providerOptions && "model" in request.providerOptions && (typeof modelOverride !== "string" || modelOverride.trim().length === 0)) {
       fail(void 0, "invalid model providerOptions");
     }
@@ -4543,26 +4882,10 @@ var GeminiProvider = class {
     for (const message of request.history) {
       if (!Array.isArray(message.content)) continue;
       if (message.role !== "user") fail(void 0, "rich non-user history");
-      for (const part of message.content) {
-        if (part.type !== "image") continue;
-        if (part.source.type === "url")
-          fail("url", "direct URL image input; pass bytes/base64 or a Gemini provider-file");
-        if (part.source.type === "provider-file" && part.source.provider !== this.name)
-          fail("provider-file");
-        if (part.source.type === "provider-file" && !part.source.mediaType)
-          fail("provider-file", "Interactions URI image mediaType");
-      }
+      for (const part of message.content) checkPart(part);
     }
     if (Array.isArray(request.input)) {
-      for (const part of request.input) {
-        if (part.type !== "image") continue;
-        if (part.source.type === "url")
-          fail("url", "direct URL image input; pass bytes/base64 or a Gemini provider-file");
-        if (part.source.type === "provider-file" && part.source.provider !== this.name)
-          fail("provider-file");
-        if (part.source.type === "provider-file" && !part.source.mediaType)
-          fail("provider-file", "Interactions URI image mediaType");
-      }
+      for (const part of request.input) checkPart(part);
     }
     return { effectiveModel };
   }
@@ -4698,6 +5021,7 @@ var GeminiProvider = class {
   async chatInteraction(messages, options) {
     this.assertSafeInteractionOptions(options);
     const body = this.buildInteractionRequestBody(messages, options, false);
+    const pricingContext = this.interactionPricingContext(this.requestModel(options), body);
     const headers = this.buildHeaders(await this.resolveKey());
     const recorder = new CallTimingRecorder(options.requestLifecycle);
     const res = await fetchWithRetry(
@@ -4718,12 +5042,17 @@ var GeminiProvider = class {
     }
     return withChatTiming(
       recorder,
-      this.parseInteractionResponse(await res.json())
+      this.parseInteractionResponse(
+        await res.json(),
+        pricingContext,
+        res.headers
+      )
     );
   }
   async *streamInteraction(messages, options) {
     this.assertSafeInteractionOptions(options);
     const body = this.buildInteractionRequestBody(messages, options, true);
+    const pricingContext = this.interactionPricingContext(this.requestModel(options), body);
     const headers = this.buildHeaders(await this.resolveKey());
     const recorder = new CallTimingRecorder(options.requestLifecycle);
     const res = await fetchWithRetry(
@@ -4745,7 +5074,7 @@ var GeminiProvider = class {
     if (!res.body) throw new Error("Gemini Interactions stream has no body");
     yield* withCallTiming(
       recorder,
-      this.parseInteractionSSEStream(res.body, res.headers, recorder)
+      this.parseInteractionSSEStream(res.body, res.headers, recorder, pricingContext)
     );
   }
   buildInteractionRequestBody(messages, options, stream) {
@@ -4840,6 +5169,7 @@ var GeminiProvider = class {
   }
   mapInteractionInput(messages, model) {
     const steps = [];
+    const functionNames = /* @__PURE__ */ new Map();
     for (const message of messages) {
       if (message.role === "user") {
         steps.push({
@@ -4849,6 +5179,10 @@ var GeminiProvider = class {
       } else if (message.role === "assistant") {
         const nativeSteps = message.providerMetadata?.geminiInteractionSteps;
         if (Array.isArray(nativeSteps) && nativeSteps.every(isGeminiInteractionStep)) {
+          for (const step of nativeSteps) {
+            if (step.type === "function_call" && typeof step.id === "string" && step.name)
+              functionNames.set(step.id, step.name);
+          }
           steps.push(...nativeSteps);
           continue;
         }
@@ -4856,6 +5190,7 @@ var GeminiProvider = class {
           steps.push({ type: "model_output", content: [{ type: "text", text: message.content }] });
         }
         for (const call of message.tool_calls ?? []) {
+          functionNames.set(call.id, call.function.name);
           steps.push({
             type: "function_call",
             id: call.id,
@@ -4864,9 +5199,17 @@ var GeminiProvider = class {
           });
         }
       } else if (message.role === "tool") {
+        const callId = message.tool_call_id ?? "";
+        const name = functionNames.get(callId) ?? message.name;
+        if (!name) {
+          throw new InvalidModelInputError(
+            `Gemini Interactions tool result for call '${callId}' has no function name: no preceding function_call carries that id and the tool message has no 'name'`
+          );
+        }
         steps.push({
           type: "function_result",
-          call_id: message.tool_call_id ?? "",
+          name,
+          call_id: callId,
           result: [
             { type: "text", text: typeof message.content === "string" ? message.content : "" }
           ]
@@ -4875,7 +5218,7 @@ var GeminiProvider = class {
     }
     return steps;
   }
-  parseInteractionResponse(json) {
+  parseInteractionResponse(json, pricingContext, responseHeaders) {
     let content = "";
     let thinkingContent = "";
     const toolCalls = [];
@@ -4891,18 +5234,63 @@ var GeminiProvider = class {
       }
     }
     this.assertInteractionTerminalStatus(json.status, toolCalls.length);
-    const usage = normalizeInteractionUsage(json.usage);
+    const normalized = normalizeInteractionUsage(json.usage);
     return {
       content,
       thinking_content: thinkingContent || void 0,
       tool_calls: toolCalls.length ? toolCalls : void 0,
-      usage,
-      // Interactions returns modality usage but no authoritative monetary cost.
