smoltalk 0.17.0 → 0.18.0

package/dist/classes/message/BaseMessage.d.ts

@@ -2,8 +2,12 @@ import { Content } from "@google/genai";
 import { Message } from "ollama";
 import { ChatCompletionMessageParam } from "openai/resources";
 import type { ResponseInputItem } from "openai/resources/responses/responses.js";
+/** The text a provider sees for a message's content: a string as is, any
+ *  other value as its JSON text, nothing for null. Exported so a consumer
+ *  that compares saved messages can apply the same rule. */
+export declare function contentToString(content: unknown): string;
 export declare class BaseMessage {
-    protected contentToString(content: string | unknown[] | null | undefined): string;
+    protected contentToString(content: unknown): string;
 }
 export interface MessageClass {
     get content(): string;

package/dist/classes/message/BaseMessage.js

@@ -1,7 +1,13 @@
+/** The text a provider sees for a message's content: a string as is, any
+ *  other value as its JSON text, nothing for null. Exported so a consumer
+ *  that compares saved messages can apply the same rule. */
+export function contentToString(content) {
+    if (content === null || content === undefined)
+        return "";
+    return typeof content === "string" ? content : JSON.stringify(content);
+}
 export class BaseMessage {
     contentToString(content) {
-        if (content === null || content === undefined)
-            return "";
-        return typeof content === "string" ? content : JSON.stringify(content);
+        return contentToString(content);
     }
 }

package/dist/classes/message/ToolMessage.d.ts

@@ -1,6 +1,5 @@
 import { z } from "zod";
 import { BaseMessage, MessageClass } from "./BaseMessage.js";
-import { TextPart } from "../../types.js";
 import { ChatCompletionMessageParam } from "openai/resources";
 import { Content } from "@google/genai";
 import { Message } from "ollama";
@@ -21,11 +20,11 @@ export type ToolMessageJSON = {
 };
 export declare class ToolMessage extends BaseMessage implements MessageClass {
     _role: "tool";
-    _content: string | Array<TextPart>;
+    _content: unknown;
     _tool_call_id: string;
     _rawData?: any;
     _name: string;
-    constructor(content: string | Array<TextPart>, options: {
+    constructor(content: unknown, options: {
         tool_call_id: string;
         rawData?: any;
         name: string;

package/dist/classes/message/ToolMessage.js

@@ -1,6 +1,5 @@
 import { z } from "zod";
 import { BaseMessage } from "./BaseMessage.js";
-import { TextPartSchema } from "../../types.js";
 import { getLogger } from "../../util/logger.js";
 export const ToolMessageJSONSchema = z.object({
     role: z.literal("tool"),
@@ -12,6 +11,9 @@ export const ToolMessageJSONSchema = z.object({
 });
 export class ToolMessage extends BaseMessage {
     _role = "tool";
+    // A tool result is whatever the tool returned: text, text parts, or a
+    // plain value such as an array of records. Every provider conversion
+    // reads it through the `content` getter, which turns any value into text.
     _content;
     _tool_call_id;
     _rawData;
@@ -59,18 +61,10 @@ export class ToolMessage extends BaseMessage {
             logger.debug("ToolMessage payload that failed to parse:", JSON.stringify(json, null, 2));
             throw new Error("Failed to parse ToolMessage");
         }
-        const TextPartArraySchema = z.array(TextPartSchema);
-        const textPartArrayResult = TextPartArraySchema.safeParse(result.data.content);
-        if (textPartArrayResult.success) {
-            result.data.content = textPartArrayResult.data;
-        }
-        else if (typeof result.data.content === "string") {
-            // do nothing, it's already a string
-        }
-        else {
-            getLogger().debug("ToolMessage content is neither a string nor an array of TextParts. Converting to string using JSON.stringify.");
-            result.data.content = JSON.stringify(result.data.content);
-        }
+        // Content is kept exactly as saved. Rebuilding a non-text value as its
+        // JSON text here would make a saved-and-rebuilt thread differ from one
+        // that was never saved, while what a provider sees is the same either
+        // way (the `content` getter stringifies both).
         return new ToolMessage(result.data.content, {
             tool_call_id: result.data.tool_call_id,
             name: result.data.name,

package/dist/classes/message/index.d.ts

@@ -51,7 +51,7 @@ export declare function systemMessage(content: string | Array<TextPart>, options
     name?: string;
     rawData?: any;
 }): SystemMessage;
-export declare function toolMessage(content: string | Array<TextPart>, options: {
+export declare function toolMessage(content: unknown, options: {
     tool_call_id: string;
     rawData?: any;
     name: string;

package/dist/clients/google.js

@@ -223,9 +223,12 @@ export class SmolGoogle extends BaseClient {
         let cost;
         if (usageMetadata) {
             const cached = usageMetadata.cachedContentTokenCount ?? 0;
+            // Thinking is reported in thoughtsTokenCount, separate from
+            // candidatesTokenCount, and billed at the output rate.
+            const thoughts = usageMetadata.thoughtsTokenCount ?? 0;
             usage = {
                 inputTokens: Math.max(0, (usageMetadata.promptTokenCount || 0) - cached),
-                outputTokens: usageMetadata.candidatesTokenCount || 0,
+                outputTokens: (usageMetadata.candidatesTokenCount || 0) + thoughts,
                 totalTokens: usageMetadata.totalTokenCount,
             };
             if (cached > 0) {

package/dist/embed/ollama.js

@@ -1,4 +1,4 @@
-import { Ollama } from "ollama";
+import { Ollama } from "#ollama";
 import { success, failure } from "../types/result.js";
 import { DEFAULT_OLLAMA_HOST } from "../clients/ollama.js";
 export async function ollamaEmbed(inputs, config, apiKey, ollamaHost) {

package/dist/model.js

@@ -43,7 +43,27 @@ export class Model {
             return null;
         }
         // BaseModel token-rate fields, read structurally across the model union.
-        const rates = model;
+        const baseRates = model;
+        const rates = { ...baseRates };
+        // A prompt over the long-context threshold is billed at the long-context
+        // rates for the whole request, input and output alike. The prompt size
+        // counts every input bucket, cached ones included.
+        const promptTokens = usage.inputTokens +
+            (usage.cachedInputTokens ?? 0) +
+            (usage.cacheCreationInputTokens ?? 0) +
+            (usage.inputAudioTokens ?? 0);
+        const longContext = baseRates.longContext;
+        if (longContext !== undefined && promptTokens > longContext.thresholdTokens) {
+            if (longContext.inputTokenCost !== undefined) {
+                rates.inputTokenCost = longContext.inputTokenCost;
+            }
+            if (longContext.outputTokenCost !== undefined) {
+                rates.outputTokenCost = longContext.outputTokenCost;
+            }
+            if (longContext.cachedInputTokenCost !== undefined) {
+                rates.cachedInputTokenCost = longContext.cachedInputTokenCost;
+            }
+        }
         // Price only models that carry at least one token rate; those without
         // (per-minute STT, per-char TTS) return null so their dedicated helpers apply.
         if (rates.inputTokenCost === undefined &&

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "smoltalk",
-  "version": "0.17.0",
+  "version": "0.18.0",
   "description": "A common interface for LLM APIs",
   "homepage": "https://github.com/egonSchiele/smoltalk",
   "files": [