@codehz/ai 0.4.3 → 0.4.5

package/dist/index.mjs

@@ -80,6 +80,14 @@ const TOOL_RESULT_OUTCOMES = /* @__PURE__ */ new Set([
 	"rejected"
 ]);
 const INCLUDE_MODES = /* @__PURE__ */ new Set(["off", "best_effort"]);
+const REASONING_LEVELS$1 = /* @__PURE__ */ new Set([
+	"none",
+	"minimal",
+	"low",
+	"medium",
+	"high",
+	"xhigh"
+]);
 function isRecord(value) {
 	return typeof value === "object" && value !== null;
 }
@@ -248,6 +256,9 @@ function validateRequest(request) {
 			message: "maxOutputTokens must be a positive integer"
 		});
 	}
+	if (request.reasoningLevel !== void 0) {
+		if (typeof request.reasoningLevel !== "string" || !REASONING_LEVELS$1.has(request.reasoningLevel)) pushIssue(issues, "reasoningLevel", "REASONING_LEVEL_INVALID", "reasoningLevel must be one of: none, minimal, low, medium, high, xhigh");
+	}
 	if (request.include !== void 0) validateInclude(request.include, issues);
 	if (request.metadata !== void 0) {
 		if (!isRecord(request.metadata)) pushIssue(issues, "metadata", "METADATA_INVALID", "metadata must be an object");
@@ -1660,6 +1671,102 @@ function createCompletionGate() {
 	};
 }
 //#endregion
+//#region src/helpers/provider-request-options.ts
+/**
+* Provider 请求 headers / body 扩展合并
+*
+* 供真实后端 adapter 构造选项 `headers` / `extraBody` 使用：
+* - headers：内置鉴权头为基，自定义后写覆盖
+* - extraBody：已构建 body 为基，额外字段浅层 spread，同名顶层键可覆盖
+*/
+/** 合并内置 headers 与自定义 headers；自定义后写覆盖同名键。 */
+function mergeProviderHeaders(base, custom) {
+	if (!custom) return base;
+	return {
+		...base,
+		...custom
+	};
+}
+/**
+* 将构造期 extraBody 浅层合并到已构建的 provider body。
+* 无 extraBody 时原样返回；有则允许覆盖同名顶层键。
+*/
+function applyExtraBody(body, extraBody) {
+	if (!extraBody) return body;
+	return {
+		...body,
+		...extraBody
+	};
+}
+//#endregion
+//#region src/helpers/reasoning-level.ts
+/**
+* Portable reasoningLevel → provider wire 字段映射
+*
+* 第一版只处理 level 枚举；budget/summary 等特化字段不在此层。
+* 无法映射的 level 抛 AIRequestError(UNSUPPORTED_REASONING_LEVEL)。
+*/
+const REASONING_LEVELS = [
+	"none",
+	"minimal",
+	"low",
+	"medium",
+	"high",
+	"xhigh"
+];
+const REASONING_LEVEL_SET = new Set(REASONING_LEVELS);
+const MESSAGES_BUDGET_RATIOS = {
+	minimal: .02,
+	low: .1,
+	medium: .3,
+	high: .6,
+	xhigh: .9
+};
+const OLLAMA_SUPPORTED = /* @__PURE__ */ new Set([
+	"none",
+	"low",
+	"medium",
+	"high"
+]);
+/** 若 level 不在 supported 集合内则抛 AIRequestError。 */
+function assertSupportedReasoningLevel(level, supported, adapterKind) {
+	if (supported.has(level)) return;
+	throw new AIRequestError(`reasoningLevel "${level}" is not supported by the ${adapterKind} adapter`, "UNSUPPORTED_REASONING_LEVEL");
+}
+/** Responses API：`reasoning: { effort }` */
+function mapResponsesReasoning(level) {
+	return { effort: level };
+}
+/** Chat Completions：顶层 `reasoning_effort` */
+function mapChatCompletionsReasoningEffort(level) {
+	return level;
+}
+/**
+* Messages thinking budget。
+* 基于 maxTokens 按比例推导，clamp 到 [1024, max(1024, maxTokens - 1)]，
+* 满足 Anthropic budget_tokens < max_tokens。
+*/
+function mapMessagesThinkingBudget(level, maxTokens) {
+	const ratio = MESSAGES_BUDGET_RATIOS[level];
+	const raw = Math.round(maxTokens * ratio);
+	const upper = Math.max(1024, maxTokens - 1);
+	return Math.min(Math.max(raw, 1024), upper);
+}
+/** Messages API：`thinking` 字段 */
+function mapMessagesThinking(level, maxTokens) {
+	if (level === "none") return { type: "disabled" };
+	return {
+		type: "enabled",
+		budget_tokens: mapMessagesThinkingBudget(level, maxTokens)
+	};
+}