-      cost: void 0,
+      usage: normalized?.usage,
+      cost: this.interactionCost({
+        normalized,
+        steps: json.steps,
+        responseTierEvidence: geminiResponseTierEvidence(
+          responseHeaders,
+          json.service_tier,
+          json.usage?.service_tier
+        ),
+        effectiveModel: typeof json.model === "string" ? json.model : pricingContext.model,
+        pricingContext,
+        // The whole body is in hand: every part is on `json.steps`, so there
+        // is nothing this path could have failed to materialize.
+        unmodeledOutput: false
+      }),
       providerMetadata: json.steps?.length ? { geminiInteractionSteps: json.steps.filter(isGeminiInteractionStep) } : void 0
     };
   }
-  async *parseInteractionSSEStream(body, responseHeaders, timing) {
+  /**
+   * Cost for one Interactions call, or `undefined` when any precondition
+   * fails. Streaming and non-streaming share this one body, so identical usage
+   * always produces an identical number on both paths.
+   */
+  interactionCost(call) {
+    const { normalized, steps, responseTierEvidence, effectiveModel, pricingContext } = call;
+    if (!normalized?.pricingUsage || normalized.hasUnmodeledBilledUsage) return void 0;
+    if (call.unmodeledOutput || !this.isTextOnlyInteractionOutput(steps)) return void 0;
+    if (!isEligibleGeminiPricing(
+      pricingContext,
+      !responseTierEvidence.reportsTier || responseTierEvidence.hasDefinitiveStandardTier,
+      responseTierEvidence.hasInvalidTier
+    )) {
+      return void 0;
+    }
+    return estimateGeminiCost(effectiveModel, normalized.pricingUsage);
+  }
+  /**
+   * Whether every model-produced step is text. A non-text reply part (image,
+   * audio, inline data) is billed on an output track this adapter carries no
+   * rate for, so the call stays unpriced. Echoed `user_input` and
+   * Axl-authored `function_result` steps are not model output and are ignored.
+   */
+  isTextOnlyInteractionOutput(steps) {
+    return (steps ?? []).every((step) => {
+      if (step.type !== "model_output" && step.type !== "thought") return true;
+      const parts = [...step.content ?? [], ...step.summary ?? []];
+      return parts.every(
+        (part) => part !== null && typeof part === "object" && part.type === "text"
+      );
+    });
+  }
+  async *parseInteractionSSEStream(body, responseHeaders, timing, pricingContext) {
     const reader = body.getReader();
     const decoder = new TextDecoder();
     let buffer = "";
@@ -4910,6 +5298,8 @@ var GeminiProvider = class {
     const argumentBuffers = /* @__PURE__ */ new Map();
     const stopped = /* @__PURE__ */ new Set();
     let completed = false;
+    const responseTierEvidence = geminiResponseTierEvidence(responseHeaders);
+    let unmodeledOutput = false;
     try {
       while (true) {
         const { done, value } = await timing.read(reader);
@@ -4927,6 +5317,8 @@ var GeminiProvider = class {
           } catch {
             continue;
           }
+          recordGeminiResponseTier(responseTierEvidence, event.interaction?.service_tier);
+          recordGeminiResponseTier(responseTierEvidence, event.interaction?.usage?.service_tier);
           if (event.event_type === "error") {
             const status = geminiInteractionErrorStatus(event.error?.code);
             const message = event.error?.message ?? "Gemini Interactions stream failed";
@@ -4954,6 +5346,7 @@ var GeminiProvider = class {
             }
             const step = cloneInteractionStep(event.step);
             steps.set(index, step);
+            if (!this.isTextOnlyInteractionOutput([step])) unmodeledOutput = true;
             if (step.type === "model_output") {
               for (const text of interactionTextChunks(step.content)) {
                 yield { type: "text_delta", content: text };
@@ -4976,6 +5369,12 @@ var GeminiProvider = class {
             const step = steps.get(index);
             if (!step)
               throw new Error("Gemini Interactions stream delta arrived before step.start");
+            if (step.type === "model_output" && !(delta.type === "text" && typeof delta.text === "string")) {
+              unmodeledOutput = true;
+            }
+            if (step.type === "thought" && !(delta.type === "thought_summary" && !!delta.content) && !(delta.type === "thought_signature" && typeof delta.signature === "string")) {
+              unmodeledOutput = true;
+            }
             if (delta.type === "text" && typeof delta.text === "string" && step.type === "model_output") {
               appendInteractionContent(step, "content", { type: "text", text: delta.text });
               yield { type: "text_delta", content: delta.text };
@@ -5023,12 +5422,20 @@ var GeminiProvider = class {
             ).length;
             this.assertInteractionTerminalStatus(event.interaction?.status, toolCallCount);
             completed = true;
+            const orderedSteps = [...steps.entries()].sort(([a], [b]) => a - b).map(([, step]) => step);
+            const normalized = normalizeInteractionUsage(event.interaction?.usage);
             yield {
               type: "done",
-              usage: normalizeInteractionUsage(event.interaction?.usage),
-              providerMetadata: steps.size ? {
-                geminiInteractionSteps: [...steps.entries()].sort(([a], [b]) => a - b).map(([, step]) => step)
-              } : void 0
+              usage: normalized?.usage,
+              cost: this.interactionCost({
+                normalized,
+                steps: orderedSteps,
+                responseTierEvidence,
+                effectiveModel: typeof event.interaction?.model === "string" ? event.interaction.model : pricingContext.model,
+                pricingContext,
+                unmodeledOutput
+              }),
+              providerMetadata: steps.size ? { geminiInteractionSteps: orderedSteps } : void 0
             };
             return;
           }
@@ -5070,6 +5477,51 @@ var GeminiProvider = class {
     }
     return `Gemini API error (${status}): ${body}`;
   }
+  /**
+   * Pricing context for the stateless Interactions transport.
