@byfriends/kosong 0.1.0 → 0.2.0

package/dist/{openai-common-08qin3UI.mjs → openai-common-Dl42y_vn.mjs}

@@ -171,16 +171,41 @@ function isFunctionToolCall(tc) {
 	return tc.type === "function";
 }
 /**
+* Model name prefixes / exact names known to support the `xhigh` reasoning
+* effort level.  All other OpenAI-compatible models clamp `xhigh` / `max`
+* down to `high`.
+*/
+const XHIGH_SUPPORT_PREFIXES = [
+	"gpt-5.",
+	"gpt-5-",
+	"o3-pro",
+	"o4-mini"
+];
+function supportsXhighReasoningEffort(model) {
+	const normalized = model.toLowerCase();
+	return XHIGH_SUPPORT_PREFIXES.some((prefix) => normalized.startsWith(prefix));
+}
+/**
 * Map kosong `ThinkingEffort` to OpenAI `reasoning_effort` string.
+*
+* When `model` is provided, `xhigh` / `max` are clamped to `'high'` with a
+* `console.warn` if the model is not known to support the `xhigh` effort
+* level.  When `model` is omitted the mapping is pass-through (backward
+* compatible).
 */
-function thinkingEffortToReasoningEffort(effort) {
+function thinkingEffortToReasoningEffort(effort, model) {
 	switch (effort) {
 		case "off": return;
 		case "low": return "low";
 		case "medium": return "medium";
 		case "high": return "high";
 		case "xhigh":
-		case "max": return "xhigh";
+		case "max":
+			if (model !== void 0 && !supportsXhighReasoningEffort(model)) {
+				console.warn(`effort '${effort}' clamped to 'high' for model ${model}`);
+				return "high";
+			}
+			return "xhigh";
 		default: throw new Error(`Unknown thinking effort: ${String(effort)}`);
 	}
 }

package/dist/{openai-common-B6cK2ig3.d.mts → openai-common-DwkxUSyI.d.mts}

@@ -1,5 +1,4 @@
-import { b as Message, c as TokenUsage, n as FinishReason, o as ThinkingEffort, p as Tool, v as ContentPart } from "./provider-DiJKWMsQ.mjs";
-import { o as ChatProviderError } from "./errors-DweKbIOf.mjs";
+import { M as ContentPart, P as Message, c as FinishReason, f as ThinkingEffort, h as TokenUsage, o as ChatProviderError, x as Tool } from "./errors-wlT14tC4.mjs";
 
 //#region src/providers/openai-common.d.ts
 interface OpenAIContentPart {
@@ -58,8 +57,13 @@ declare function isFunctionToolCall<T extends {
 }>(tc: T): tc is T & FunctionToolCallShape;
 /**
  * Map kosong `ThinkingEffort` to OpenAI `reasoning_effort` string.
+ *
+ * When `model` is provided, `xhigh` / `max` are clamped to `'high'` with a
+ * `console.warn` if the model is not known to support the `xhigh` effort
+ * level.  When `model` is omitted the mapping is pass-through (backward
+ * compatible).
  */
-declare function thinkingEffortToReasoningEffort(effort: ThinkingEffort): string | undefined;
+declare function thinkingEffortToReasoningEffort(effort: ThinkingEffort, model?: string): string | undefined;
 /**
  * Map OpenAI `reasoning_effort` string back to kosong `ThinkingEffort`.
  */

package/dist/{openai-responses-BxOwxtd3.d.mts → openai-responses-DZ9mQ5RA.d.mts}

@@ -1,5 +1,5 @@
-import { S as StreamedMessagePart, a as StreamedMessage, b as Message, c as TokenUsage, i as ProviderRequestAuth, m as ModelCapability, n as FinishReason, o as ThinkingEffort, p as Tool, r as GenerateOptions, t as ChatProvider } from "./provider-DiJKWMsQ.mjs";
-import { i as ToolMessageConversion } from "./openai-common-B6cK2ig3.mjs";
+import { I as StreamedMessagePart, P as Message, S as ModelCapability, c as FinishReason, d as StreamedMessage, f as ThinkingEffort, h as TokenUsage, l as GenerateOptions, s as ChatProvider, u as ProviderRequestAuth, x as Tool } from "./errors-wlT14tC4.mjs";
+import { i as ToolMessageConversion } from "./openai-common-DwkxUSyI.mjs";
 import OpenAI from "openai";
 