+/** Ollama：`think` 字段；minimal/xhigh 不支持 */
+function mapOllamaThink(level) {
+	assertSupportedReasoningLevel(level, OLLAMA_SUPPORTED, "ollama");
+	if (level === "none") return false;
+	return level;
+}
+//#endregion
 //#region src/helpers/request-mapper.ts
 var NormalizedRequestMapper = class {
 	kind;
@@ -1795,16 +1902,40 @@ function hasReplayCanonicalInput(input) {
 function extractFailureMessage(response) {
 	return response.error?.message ?? response.failure?.message ?? "unknown";
 }
-function canonicalToResponsesBlock(b) {
-	if (b.type === "text") return {
-		type: "text",
-		text: b.text
+function readNonEmptyString(value, maxLen = 256) {
+	if (typeof value !== "string" || value.length === 0 || value.length > maxLen) return void 0;
+	return value;
+}
+/** 将 canonical text/json blocks 压成 EasyInputMessage 的 string content。 */
+function messageContentAsString(blocks, field) {
+	return mapper$3.textFromBlocks(blocks, field);
+}
+function mapReasoningInput(item, index) {
+	const text = mapper$3.textFromBlocks(mapper$3.ensureReasoningBlocks(item.content, "reasoning content"), "reasoning content");
+	const id = item.id && item.id.length > 0 ? item.id : `reasoning_replay_${index}`;
+	if (item.visibility === "full") return {
+		type: "reasoning",
+		id,
+		summary: [],
+		content: text ? [{
+			type: "reasoning_text",
+			text
+		}] : void 0
 	};
-	if (b.type === "json") return {
-		type: "text",
-		text: JSON.stringify(b.json)
+	return {
+		type: "reasoning",
+		id,
+		summary: text ? [{
+			type: "summary_text",
+			text
+		}] : []
+	};
+}
+function extractOpaqueContinuationId(payload) {
+	return {
+		previousResponseId: readNonEmptyString(payload.previous_response_id) ?? (typeof payload.item_id === "string" ? void 0 : readNonEmptyString(payload.id)),
+		itemReferenceId: readNonEmptyString(payload.item_id)
 	};
-	throw new AIRequestError(`responses does not support content block type "${b.type}" in canonical mapping`, "UNSUPPORTED_CONTENT_BLOCK");
 }
 var ResponsesAdapter = class extends AdapterBase {
 	kind = "responses";
@@ -1812,44 +1943,35 @@ var ResponsesAdapter = class extends AdapterBase {
 	apiKey;
 	baseUrl;
 	fetchFn;
+	headers;
+	extraBody;
 	constructor(options) {
 		super();
 		this.apiKey = options.apiKey;
 		this.baseUrl = options.baseUrl ?? "https://api.openai.com/v1";
 		this.fetchFn = options.fetch ?? globalThis.fetch;
+		this.headers = options.headers;
+		this.extraBody = options.extraBody;
 	}
 	buildRequest(request) {
 		const input = [];
+		let previousResponseId;
+		let reasoningIndex = 0;
 		for (const item of request.input) switch (item.type) {
 			case "message":
-				if (item.role === "assistant") {
-					const blocks = mapper$3.ensureTextBlocks(item.content, `assistant message (${item.role}) content`).map(canonicalToResponsesBlock);
-					input.push({
-						type: "message",
-						role: item.role,
-						content: blocks
-					});
-				} else input.push({
+				input.push({
 					type: "message",
 					role: item.role,
-					content: mapper$3.textFromBlocks(item.content, `input message (${item.role}) content`)
+					content: messageContentAsString(item.content, `input message (${item.role}) content`)
 				});
 				break;
-			case "reasoning": {
-				const blocks = mapper$3.ensureReasoningBlocks(item.content, "reasoning content").map((b) => ({
-					type: "reasoning",
-					text: b.text
-				}));
-				input.push({
-					type: "reasoning",
-					content: blocks
-				});
+			case "reasoning":
+				input.push(mapReasoningInput(item, reasoningIndex++));