+   *
+   * The endpoint reports a per-modality input split, so a rich call IS
+   * priceable — but only from the canonical endpoint, on a standard tier, with
+   * a request whose billing is fully represented by the token totals.
+   */
+  interactionPricingContext(fallbackModel, body) {
+    return {
+      model: typeof body.model === "string" ? body.model : fallbackModel,
+      serviceTier: body.service_tier ?? body.serviceTier,
+      eligibleRequest: this.baseUrl === CANONICAL_GEMINI_BASE_URL && isStandardGeminiRequestTier(body.serviceTier) && this.isModeledInteractionRequest(body)
+    };
+  }
+  /**
+   * Whether an Interactions request body's billing is fully represented by the
+   * reported token totals. `providerOptions` merges arbitrary keys into the
+   * body, so anything that can move billing off the local totals — a
+   * server-side cache, a server-executed tool, a residency/region override, a
+   * non-text response modality — makes the call unpriced.
+   *
+   * Input CONTENT is deliberately not inspected: `input_tokens_by_modality` is
+   * the authority on what the input billed as, and it unprices any modality
+   * without a billing category (video, documents, anything new).
+   */
+  isModeledInteractionRequest(body) {
+    if ([
+      "cached_content",
+      "cachedContent",
+      "previous_interaction_id",
+      "region",
+      "inference_geo",
+      "data_residency"
+    ].some((key) => body[key] !== void 0)) {
+      return false;
+    }
+    if (body.tools !== void 0 && (!Array.isArray(body.tools) || body.tools.some(
+      (tool2) => tool2 === null || typeof tool2 !== "object" || tool2.type !== "function"
+    ))) {
+      return false;
+    }
+    const generationConfig = body.generation_config;
+    const modalities = generationConfig !== null && typeof generationConfig === "object" ? generationConfig.response_modalities : void 0;
+    return modalities === void 0 || modalities === "TEXT" || Array.isArray(modalities) && modalities.length === 1 && modalities[0] === "TEXT";
+  }
   pricingContext(fallbackModel, body) {
     return {
       // Gemini's model is normally selected in the URL. providerOptions is an
@@ -5077,7 +5529,7 @@ var GeminiProvider = class {
       // the URL and pricing rather than pricing a different request.
       model: typeof body.model === "string" ? body.model : fallbackModel,
       serviceTier: body.serviceTier,
-      eligibleRequest: this.baseUrl === "https://generativelanguage.googleapis.com/v1beta" && this.isTextOnlyClientRequest(body)
+      eligibleRequest: this.baseUrl === CANONICAL_GEMINI_BASE_URL && this.isTextOnlyClientRequest(body)
     };
   }
   requestModel(options) {
@@ -5400,7 +5852,7 @@ var GeminiProvider = class {
     }
     const normalized = json.usageMetadata ? normalizeGeminiUsage(json.usageMetadata) : void 0;
     const effectiveModel = typeof json.modelVersion === "string" ? json.modelVersion : pricingContext.model;
-    const cost = normalized && !normalized.hasUnmodeledBilledUsage && this.isTextOnlyGeminiResponse(candidate) && isEligibleGeminiPricing(
+    const cost = normalized?.pricingUsage && !normalized.hasUnmodeledBilledUsage && this.isTextOnlyGeminiResponse(candidate) && isEligibleGeminiPricing(
       pricingContext,
       isDefinitiveStandardGeminiResponseTier(json.usageMetadata?.serviceTier),
       json.usageMetadata?.serviceTier !== void 0 && !isDefinitiveStandardGeminiResponseTier(json.usageMetadata.serviceTier)
@@ -5493,7 +5945,7 @@ var GeminiProvider = class {
       yield {
         type: "done",
         usage: normalizedUsage?.usage,
-        cost: normalizedUsage && !normalizedUsage.hasUnmodeledBilledUsage && responseIsTextOnly && isEligibleGeminiPricing(
+        cost: normalizedUsage?.pricingUsage && !normalizedUsage.hasUnmodeledBilledUsage && responseIsTextOnly && isEligibleGeminiPricing(
           pricingContext,
           hasDefinitiveStandardResponseTier,
           hasInvalidResponseTier
@@ -5539,20 +5991,73 @@ function safeJsonObject(value) {
     return {};
   }
 }