 //#region src/providers/openai-responses.d.ts

package/dist/providers/anthropic.d.mts

@@ -1,2 +1,2 @@
-import { i as resolveDefaultMaxTokens, n as AnthropicOptions, r as convertAnthropicError, t as AnthropicChatProvider } from "../anthropic-Dm_GqFgS.mjs";
+import { i as resolveDefaultMaxTokens, n as AnthropicOptions, r as convertAnthropicError, t as AnthropicChatProvider } from "../anthropic-0CVG5rBE.mjs";
 export { AnthropicChatProvider, AnthropicOptions, convertAnthropicError, resolveDefaultMaxTokens };

package/dist/providers/anthropic.mjs

@@ -1,6 +1,5 @@
 import { a as APITimeoutError, o as ChatProviderError, s as normalizeAPIStatusError, t as APIConnectionError$1 } from "../errors-WFxxzL1B.mjs";
-import { n as requireProviderApiKey, r as resolveAuthBackedClient, t as mergeRequestHeaders } from "../request-auth-DCWSyCKI.mjs";
-import { t as getAnthropicModelCapability } from "../capability-registry-CMBuEYcf.mjs";
+import { i as getAnthropicModelCapability, n as requireProviderApiKey, r as resolveAuthBackedClient, t as mergeRequestHeaders } from "../request-auth-BMXt8jRu.mjs";
 import Anthropic, { APIConnectionError, APIConnectionTimeoutError, APIError, AnthropicError } from "@anthropic-ai/sdk";
 //#region src/providers/anthropic.ts
 /**
@@ -170,34 +169,24 @@ function supportsAdaptiveThinking(model) {
 	if (match === null) return false;
 	return versionAtLeast(parseVersion(match), ADAPTIVE_MIN_VERSION);
 }
-function isOpus47(model) {
+function supportsXhigh(model) {
 	const match = OPUS_VERSION_RE.exec(model.toLowerCase());
 	if (match === null) return false;
 	const version = parseVersion(match);
-	return version.major === 4 && version.minor === 7;
-}
-function supportsEffortParam(model) {
-	if (supportsAdaptiveThinking(model)) return true;
-	const normalized = model.toLowerCase();
-	return normalized.includes("opus-4-5") || normalized.includes("opus-4.5");
+	return version.major === 4 && (version.minor === 7 || version.minor === 8);
 }
 function clampEffort(effort, model) {
 	if (effort === "off") return effort;
-	if (effort === "xhigh" && !isOpus47(model)) return "high";
-	if (effort === "max" && !supportsAdaptiveThinking(model)) return "high";
+	if (effort === "xhigh" && !supportsXhigh(model)) {
+		console.warn(`effort 'xhigh' clamped to 'high' for model ${model}`);
+		return "high";
+	}
+	if (effort === "max" && !supportsAdaptiveThinking(model)) {
+		console.warn(`effort 'max' clamped to 'high' for model ${model}`);
+		return "high";
+	}
 	return effort;
 }
-function budgetTokensForEffort(effort) {
-	switch (effort) {
-		case "low": return 1024;
-		case "medium": return 4096;
-		case "high": return 32e3;
-		case "off":
-		case "xhigh":
-		case "max": throw new Error(`Unsupported budget-based thinking effort: ${effort}`);
-	}
-	throw new Error(`Unknown thinking effort: ${String(effort)}`);
-}
 const CACHE_CONTROL = { type: "ephemeral" };
 /**
 * Content block types that support cache_control injection.
@@ -212,14 +201,22 @@ const CACHEABLE_TYPES = new Set([
 	"server_tool_use",
 	"web_search_tool_result"
 ]);
-function injectCacheControlOnLastBlock(messages) {
-	const lastMessage = messages.at(-1);
-	if (lastMessage === void 0) return;
-	const content = lastMessage.content;
-	if (!Array.isArray(content) || content.length === 0) return;
-	const lastBlock = content.at(-1);
-	if (lastBlock === void 0) return;
-	if (CACHEABLE_TYPES.has(lastBlock.type)) lastBlock.cache_control = CACHE_CONTROL;
+/**
+* Convert a PromptPlan to Anthropic TextBlockParam[] with cache control.
+*
+* Injects cache_control on blocks with cacheScope other than 'none'.
+*/
+function promptPlanToSystemBlocks(promptPlan) {
+	const blocks = [];
+	for (const block of promptPlan.blocks) {
+		const textBlock = {
+			type: "text",
+			text: block.text
+		};
+		if (block.cacheScope !== "none") textBlock.cache_control = CACHE_CONTROL;
+		blocks.push(textBlock);
+	}
+	return blocks;
 }
 /**
 * Check whether a MessageParam is a user message whose content consists
@@ -300,9 +297,11 @@ function convertMessage(message) {
 	};
 	if (role === "tool") {
 		if (message.toolCallId === void 0) throw new ChatProviderError("Tool message missing `toolCallId`.");
+		const block = toolResultToBlock(message.toolCallId, message.content);
+		if (message.cacheHint?.isLastTurnEnd || message.cacheHint?.isSuddenLargeContext) block.cache_control = CACHE_CONTROL;
 		return {
 			role: "user",