 				break;
-			}
 			case "tool_call":
 				input.push({
 					type: "function_call",
-					id: item.id,
+					call_id: item.id,
 					name: item.name,
 					arguments: item.argumentsText
 				});
@@ -1867,11 +1989,17 @@ var ResponsesAdapter = class extends AdapterBase {
 				if (item.source !== "responses" || item.purpose !== "replay") break;
 				assertOpaqueReplayEnvelope(item.payload);
 				const payload = item.payload;
-				if ("id" in payload) {
-					if (typeof payload.id !== "string" || payload.id.length === 0 || payload.id.length > 256) throw new AIRequestError("Invalid opaque replay payload: id must be a non-empty string (max 256)", "INVALID_OPAQUE_REPLAY");
-					if (!hasReplayCanonicalInput(input)) input.push({
+				for (const key of [
+					"id",
+					"previous_response_id",
+					"item_id"
+				]) if (key in payload && (typeof payload[key] !== "string" || payload[key].length === 0 || payload[key].length > 256)) throw new AIRequestError(`Invalid opaque replay payload: ${key} must be a non-empty string (max 256)`, "INVALID_OPAQUE_REPLAY");
+				const { previousResponseId: prevId, itemReferenceId } = extractOpaqueContinuationId(payload);
+				if (!hasReplayCanonicalInput(input)) {
+					if (prevId && !previousResponseId) previousResponseId = prevId;
+					else if (itemReferenceId) input.push({
 						type: "item_reference",
-						id: payload.id
+						id: itemReferenceId
 					});
 				}
 				break;
@@ -1882,6 +2010,7 @@ var ResponsesAdapter = class extends AdapterBase {
 			input,
 			stream: true
 		};
+		if (previousResponseId) body.previous_response_id = previousResponseId;
 		if (request.instructions) body.instructions = mapper$3.mapInstructions(request.instructions);
 		body.tools = mapper$3.mapToolsIfPresent(request.tools, (t) => ({
 			type: "function",
@@ -1900,7 +2029,8 @@ var ResponsesAdapter = class extends AdapterBase {
 		if (request.temperature !== void 0) body.temperature = request.temperature;
 		if (request.maxOutputTokens !== void 0) body.max_output_tokens = request.maxOutputTokens;
 		if (request.metadata) body.metadata = request.metadata;
-		return body;
+		if (request.reasoningLevel !== void 0) body.reasoning = mapResponsesReasoning(request.reasoningLevel);
+		return applyExtraBody(body, this.extraBody);
 	}
 	async *runStream(providerRequest, factory, request) {
 		const auxiliary = this.createAuxiliaryState(request);
@@ -1908,10 +2038,10 @@ var ResponsesAdapter = class extends AdapterBase {
 		const { reader } = await openProviderJsonStream({
 			fetchFn: this.fetchFn,
 			url: `${this.baseUrl}/responses`,
-			headers: {
+			headers: mergeProviderHeaders({
 				"Content-Type": "application/json",
 				Authorization: `Bearer ${this.apiKey}`
-			},
+			}, this.headers),
 			body: providerRequest,
 			signal: request.signal
 		});
@@ -1920,8 +2050,12 @@ var ResponsesAdapter = class extends AdapterBase {
 		let completedResponse;
 		let unknownEventsWarned = false;
 		const messageItemsWithDelta = /* @__PURE__ */ new Set();
+		/** item_id → function name */
 		const toolCallNames = /* @__PURE__ */ new Map();
+		/** item_id → call_id（canonical ToolCallItem.id / function_call_output.call_id） */
+		const toolCallIds = /* @__PURE__ */ new Map();
 		const reasoningStates = /* @__PURE__ */ new Map();
+		const resolveToolCallId = (itemId) => toolCallIds.get(itemId) ?? itemId;
 		const ensureReasoningState = (itemId, visibility = "summary") => {
 			let state = reasoningStates.get(itemId);