+var KNOWN_GEMINI_MODALITIES = /* @__PURE__ */ new Set(["text", "image", "audio"]);
+function foldGeminiModalities(raw, totalInputTokens) {
+  if (!Array.isArray(raw)) return void 0;
+  let audioTokens = 0;
+  let nonAudioTokens = 0;
+  for (const entry of raw) {
+    if (entry === null || typeof entry !== "object") return void 0;
+    const { modality, tokens } = entry;
+    if (typeof modality !== "string" || !isValidTokenCount2(tokens)) return void 0;
+    const normalized = modality.toLowerCase();
+    if (!KNOWN_GEMINI_MODALITIES.has(normalized)) return void 0;
+    if (normalized === "audio") audioTokens += tokens;
+    else nonAudioTokens += tokens;
+  }
+  const total = safeTokenSum2(audioTokens, nonAudioTokens);
+  if (total === void 0 || total !== totalInputTokens) return void 0;
+  return { audioTokens, nonAudioTokens };
+}
+function reportedGeminiAudioTokens(raw, totalInputTokens) {
+  if (!Array.isArray(raw)) return void 0;
+  let audioTokens;
+  for (const entry of raw) {
+    if (entry === null || typeof entry !== "object") return void 0;
+    const { modality, tokens } = entry;
+    if (typeof modality !== "string" || !isValidTokenCount2(tokens)) return void 0;
+    if (modality.toLowerCase() === "audio") audioTokens = (audioTokens ?? 0) + tokens;
+  }
+  return audioTokens !== void 0 && audioTokens <= totalInputTokens ? audioTokens : void 0;
+}
 function normalizeInteractionUsage(usage) {
   if (!usage || !isValidTokenCount2(usage.total_input_tokens) || !isValidTokenCount2(usage.total_output_tokens) || !isValidTokenCount2(usage.total_tokens)) {
     return void 0;
   }
   const cached = usage.total_cached_tokens;
   const thinking = usage.total_thought_tokens;
-  if (cached !== void 0 && (!isValidTokenCount2(cached) || cached > usage.total_input_tokens) || thinking !== void 0 && !isValidTokenCount2(thinking))
+  const toolUse = usage.total_tool_use_tokens;
+  if (cached !== void 0 && (!isValidTokenCount2(cached) || cached > usage.total_input_tokens) || thinking !== void 0 && !isValidTokenCount2(thinking) || toolUse !== void 0 && !isValidTokenCount2(toolUse))
     return void 0;
+  const inputTokens = usage.total_input_tokens;
+  const audioInputTokens = reportedGeminiAudioTokens(usage.input_tokens_by_modality, inputTokens);
+  const modalities = foldGeminiModalities(usage.input_tokens_by_modality, inputTokens);
+  const billedOutputTokens = safeTokenSum2(usage.total_output_tokens, thinking ?? 0);
+  const reportedTotal = safeTokenSum2(
+    inputTokens,
+    usage.total_output_tokens,
+    thinking ?? 0,
+    toolUse ?? 0
+  );
+  const reconciles = reportedTotal !== void 0 && usage.total_tokens === reportedTotal;
   return {
-    prompt_tokens: usage.total_input_tokens,
-    completion_tokens: usage.total_output_tokens,
-    total_tokens: usage.total_tokens,
-    ...cached ? { cached_tokens: cached } : {},
-    ...thinking ? { reasoning_tokens: thinking } : {}
+    usage: {
+      prompt_tokens: inputTokens,
+      completion_tokens: usage.total_output_tokens,
+      total_tokens: usage.total_tokens,
+      ...cached ? { cached_tokens: cached } : {},
+      ...thinking ? { reasoning_tokens: thinking } : {},
+      ...audioInputTokens !== void 0 ? { audio_input_tokens: audioInputTokens } : {}
+    },
+    ...modalities !== void 0 && billedOutputTokens !== void 0 && reconciles ? {
+      pricingUsage: {
+        inputTokens,
+        outputTokens: billedOutputTokens,
+        cachedTokens: cached ?? 0,
+        modalities
+      }
+    } : {},
+    hasUnmodeledBilledUsage: toolUse !== void 0 && toolUse !== 0
   };
 }
 
@@ -5572,6 +6077,14 @@ var OPENROUTER_PROFILE = {
     capture: "reasoning_details",
     roundTrip: "on-tool-call-turns"
   },
+  capabilities: {
+    // Images: `image_url` data URLs or remote URLs. Audio: `input_audio`, which
+    // OpenRouter documents as base64-only — bytes are encoded by the engine.
+    inputModalities: {
+      image: { sources: ["url", "bytes", "base64"] },
+      audio: { sources: ["bytes", "base64"], formats: OPENROUTER_AUDIO_FORMATS }
+    }
+  },
   // OpenRouter normalizes to max_tokens across its backends.