-			content: [toolResultToBlock(message.toolCallId, message.content)]
+			content: [block]
 		};
 	}
 	const blocks = [];
@@ -336,6 +335,10 @@ function convertMessage(message) {
 			input: toolInput
 		});
 	}
+	if (message.cacheHint?.isLastTurnEnd || message.cacheHint?.isSuddenLargeContext) {
+		const lastBlock = blocks.at(-1);
+		if (lastBlock !== void 0 && CACHEABLE_TYPES.has(lastBlock.type)) lastBlock.cache_control = CACHE_CONTROL;
+	}
 	return {
 		role,
 		content: blocks
@@ -563,20 +566,15 @@ var AnthropicChatProvider = class {
 		const thinkingConfig = this._generationKwargs.thinking;
 		if (thinkingConfig === void 0 || thinkingConfig === null) return null;
 		if (thinkingConfig.type === "disabled") return "off";
-		if (thinkingConfig.type === "adaptive") {
-			const effort = this._generationKwargs.output_config?.effort;
-			if (effort === void 0 || effort === null) return "high";
-			switch (effort) {
-				case "low":
-				case "medium":
-				case "high":
-				case "xhigh":
-				case "max": return effort;
-			}
+		const effort = this._generationKwargs.output_config?.effort;
+		if (effort === void 0 || effort === null) return "high";
+		switch (effort) {
+			case "low":
+			case "medium":
+			case "high":
+			case "xhigh":
+			case "max": return effort;
 		}
-		const budget = thinkingConfig.budget_tokens ?? 0;
-		if (budget <= 1024) return "low";
-		if (budget <= 4096) return "medium";
 		return "high";
 	}
 	get modelParameters() {
@@ -589,11 +587,16 @@ var AnthropicChatProvider = class {
 		return getAnthropicModelCapability(model ?? this._model);
 	}
 	async generate(systemPrompt, tools, history, options) {
-		const system = systemPrompt ? [{
+		let system;
+		if (options?.promptPlan) {
+			const systemBlocks = promptPlanToSystemBlocks(options.promptPlan);
+			system = systemBlocks.length > 0 ? systemBlocks : void 0;
+		} else if (systemPrompt) system = [{
 			type: "text",
 			text: systemPrompt,
 			cache_control: CACHE_CONTROL
-		}] : void 0;
+		}];
+		else system = void 0;
 		const messages = [];
 		for (const msg of history) {
 			const converted = convertMessage(msg);
@@ -601,7 +604,6 @@ var AnthropicChatProvider = class {
 			if (last !== void 0 && isToolResultOnly(last) && isToolResultOnly(converted)) last.content = [...last.content, ...converted.content];
 			else messages.push(converted);
 		}
-		injectCacheControlOnLastBlock(messages);
 		const kwargs = {};
 		if (this._generationKwargs.max_tokens !== void 0) kwargs["max_tokens"] = this._generationKwargs.max_tokens;
 		if (this._generationKwargs.temperature !== void 0) kwargs["temperature"] = this._generationKwargs.temperature;
@@ -664,7 +666,7 @@ var AnthropicChatProvider = class {
 	withThinking(effort) {
 		if (effort === "off") {
 			let newBetas = [...this._generationKwargs.betaFeatures ?? []];
-			if (supportsAdaptiveThinking(this._model)) newBetas = newBetas.filter((b) => b !== INTERLEAVED_THINKING_BETA);
+			newBetas = newBetas.filter((b) => b !== INTERLEAVED_THINKING_BETA);
 			const clone = this._withGenerationKwargs({
 				thinking: { type: "disabled" },
 				betaFeatures: newBetas
@@ -675,29 +677,15 @@ var AnthropicChatProvider = class {
 		const effectiveEffort = clampEffort(effort, this._model);
 		if (effectiveEffort === "off") throw new Error("Non-off thinking effort unexpectedly clamped to off.");
 		let newBetas = [...this._generationKwargs.betaFeatures ?? []];
-		if (supportsAdaptiveThinking(this._model)) {
-			newBetas = newBetas.filter((b) => b !== INTERLEAVED_THINKING_BETA);
-			return this._withGenerationKwargs({
-				thinking: {
-					type: "adaptive",
-					display: "summarized"
-				},
-				output_config: { effort: effectiveEffort },
-				betaFeatures: newBetas
-			});
-		}
-		const kwargs = {
+		newBetas = newBetas.filter((b) => b !== INTERLEAVED_THINKING_BETA);
+		return this._withGenerationKwargs({
 			thinking: {
-				type: "enabled",
-				budget_tokens: budgetTokensForEffort(effectiveEffort)
+				type: "adaptive",
+				display: "summarized"
 			},
+			output_config: { effort: effectiveEffort },
 			betaFeatures: newBetas
-		};
-		if (supportsEffortParam(this._model)) kwargs.output_config = { effort: effectiveEffort };
-		else kwargs.output_config = void 0;
-		const clone = this._withGenerationKwargs(kwargs);
-		if (!supportsEffortParam(this._model)) delete clone._generationKwargs.output_config;
-		return clone;
+		});
 	}
 	withGenerationKwargs(kwargs) {
 		return this._withGenerationKwargs(kwargs);