 			if (!state) {
@@ -1967,8 +2101,10 @@ var ResponsesAdapter = class extends AdapterBase {
 						}
 						case "function_call": {
 							const name = typeof item.name === "string" ? item.name : "unknown";
+							const callId = typeof item.call_id === "string" && item.call_id.length > 0 ? item.call_id : item.id;
 							toolCallNames.set(item.id, name);
-							yield factory.toolCallStarted(item.id, name);
+							toolCallIds.set(item.id, callId);
+							yield factory.toolCallStarted(callId, name);
 							break;
 						}
 					}
@@ -2057,13 +2193,15 @@ var ResponsesAdapter = class extends AdapterBase {
 				}
 				if (sseEvent.type === "response.function_call_arguments.delta") {
 					const data = sseEvent.data;
-					if (data.delta) yield factory.toolCallDelta(data.item_id, { argumentsText: data.delta });
+					if (data.delta) yield factory.toolCallDelta(resolveToolCallId(data.item_id), { argumentsText: data.delta });
 					continue;
 				}
 				if (sseEvent.type === "response.function_call_arguments.done") {
 					const data = sseEvent.data;
-					const tcItem = toolCallItem(data.item_id, toolCallNames.get(data.item_id) ?? "unknown", data.arguments);
-					yield factory.toolCallCompleted(data.item_id);
+					const callId = resolveToolCallId(data.item_id);
+					if (!toolCallIds.has(data.item_id)) toolCallIds.set(data.item_id, callId);
+					const tcItem = toolCallItem(callId, toolCallNames.get(data.item_id) ?? "unknown", data.arguments);
+					yield factory.toolCallCompleted(callId);
 					output.push(tcItem);
 					continue;
 				}
@@ -2097,7 +2235,10 @@ var ResponsesAdapter = class extends AdapterBase {
 			if (completedResponse.usage) auxiliary.recordUsage(usageFromOpenAIResponses(completedResponse.usage), "final", completedResponse.usage);
 		}
 		const replay = [...replayFromOutput(output)];
-		if (completedResponse?.id) replay.push(opaqueItem("responses", "replay", { id: completedResponse.id }));
+		if (completedResponse?.id) replay.push(opaqueItem("responses", "replay", {
+			id: completedResponse.id,
+			previous_response_id: completedResponse.id
+		}));
 		const stopReason = completedResponse ? this.inferStopReason(completedResponse) : void 0;
 		if (gate.tryComplete()) yield* this.emitStreamCompleted(factory, request, auxiliary, {
 			output,
@@ -2211,12 +2352,16 @@ var MessagesAdapter = class extends AdapterBase {
 	apiVersion;
 	baseUrl;
 	fetchFn;
+	headers;
+	extraBody;
 	constructor(options) {
 		super();
 		this.apiKey = options.apiKey;
 		this.apiVersion = options.apiVersion ?? "2023-06-01";
 		this.baseUrl = options.baseUrl ?? "https://api.anthropic.com/v1";
 		this.fetchFn = options.fetch ?? globalThis.fetch;
+		this.headers = options.headers;
+		this.extraBody = options.extraBody;
 	}
 	buildRequest(request) {
 		const messages = [];
@@ -2322,7 +2467,8 @@ var MessagesAdapter = class extends AdapterBase {
 			})
 		});
 		if (request.temperature !== void 0) body.temperature = request.temperature;
-		return body;
+		if (request.reasoningLevel !== void 0) body.thinking = mapMessagesThinking(request.reasoningLevel, body.max_tokens);
+		return applyExtraBody(body, this.extraBody);
 	}
 	async *runStream(providerRequest, factory, request) {
 		const auxiliary = this.createAuxiliaryState(request);
@@ -2331,11 +2477,11 @@ var MessagesAdapter = class extends AdapterBase {
 		const { reader, headers } = await openProviderJsonStream({
 			fetchFn: this.fetchFn,
 			url: `${this.baseUrl}/messages`,
-			headers: {