   maxTokensField: "max_tokens"
 };
@@ -7067,87 +7580,98 @@ function validBase64(value) {
 function decodedBase64Bytes(value) {
   return Math.floor(value.length * 3 / 4) - (value.endsWith("==") ? 2 : value.endsWith("=") ? 1 : 0);
 }
+var INLINE_MEDIA_LIMIT_MESSAGE = "Inline media data must not exceed 25 MiB total; use a provider-file source, or a URL for images, where supported";
+var IMAGE_SOURCE_TYPES = ["url", "bytes", "base64", "provider-file"];
+var AUDIO_SOURCE_TYPES = ["bytes", "base64", "provider-file"];
+function cloneMediaSource(source, index, allowed, reserveInlineBytes) {
+  if (!source || typeof source !== "object") invalid(`part ${index}.source must be an object`);
+  const raw = source;
+  const kind = raw.type;
+  if (typeof kind !== "string" || !allowed.includes(kind))
+    invalid(`part ${index}.source.type is unsupported`);
+  let clone;
+  switch (kind) {
+    case "url": {
+      const url = nonEmptyString2(raw.url, `part ${index}.source.url`);
+      let parsed;
+      try {
+        parsed = new URL(url);
+      } catch {
+        invalid(`part ${index}.source.url must be an http(s) URL`);
+      }
+      if (parsed.protocol !== "http:" && parsed.protocol !== "https:")
+        invalid(`part ${index}.source.url must be an http(s) URL`);
+      clone = {
+        type: "url",
+        url,
+        ...raw.mediaType === void 0 ? {} : { mediaType: mediaType(raw.mediaType, `part ${index}.source.mediaType`) }
+      };
+      break;
+    }
+    case "bytes": {
+      if (!(raw.data instanceof Uint8Array) || raw.data.byteLength === 0)
+        invalid(`part ${index}.source.data must be a non-empty Uint8Array`);
+      reserveInlineBytes(raw.data.byteLength);
+      clone = {
+        type: "bytes",
+        data: new Uint8Array(raw.data),
+        mediaType: mediaType(raw.mediaType, `part ${index}.source.mediaType`)
+      };
+      break;
+    }
+    case "base64": {
+      const data = nonEmptyString2(raw.data, `part ${index}.source.data`);
+      if (data.length > 4 * Math.ceil(MAX_INLINE_MODEL_INPUT_BYTES / 3))
+        invalid(INLINE_MEDIA_LIMIT_MESSAGE);
+      if (!validBase64(data)) invalid(`part ${index}.source.data must be valid base64`);
+      reserveInlineBytes(decodedBase64Bytes(data));
+      clone = {
+        type: "base64",
+        data,
+        mediaType: mediaType(raw.mediaType, `part ${index}.source.mediaType`)
+      };
+      break;
+    }
+    default:
+      clone = {
+        type: "provider-file",
+        provider: nonEmptyString2(raw.provider, `part ${index}.source.provider`),
+        reference: nonEmptyString2(raw.reference, `part ${index}.source.reference`),
+        ...raw.mediaType === void 0 ? {} : { mediaType: mediaType(raw.mediaType, `part ${index}.source.mediaType`) }
+      };
+  }
+  return clone;
+}
 function normalizeModelInput(input) {
   if (typeof input === "string") return input;
   if (!Array.isArray(input) || input.length === 0)
     invalid("ModelInput parts must be a non-empty array");
   let inlineBytes = 0;
   const reserveInlineBytes = (bytes) => {
-    if (bytes > MAX_INLINE_MODEL_INPUT_BYTES - inlineBytes) {
-      invalid(
-        "Inline image data must not exceed 25 MiB total; use a URL or provider-file source where supported"
-      );
-    }
+    if (bytes > MAX_INLINE_MODEL_INPUT_BYTES - inlineBytes) invalid(INLINE_MEDIA_LIMIT_MESSAGE);
     inlineBytes += bytes;
   };
   return input.map((part, index) => {
     if (!part || typeof part !== "object") invalid(`ModelInput part ${index} must be an object`);
-    if (part.type === "text")
-      return { type: "text", text: nonEmptyString2(part.text, `part ${index}.text`) };
-    if (part.type !== "image") invalid(`part ${index}.type must be 'text' or 'image'`);
-    const source = part.source;
-    if (!source || typeof source !== "object") invalid(`part ${index}.source must be an object`);
-    let clone;
-    switch (source.type) {
-      case "url": {
-        const url = nonEmptyString2(source.url, `part ${index}.source.url`);
-        let parsed;
-        try {
-          parsed = new URL(url);
-        } catch {
-          invalid(`part ${index}.source.url must be an http(s) URL`);
-        }
-        if (parsed.protocol !== "http:" && parsed.protocol !== "https:")
-          invalid(`part ${index}.source.url must be an http(s) URL`);
-        clone = {
-          type: "url",
-          url,
-          ...source.mediaType === void 0 ? {} : { mediaType: mediaType(source.mediaType, `part ${index}.source.mediaType`) }
-        };
-        break;
-      }
-      case "bytes":
-        if (!(source.data instanceof Uint8Array) || source.data.byteLength === 0)
-          invalid(`part ${index}.source.data must be a non-empty Uint8Array`);
-        reserveInlineBytes(source.data.byteLength);
-        clone = {
-          type: "bytes",
-          data: source.data.slice(),
-          mediaType: mediaType(source.mediaType, `part ${index}.source.mediaType`)
-        };
-        break;
-      case "base64": {
-        const data = nonEmptyString2(source.data, `part ${index}.source.data`);
-        if (data.length > 4 * Math.ceil(MAX_INLINE_MODEL_INPUT_BYTES / 3)) {
-          invalid(
-            "Inline image data must not exceed 25 MiB total; use a URL or provider-file source where supported"
-          );
-        }
-        if (!validBase64(data)) invalid(`part ${index}.source.data must be valid base64`);
-        reserveInlineBytes(decodedBase64Bytes(data));
-        clone = {
-          type: "base64",
-          data,
-          mediaType: mediaType(source.mediaType, `part ${index}.source.mediaType`)
+    const mediaLabel = () => part.label === void 0 ? {} : { label: nonEmptyString2(part.label, `part ${index}.label`) };
+    switch (part.type) {