package/dist/providers/google-genai.d.mts

@@ -1,2 +1,2 @@
-import { a as convertGoogleGenAIError, i as GoogleGenAIStreamedMessage, n as GoogleGenAIGenerationKwargs, o as messagesToGoogleGenAIContents, r as GoogleGenAIOptions, t as GoogleGenAIChatProvider } from "../google-genai-hX0X6CF3.mjs";
+import { a as convertGoogleGenAIError, i as GoogleGenAIStreamedMessage, n as GoogleGenAIGenerationKwargs, o as messagesToGoogleGenAIContents, r as GoogleGenAIOptions, t as GoogleGenAIChatProvider } from "../google-genai-xKK8lI_R.mjs";
 export { GoogleGenAIChatProvider, GoogleGenAIGenerationKwargs, GoogleGenAIOptions, GoogleGenAIStreamedMessage, convertGoogleGenAIError, messagesToGoogleGenAIContents };

package/dist/providers/google-genai.mjs

@@ -1,6 +1,5 @@
 import { a as APITimeoutError, o as ChatProviderError, s as normalizeAPIStatusError, t as APIConnectionError } from "../errors-WFxxzL1B.mjs";
-import { n as requireProviderApiKey, r as resolveAuthBackedClient } from "../request-auth-DCWSyCKI.mjs";
-import { n as getGoogleGenAIModelCapability } from "../capability-registry-CMBuEYcf.mjs";
+import { a as getGoogleGenAIModelCapability, n as requireProviderApiKey, r as resolveAuthBackedClient } from "../request-auth-BMXt8jRu.mjs";
 import { ApiError, GoogleGenAI } from "@google/genai";
 //#region src/providers/google-genai.ts
 /**

package/dist/providers/openai-common.d.mts

@@ -1,2 +1,2 @@
-import { a as convertContentPart, c as extractUsage, d as reasoningEffortToThinkingEffort, f as thinkingEffortToReasoningEffort, i as ToolMessageConversion, l as isFunctionToolCall, n as OpenAIContentPart, o as convertOpenAIError, p as toolToOpenAI, r as OpenAIToolParam, s as convertToolMessageContent, t as FunctionToolCallShape, u as normalizeOpenAIFinishReason } from "../openai-common-B6cK2ig3.mjs";
+import { a as convertContentPart, c as extractUsage, d as reasoningEffortToThinkingEffort, f as thinkingEffortToReasoningEffort, i as ToolMessageConversion, l as isFunctionToolCall, n as OpenAIContentPart, o as convertOpenAIError, p as toolToOpenAI, r as OpenAIToolParam, s as convertToolMessageContent, t as FunctionToolCallShape, u as normalizeOpenAIFinishReason } from "../openai-common-DwkxUSyI.mjs";
 export { FunctionToolCallShape, OpenAIContentPart, OpenAIToolParam, ToolMessageConversion, convertContentPart, convertOpenAIError, convertToolMessageContent, extractUsage, isFunctionToolCall, normalizeOpenAIFinishReason, reasoningEffortToThinkingEffort, thinkingEffortToReasoningEffort, toolToOpenAI };

package/dist/providers/openai-common.mjs

@@ -1,2 +1,2 @@
-import { a as isFunctionToolCall, c as thinkingEffortToReasoningEffort, i as extractUsage, l as toolToOpenAI, n as convertOpenAIError, o as normalizeOpenAIFinishReason, r as convertToolMessageContent, s as reasoningEffortToThinkingEffort, t as convertContentPart } from "../openai-common-08qin3UI.mjs";
+import { a as isFunctionToolCall, c as thinkingEffortToReasoningEffort, i as extractUsage, l as toolToOpenAI, n as convertOpenAIError, o as normalizeOpenAIFinishReason, r as convertToolMessageContent, s as reasoningEffortToThinkingEffort, t as convertContentPart } from "../openai-common-Dl42y_vn.mjs";
 export { convertContentPart, convertOpenAIError, convertToolMessageContent, extractUsage, isFunctionToolCall, normalizeOpenAIFinishReason, reasoningEffortToThinkingEffort, thinkingEffortToReasoningEffort, toolToOpenAI };