+			headers: mergeProviderHeaders({
 				"Content-Type": "application/json",
 				"x-api-key": this.apiKey,
 				"anthropic-version": this.apiVersion
-			},
+			}, this.headers),
 			body: providerRequest,
 			signal: request.signal
 		});
@@ -2610,11 +2756,15 @@ var ChatCompletionsAdapter = class extends AdapterBase {
 	apiKey;
 	baseUrl;
 	fetchFn;
+	headers;
+	extraBody;
 	constructor(options) {
 		super();
 		this.apiKey = options.apiKey;
 		this.baseUrl = options.baseUrl ?? "https://api.openai.com/v1";
 		this.fetchFn = options.fetch ?? globalThis.fetch;
+		this.headers = options.headers;
+		this.extraBody = options.extraBody;
 	}
 	buildRequest(request) {
 		const messages = [];
@@ -2708,7 +2858,8 @@ var ChatCompletionsAdapter = class extends AdapterBase {
 		if (request.temperature !== void 0) body.temperature = request.temperature;
 		if (request.maxOutputTokens !== void 0) body.max_tokens = request.maxOutputTokens;
 		if (request.metadata) body.metadata = request.metadata;
-		return body;
+		if (request.reasoningLevel !== void 0) body.reasoning_effort = mapChatCompletionsReasoningEffort(request.reasoningLevel);
+		return applyExtraBody(body, this.extraBody);
 	}
 	async *runStream(providerRequest, factory, request) {
 		const auxiliary = this.createAuxiliaryState(request);
@@ -2716,10 +2867,10 @@ var ChatCompletionsAdapter = class extends AdapterBase {
 		const { reader } = await openProviderJsonStream({
 			fetchFn: this.fetchFn,
 			url: `${this.baseUrl}/chat/completions`,
-			headers: {
+			headers: mergeProviderHeaders({
 				"Content-Type": "application/json",
 				Authorization: `Bearer ${this.apiKey}`
-			},
+			}, this.headers),
 			body: providerRequest,
 			signal: request.signal
 		});
@@ -2945,11 +3096,15 @@ var OllamaAdapter = class extends AdapterBase {
 	baseUrl;
 	apiKey;
 	fetchFn;
+	headers;
+	extraBody;
 	constructor(options = {}) {
 		super();
 		this.baseUrl = options.baseUrl ?? "http://localhost:11434";
 		this.apiKey = options.apiKey;
 		this.fetchFn = options.fetch ?? globalThis.fetch;
+		this.headers = options.headers;
+		this.extraBody = options.extraBody;
 	}
 	buildRequest(request) {
 		const messages = [];
@@ -3047,7 +3202,8 @@ var OllamaAdapter = class extends AdapterBase {
 			if (request.temperature !== void 0) body.options.temperature = request.temperature;
 			if (request.maxOutputTokens !== void 0) body.options.num_predict = request.maxOutputTokens;
 		}
-		return body;
+		if (request.reasoningLevel !== void 0) body.think = mapOllamaThink(request.reasoningLevel);
+		return applyExtraBody(body, this.extraBody);
 	}
 	async *runStream(providerRequest, factory, request) {
 		const auxiliary = this.createAuxiliaryState(request);
@@ -3059,7 +3215,7 @@ var OllamaAdapter = class extends AdapterBase {
 		const { reader } = await openProviderJsonStream({
 			fetchFn: this.fetchFn,
 			url: `${this.baseUrl}/api/chat`,
-			headers,
+			headers: mergeProviderHeaders(headers, this.headers),
 			body: providerRequest,
 			signal: request.signal
 		});
@@ -3228,7 +3384,7 @@ var MockAdapter = class extends AdapterBase {
 	}
 	async buildRequest(request) {
 		const turnIndex = this.cursor;
-		const context = this.buildHandlerContext(turnIndex, request.signal);
+		const context = this.buildHandlerContext(turnIndex, request);
 		const remainingPendingToolCalls = consumePendingToolCalls(this.pendingToolCalls, request.input);
 		const handlerResult = this.handler(request, context);
 		this.cursor += 1;