+      case "text":
+        return { type: "text", text: nonEmptyString2(part.text, `part ${index}.text`) };
+      case "image":
+        return {
+          type: "image",
+          source: cloneMediaSource(part.source, index, IMAGE_SOURCE_TYPES, reserveInlineBytes),
+          ...mediaLabel()
         };
-        break;
-      }
-      case "provider-file":
-        clone = {
-          type: "provider-file",
-          provider: nonEmptyString2(source.provider, `part ${index}.source.provider`),
-          reference: nonEmptyString2(source.reference, `part ${index}.source.reference`),
-          ...source.mediaType === void 0 ? {} : { mediaType: mediaType(source.mediaType, `part ${index}.source.mediaType`) }
+      case "audio":
+        return {
+          type: "audio",
+          source: cloneMediaSource(part.source, index, AUDIO_SOURCE_TYPES, reserveInlineBytes),
+          ...mediaLabel()
         };
-        break;
       default:
-        invalid(`part ${index}.source.type is unsupported`);
+        invalid(`part ${index}.type must be 'text', 'image', or 'audio'`);
     }
-    return {
-      type: "image",
-      source: clone,
-      ...part.label === void 0 ? {} : { label: nonEmptyString2(part.label, `part ${index}.label`) }
-    };
   });
 }
 function inputText(input) {
@@ -7156,33 +7680,54 @@ function inputText(input) {
 function cloneModelInput(input) {
   return typeof input === "string" ? input : normalizeModelInput(input);
 }
+function unsupportedPart(part) {
+  return invalid(
+    `Unsupported model input part type '${String(part.type)}'`
+  );
+}
+function describeSource(source, label) {
+  const media = source;
+  const locator = media.type === "url" ? media.url : media.type === "provider-file" ? media.reference : void 0;
+  return {
+    source: source.type,
+    ...media.mediaType ? { mediaType: media.mediaType } : {},
+    ...media.type === "bytes" ? { bytes: media.data.byteLength } : {},
+    ...media.type === "base64" ? { bytes: decodedBase64Bytes(media.data) } : {},
+    ...locator ? { locator } : {},
+    ...label ? { label } : {}
+  };
+}
 function describeModelInput(input) {
   if (typeof input === "string") return void 0;
   return {
     parts: input.map((part) => {
-      if (part.type === "text") return { type: "text", characters: part.text.length };
-      const { source } = part;
-      return {
-        type: "image",
-        source: source.type,
-        ...source.mediaType ? { mediaType: source.mediaType } : {},
-        ...source.type === "bytes" ? { bytes: source.data.byteLength } : {},
-        ...source.type === "base64" ? {
-          bytes: decodedBase64Bytes(source.data)
-        } : {},
-        ...(source.type === "url" ? source.url : source.type === "provider-file" ? source.reference : void 0) ? {
-          locator: source.type === "url" ? source.url : source.type === "provider-file" ? source.reference : void 0
-        } : {},
-        ...part.label ? { label: part.label } : {}
-      };
+      switch (part.type) {
+        case "text":
+          return { type: "text", characters: part.text.length };
+        case "image":
+          return { type: "image", ...describeSource(part.source, part.label) };
+        case "audio":
+          return { type: "audio", ...describeSource(part.source, part.label) };
+        default:
+          return unsupportedPart(part);
+      }
     })
   };
 }
 function summarizeModelInput(input) {
   if (typeof input === "string") return input;
-  return input.map(
-    (part) => part.type === "text" ? part.text : `[image ${part.source.mediaType ?? "media"}]`
-  ).join("\n");
+  return input.map((part) => {
+    switch (part.type) {
+      case "text":
+        return part.text;
+      case "image":
+        return `[image ${part.source.mediaType ?? "media"}]`;
+      case "audio":
+        return `[audio ${part.source.mediaType ?? "media"}]`;
+      default:
+        return unsupportedPart(part);
+    }
+  }).join("\n");
 }
 function sanitizeModelInputForTrace(input) {
   return inputText(input);
@@ -7196,15 +7741,86 @@ function normalizePersistedSessionHistory(history) {
       );
     }
     const content = normalizeModelInput(message.content);
-    if (typeof content !== "string" && content.some((part) => part.type === "image" && part.source.type === "bytes")) {
+    if (typeof content !== "string" && content.some((part) => part.type !== "text" && part.source.type === "bytes")) {
       throw new InvalidModelInputError(
-        "Uint8Array image input cannot be persisted in session history"
+        "Uint8Array media input cannot be persisted in session history"
       );
     }
     return { ...message, content };
   });
 }
 
+// src/session-input.ts
+var MODEL_INPUT_PART_TYPES = /* @__PURE__ */ new Set(["text", "image", "audio"]);
+var sessionInputMarkers = /* @__PURE__ */ new WeakMap();
+function isModelInputParts(input) {
+  return Array.isArray(input) && input.length > 0 && input.every(
+    (part) => typeof part === "object" && part !== null && MODEL_INPUT_PART_TYPES.has(part.type)
+  );
+}
+function prepareSessionInput(input) {
+  if (typeof input === "string") return { content: input, normalizedInput: input };
+  if (isModelInputParts(input)) {
+    const normalizedInput = normalizeModelInput(input);
+    return { content: summarizeModelInput(normalizedInput), normalizedInput };
+  }
+  const content = JSON.stringify(input);
+  return {
+    content,
+    ...typeof content === "string" ? { normalizedInput: content } : {}
+  };
+}
+function registerSessionInput(history, message, normalizedInput) {
+  sessionInputMarkers.set(history, {
+    message,
+    content: message.content,
+    normalizedInput: cloneModelInput(normalizedInput)
+  });
+}
+function clearSessionInput(history) {