package/dist/providers/openai-responses.d.mts

@@ -1,2 +1,2 @@
-import { i as OpenAIResponsesStreamedMessage, n as OpenAIResponsesGenerationKwargs, r as OpenAIResponsesOptions, t as OpenAIResponsesChatProvider } from "../openai-responses-BxOwxtd3.mjs";
+import { i as OpenAIResponsesStreamedMessage, n as OpenAIResponsesGenerationKwargs, r as OpenAIResponsesOptions, t as OpenAIResponsesChatProvider } from "../openai-responses-DZ9mQ5RA.mjs";
 export { OpenAIResponsesChatProvider, OpenAIResponsesGenerationKwargs, OpenAIResponsesOptions, OpenAIResponsesStreamedMessage };

package/dist/providers/openai-responses.mjs

@@ -1,8 +1,8 @@
-import { c as thinkingEffortToReasoningEffort, n as convertOpenAIError, p as extractText, s as reasoningEffortToThinkingEffort } from "../openai-common-08qin3UI.mjs";
+import { c as thinkingEffortToReasoningEffort, n as convertOpenAIError, p as extractText, s as reasoningEffortToThinkingEffort } from "../openai-common-Dl42y_vn.mjs";
 import { o as ChatProviderError } from "../errors-WFxxzL1B.mjs";
-import { n as requireProviderApiKey, r as resolveAuthBackedClient, t as mergeRequestHeaders } from "../request-auth-DCWSyCKI.mjs";
-import { a as usesOpenAIResponsesDeveloperRole, i as getOpenAIResponsesModelCapability } from "../capability-registry-CMBuEYcf.mjs";
+import { l as usesOpenAIResponsesDeveloperRole, n as requireProviderApiKey, r as resolveAuthBackedClient, s as getOpenAIResponsesModelCapability, t as mergeRequestHeaders } from "../request-auth-BMXt8jRu.mjs";
 import OpenAI from "openai";
+import { createHash } from "node:crypto";
 //#region src/providers/openai-responses.ts
 /**
 * Normalize the Responses API status / incomplete_details into the unified
@@ -45,6 +45,19 @@ function normalizeResponsesFinishReason(status, incompleteReason) {
 		rawFinishReason: null
 	};
 }
+/**
+* Derive a stable cache key from cacheable blocks in a PromptPlan.
+*
+* Only blocks with cacheScope 'global' are included in the hash, as OpenAI
+* only supports caching the prefix (global scope).
+*/
+function deriveCacheKeyFromPromptPlan(promptPlan) {
+	if (!promptPlan || promptPlan.blocks.length === 0) return void 0;
+	const cacheableTexts = [];
+	for (const block of promptPlan.blocks) if (block.cacheScope === "global") cacheableTexts.push(block.text);
+	if (cacheableTexts.length === 0) return void 0;
+	return createHash("sha256").update(cacheableTexts.join("")).digest("hex");
+}
 function asRawObject(value) {
 	if (value === null || typeof value !== "object" || Array.isArray(value)) return null;
 	return value;
@@ -574,6 +587,10 @@ var OpenAIResponsesChatProvider = class {
 				stream: this._stream,
 				...kwargs
 			};
+			if (options?.promptPlan) {
+				const cacheKey = deriveCacheKeyFromPromptPlan(options.promptPlan);
+				if (cacheKey) createParams["prompt_cache_key"] = cacheKey;
+			}
 			if (!("responses" in client) || typeof client.responses?.create !== "function") throw new Error("OpenAI SDK version does not support Responses API. Upgrade to >=4.x with responses support.");
 			return new OpenAIResponsesStreamedMessage(await client.responses.create(createParams, options?.signal ? { signal: options.signal } : void 0), this._stream);
 		} catch (error) {
@@ -581,7 +598,7 @@ var OpenAIResponsesChatProvider = class {
 		}
 	}
 	withThinking(effort) {
-		const reasoningEffort = thinkingEffortToReasoningEffort(effort);
+		const reasoningEffort = thinkingEffortToReasoningEffort(effort, this._model);
 		const clone = this._clone();
 		clone._generationKwargs = {
 			...clone._generationKwargs,