@@ -3367,7 +3523,7 @@ var MockAdapter = class extends AdapterBase {
 			rawResponseId: completion.rawResponseId
 		}, factory);
 	}
-	buildHandlerContext(turnIndex, signal) {
+	buildHandlerContext(turnIndex, request) {
 		return {
 			turnIndex,
 			previousReplay: this.previousReplay.map(cloneItem),
@@ -3377,7 +3533,8 @@ var MockAdapter = class extends AdapterBase {
 				replay: record.replay.map(cloneItem),
 				toolCalls: record.toolCalls.map(cloneItem)
 			})),
-			signal
+			signal: request.signal,
+			reasoningLevel: request.reasoningLevel
 		};
 	}
 };
@@ -3603,6 +3760,6 @@ function cloneItem(item) {
 	return structuredClone(item);
 }
 //#endregion
-export { AIError, AIMappingError, AIProviderError, AIRequestError, AIStreamError, AdapterAuxiliaryState, AdapterBase, AuxiliaryCollector, ChatCompletionsAdapter, IncrementalStreamParser, MAX_OPAQUE_JSON_DEPTH, MAX_OPAQUE_PAYLOAD_BYTES, MessagesAdapter, MockAdapter, NormalizedRequestMapper, OllamaAdapter, PROVIDER_ERROR_MESSAGE_MAX_LEN, PROVIDER_ERROR_RAW_BODY_THRESHOLD, ResponsesAdapter, WarningCode, aggregateEvents, assertMockRequest, assertOpaqueReplayEnvelope, assertValidRequest, blockToText, collectStream, contentBlocksToText, createAIClient, createChatCompletionsSseParser, createCompletionGate, createEventFactory, createNdjsonLineParser, createSseJsonParser, emitMalformedStreamWarning, extractProviderErrorMessage, extractText, imageBlock, iterateProviderStreamBatches, jsonBlock, mapReasoningVisibility, mapStopReason, measureJsonDepth, messageItem, normalizeRequest, opaqueBlock, opaqueItem, openProviderJsonStream, parseChatCompletionsDataLine, parseSseJsonFrame, providerHttpError, reasoningItem, replayFromOutput, splitLines, splitSSEFrames, syntheticStream, textBlock, toolCallItem, toolResultItem, usageFromAnthropicMessages, usageFromChatCompletions, usageFromOllama, usageFromOpenAIResponses, validateOpaqueReplayEnvelope, validateRequest, withMockStreaming };
+export { AIError, AIMappingError, AIProviderError, AIRequestError, AIStreamError, AdapterAuxiliaryState, AdapterBase, AuxiliaryCollector, ChatCompletionsAdapter, IncrementalStreamParser, MAX_OPAQUE_JSON_DEPTH, MAX_OPAQUE_PAYLOAD_BYTES, MessagesAdapter, MockAdapter, NormalizedRequestMapper, OllamaAdapter, PROVIDER_ERROR_MESSAGE_MAX_LEN, PROVIDER_ERROR_RAW_BODY_THRESHOLD, REASONING_LEVELS, REASONING_LEVEL_SET, ResponsesAdapter, WarningCode, aggregateEvents, applyExtraBody, assertMockRequest, assertOpaqueReplayEnvelope, assertSupportedReasoningLevel, assertValidRequest, blockToText, collectStream, contentBlocksToText, createAIClient, createChatCompletionsSseParser, createCompletionGate, createEventFactory, createNdjsonLineParser, createSseJsonParser, emitMalformedStreamWarning, extractProviderErrorMessage, extractText, imageBlock, iterateProviderStreamBatches, jsonBlock, mapChatCompletionsReasoningEffort, mapMessagesThinking, mapMessagesThinkingBudget, mapOllamaThink, mapReasoningVisibility, mapResponsesReasoning, mapStopReason, measureJsonDepth, mergeProviderHeaders, messageItem, normalizeRequest, opaqueBlock, opaqueItem, openProviderJsonStream, parseChatCompletionsDataLine, parseSseJsonFrame, providerHttpError, reasoningItem, replayFromOutput, splitLines, splitSSEFrames, syntheticStream, textBlock, toolCallItem, toolResultItem, usageFromAnthropicMessages, usageFromChatCompletions, usageFromOllama, usageFromOpenAIResponses, validateOpaqueReplayEnvelope, validateRequest, withMockStreaming };
 
 //# sourceMappingURL=index.mjs.map