+  sessionInputMarkers.delete(history);
+}
+function equalBytes(left, right) {
+  if (left.byteLength !== right.byteLength) return false;
+  for (let index = 0; index < left.byteLength; index++) {
+    if (left[index] !== right[index]) return false;
+  }
+  return true;
+}
+function equalSource(left, right) {
+  if (left.type !== right.type) return false;
+  switch (left.type) {
+    case "url":
+      return right.type === "url" && left.url === right.url && left.mediaType === right.mediaType;
+    case "bytes":
+      return right.type === "bytes" && left.mediaType === right.mediaType && equalBytes(left.data, right.data);
+    case "base64":
+      return right.type === "base64" && left.data === right.data && left.mediaType === right.mediaType;
+    case "provider-file":
+      return right.type === "provider-file" && left.provider === right.provider && left.reference === right.reference && left.mediaType === right.mediaType;
+  }
+}
+function equalPart(left, right) {
+  if (left.type !== right.type) return false;
+  if (left.type === "text") return right.type === "text" && left.text === right.text;
+  return right.type === left.type && left.label === right.label && equalSource(left.source, right.source);
+}
+function equalModelInput(left, right) {
+  if (typeof left === "string" || typeof right === "string") return left === right;
+  return left.length === right.length && left.every((part, index) => equalPart(part, right[index]));
+}
+function sessionHistoryForAsk(history, normalizedInput) {
+  const snapshot = history.slice();
+  const marker = sessionInputMarkers.get(history);
+  if (!marker) return snapshot;
+  const last = snapshot[snapshot.length - 1];
+  if (last !== marker.message || last.role !== "user" || last.content !== marker.content || !equalModelInput(marker.normalizedInput, normalizedInput)) {
+    return snapshot;
+  }
+  snapshot.pop();
+  return snapshot;
+}
+
 // src/partial-json.ts
 var MAX_DEPTH = 256;
 function parsePartialJson(input) {
@@ -10003,7 +10619,7 @@ function estimateMessagesTokens(messages) {
   for (const msg of messages) {
     total += estimateTokens(inputText(msg.content));
     if (typeof msg.content !== "string") {
-      unmeasured ||= msg.content.some((part) => part.type === "image");
+      unmeasured ||= msg.content.some((part) => part.type !== "text");
     }
     if (msg.tool_calls) {
       for (const tc of msg.tool_calls) {
@@ -10049,10 +10665,6 @@ function normalizeSessionHistory(messages) {
 function hasRichMessage(messages) {
   return messages.some((message) => typeof message.content !== "string");
 }
-function firstImageSource(input, history) {
-  const find = (value) => typeof value === "string" ? void 0 : value.find((part) => part.type === "image")?.source.type;
-  return find(input) ?? history.map((message) => find(message.content)).find(Boolean);
-}
 function appendHandoffInstruction(input, originalText, instruction) {
   if (typeof input === "string" || instruction.length === 0 || instruction === originalText)
     return input;
@@ -10610,7 +11222,8 @@ var WorkflowContext = class _WorkflowContext {
   // ── ctx.ask() ─────────────────────────────────────────────────────────
   async ask(agent2, prompt, options) {
     const normalizedInput = normalizeModelInput(prompt);
-    const normalizedHistory = normalizeSessionHistory(this.sessionHistory);
+    const sessionHistorySnapshot = sessionHistoryForAsk(this.sessionHistory, normalizedInput);
+    const normalizedHistory = normalizeSessionHistory(sessionHistorySnapshot);
     const delegateInputHolder = agent2;
     if (delegateInputHolder._delegateOriginalInput !== void 0) {
       delegateInputHolder._delegateOriginalInput = normalizeModelInput(
@@ -10820,20 +11433,31 @@ var WorkflowContext = class _WorkflowContext {
     const providerOptions = options?.providerOptions ?? agent2._config.providerOptions;
     const richRequest = typeof input !== "string" || hasRichMessage(sessionHistory);
     if (richRequest) {
+      const richPart = firstRichPart(input, sessionHistory);
+      const modality = richPart?.type ?? "image";
+      const source = richPart?.source.type;
       if (providerOptions && ("messages" in providerOptions || "input" in providerOptions)) {
         throw new UnsupportedModelInputError({
           provider: provider.name ?? modelUri.split(":", 1)[0],
           model,
-          modality: "image",
+          modality,
           feature: "raw input-container providerOptions"
         });
       }
-      const source = firstImageSource(input, sessionHistory);
+      const audioPart = firstRichPart(input, sessionHistory, "audio");
+      if (audioPart && !provider.inputCapabilities?.(model)?.audio) {
+        throw new UnsupportedModelInputError({
+          provider: provider.name ?? modelUri.split(":", 1)[0],
+          model,
+          modality: "audio",
+          source: audioPart.source.type
+        });
+      }
       if (!provider.validateInput) {
         throw new UnsupportedModelInputError({
           provider: provider.name ?? modelUri.split(":", 1)[0],
           model,
-          modality: "image",
+          modality,
           ...source ? { source } : {}
         });
       }
@@ -10852,10 +11476,20 @@ var WorkflowContext = class _WorkflowContext {
         throw new UnsupportedModelInputError({
           provider: provider.name ?? modelUri.split(":", 1)[0],
           model,
-          modality: "image",
+          modality,
           ...source ? { source } : {}
         });
       }
+      if (audioPart && validation.effectiveModel !== model) {
+        if (!provider.inputCapabilities?.(validation.effectiveModel)?.audio) {
+          throw new UnsupportedModelInputError({
+            provider: provider.name ?? modelUri.split(":", 1)[0],
+            model: validation.effectiveModel,
+            modality: "audio",
+            source: audioPart.source.type
+          });
+        }
+      }
       model = validation.effectiveModel;
     }
     const effectiveModelUri = !richRequest ? modelUri : `${modelUri.slice(0, modelUri.indexOf(":") + 1)}${model}`;
@@ -10868,25 +11502,17 @@ var WorkflowContext = class _WorkflowContext {
         (descriptor) => descriptor !== void 0
       );
       const parts = descriptors.flatMap((descriptor) => descriptor.parts);
+      const media = parts.filter((part) => part.type !== "text");
+      const withSource = (source) => media.filter((part) => part.source === source).length;
       this.spanManager?.addEventToActiveSpan("axl.model_input", {
         "axl.input.parts": parts.length,
         "axl.input.images": parts.filter((part) => part.type === "image").length,
-        "axl.input.source.bytes": parts.filter(
-          (part) => part.type === "image" && part.source === "bytes"
-        ).length,
-        "axl.input.source.base64": parts.filter(
-          (part) => part.type === "image" && part.source === "base64"
-        ).length,
-        "axl.input.source.url": parts.filter(
-          (part) => part.type === "image" && part.source === "url"
-        ).length,
-        "axl.input.source.provider_file": parts.filter(
-          (part) => part.type === "image" && part.source === "provider-file"
-        ).length,
-        "axl.input.inline_bytes": parts.reduce(
-          (sum, part) => sum + (part.type === "image" ? part.bytes ?? 0 : 0),
-          0
-        )
+        "axl.input.audio": parts.filter((part) => part.type === "audio").length,
+        "axl.input.source.bytes": withSource("bytes"),
+        "axl.input.source.base64": withSource("base64"),
+        "axl.input.source.url": withSource("url"),
+        "axl.input.source.provider_file": withSource("provider-file"),
+        "axl.input.inline_bytes": media.reduce((sum, part) => sum + (part.bytes ?? 0), 0)
       });
     }
     assertRichMessagesAreUserOnly(sessionHistory);
@@ -13453,6 +14079,11 @@ ${summaryResponse.content}`
       }
     });
     const routerInput = options?.routerInput === "text" ? inputText(prompt) : prompt;
+    if (routerInput === "") {
+      throw new InvalidModelInputError(
+        "delegate routerInput 'text' requires at least one text part; the input is media-only \u2014 route on the full input or add a text part"
+      );
+    }
     routerAgent._delegateOriginalInput = prompt;
     return this.ask(routerAgent, routerInput, {
       schema: options?.schema,
@@ -14549,9 +15180,13 @@ var Session = class _Session {
   }
   async sendImpl(workflowName, input, signal, events) {
     const { history, metadata } = await this.prepareHistory(input);
-    const result = await this.runtime.execute(workflowName, input, { metadata, signal, events });
-    await this.commitHistory(history, result);
-    return result;
+    try {
+      const result = await this.runtime.execute(workflowName, input, { metadata, signal, events });
+      await this.commitHistory(history, result);
+      return result;
+    } finally {
+      clearSessionInput(history);
+    }
   }
   /** Read history + summary, apply maxMessages limit (with optional
    *  summarization), and push the new user message. Returns the mutable
@@ -14561,6 +15196,7 @@ var Session = class _Session {
    *  `streamImpl` so future fixes (e.g., to summarization wiring) land in
    *  one place. */
   async prepareHistory(input) {
+    const preparedInput = prepareSessionInput(input);
     const history = await this.store.getSession(this.sessionId);
     let cachedSummary = await this.store.getSessionMeta(this.sessionId, "summaryCache");
     const maxMessages = this.options.history?.maxMessages;
@@ -14583,10 +15219,11 @@ var Session = class _Session {
       history.length = 0;
       history.push(...trimmed);
     }
-    history.push({
-      role: "user",
-      content: typeof input === "string" ? input : JSON.stringify(input)
-    });
+    const userMessage = { role: "user", content: preparedInput.content };
+    history.push(userMessage);
+    if (this.options.deduplicateInput !== false && preparedInput.normalizedInput !== void 0) {
+      registerSessionInput(history, userMessage, preparedInput.normalizedInput);
+    }
     const metadata = {
       sessionId: this.sessionId,
       sessionHistory: history,
@@ -14645,21 +15282,26 @@ var Session = class _Session {
         events
       });
     } catch (err) {
+      if (history) clearSessionInput(history);
       rejectReady(err);
       throw err;
     }
     resolveReady(axlStream);
-    const completion = axlStream.promise.then(
-      (result) => this.commitHistory(history, result).catch((err) => {
-        this.runtime.emit("error", {
-          type: "session_history_save_failed",
-          sessionId: this.sessionId,
-          error: err instanceof Error ? err.message : String(err)
-        });
-      }),
-      () => void 0
-    );
-    await completion;
+    try {
+      const completion = axlStream.promise.then(
+        (result) => this.commitHistory(history, result).catch((err) => {
+          this.runtime.emit("error", {
+            type: "session_history_save_failed",
+            sessionId: this.sessionId,
+            error: err instanceof Error ? err.message : String(err)
+          });
+        }),
+        () => void 0
+      );
+      await completion;
+    } finally {
+      clearSessionInput(history);
+    }
   }
   async history() {
     return this.store.getSession(this.sessionId);
@@ -18274,6 +18916,7 @@ function cosineSimilarity2(a, b) {
   resolveRuntime,
   resolveThinkingOptions,
   stringStreamFromEvents,
+  summarizeModelInput,
   tool,
   workflow,
   zodToJsonSchema