@codehz/ai 0.1.5 → 0.1.7

package/dist/index.d.mts

@@ -94,13 +94,13 @@ type AIRequest = {
 //#region src/types/response.d.ts
 type StopReason = "end_turn" | "tool_call" | "max_output_tokens" | "content_filter" | "error" | "unknown";
 type Usage = {
-  inputTokens?: number;
-  outputTokens?: number;
+  /** Provider prompt / input token count */inputTokens?: number; /** Provider completion / output token count */
+  outputTokens?: number; /** Reasoning tokens when provider exposes output breakdown (e.g. OpenAI Responses) */
   reasoningTokens?: number;
-  totalTokens?: number;
-  cachedInputTokens?: number;
-  cacheWriteInputTokens?: number;
-  billableInputTokens?: number;
+  totalTokens?: number; /** Tokens read from prompt cache (OpenAI cached_tokens, Anthropic cache_read_input_tokens) */
+  cachedInputTokens?: number; /** Tokens written to prompt cache (Anthropic cache_creation_input_tokens) */
+  cacheWriteInputTokens?: number; /** Best-effort billable input (full-rate input; excludes discounted cache reads where known) */
+  billableInputTokens?: number; /** Best-effort billable output (non-reasoning slice when provider gives reasoning breakdown) */
   billableOutputTokens?: number;
 };
 type BillingInfo = {
@@ -1046,5 +1046,49 @@ type SyntheticStreamOptions = {
  */
 declare function syntheticStream(options: SyntheticStreamOptions): AsyncIterable<AIStreamEvent>;
 //#endregion
-export { type AIClient, AIError, AIMappingError, AIProviderError, type AIRequest, AIRequestError, type AIResponse, AIStreamError, type AIStreamEvent, AdapterAuxiliaryState, AdapterBase, AuxiliaryCollector, type AuxiliaryFinalizeOptions, type AuxiliaryFinalizeResult, type AuxiliaryInfo, type BackendAdapter, type BackendTrace, type BillingInfo, type BillingPostprocessHook, type BillingSource, ChatCompletionsAdapter, type ChatCompletionsAdapterOptions, type ContentBlock, type CreateAIClientOptions, type EventFactory, type EventFactoryBackend, type EventFactoryState, type FetchFn, type IncludeSettings, type InputItem, type InstructionBlock, type JsonContentBlock, type LookupResult, type MessageCompletedEvent, type MessageDeltaEvent, type MessageItem, type MessageStartedEvent, MessagesAdapter, type MessagesAdapterOptions, MockAdapter, type MockAdapterOptions, type MockAuxiliaryStep, type MockCompleteStep, type MockErrorStep, type MockHandler, type MockHandlerContext, type MockHistoryRecord, type MockInputExpectation, type MockInterruptStep, type MockMessageStep, type MockOutputStep, type MockReasoningStep, type MockRequestExpectation, type MockStaticHandler, type MockStep, type MockTextStreamOptions, type MockThrowStep, type MockToolCallStep, type MockWarningStep, type NormalizeOptions, type NormalizedRequest, OllamaAdapter, type OllamaAdapterOptions, type OpaqueItem, type OutputItem, type ReasoningCompletedEvent, type ReasoningDeltaEvent, type ReasoningItem, type ReasoningStartedEvent, type ReplayItem, type ResponseAuxiliaryEvent, type ResponseCompletedEvent, type ResponseStartedEvent, type ResponseWarningEvent, ResponsesAdapter, type ResponsesAdapterOptions, type SSEEvent, type StopReason, type StreamEventBase, type StreamResult, type SyntheticStreamOptions, type TextContentBlock, type ToolCallCompletedEvent, type ToolCallDeltaEvent, type ToolCallItem, type ToolCallStartedEvent, type ToolChoice, type ToolDefinition, type ToolResultItem, type Usage, type UsageSource, type ValidationIssue, WarningCode, aggregateEvents, assertMockRequest, assertValidRequest, blockToText, collectStream, contentBlocksToText, createAIClient, createEventFactory, emitMalformedStreamWarning, extractText, imageBlock, instructionsToText, jsonBlock, mapReasoningVisibility, mapStopReason, messageItem, metadataSourceList, normalizeRequest, opaqueBlock, opaqueItem, parseSSEEvents, reasoningItem, replayFromOutput, syntheticStream, textBlock, toolCallItem, toolResultItem, validateRequest, withMockStreaming };
+//#region src/helpers/usage-mapping.d.ts
+/** OpenAI Chat Completions `usage` */
+declare function usageFromChatCompletions(raw: {
+  prompt_tokens?: number;
+  completion_tokens?: number;
+  total_tokens?: number;
+  prompt_tokens_details?: {
+    cached_tokens?: number;
+    [key: string]: unknown;
+  };
+  completion_tokens_details?: {
+    reasoning_tokens?: number;
+    [key: string]: unknown;
+  };
+}): Partial<Usage>;
+/** OpenAI Responses API `usage` */
+declare function usageFromOpenAIResponses(raw: {
+  input_tokens?: number;
+  output_tokens?: number;
+  total_tokens?: number;
+  input_tokens_details?: {
+    cached_tokens?: number;
+    [key: string]: unknown;
+  };
+  output_tokens_details?: {
+    reasoning_tokens?: number;
+    [key: string]: unknown;
+  };
+  [key: string]: unknown;
+}): Partial<Usage>;
+/** Anthropic Messages `usage`（message_start / message_delta） */
+declare function usageFromAnthropicMessages(raw: {
+  input_tokens?: number;
+  output_tokens?: number;
+  cache_creation_input_tokens?: number;
+  cache_read_input_tokens?: number;
+  [key: string]: unknown;
+}): Partial<Usage>;
+/** Ollama 流式 chunk（无 cache / reasoning 细分时仅填基础与 billable 镜像） */
+declare function usageFromOllama(raw: {
+  prompt_eval_count?: number;
+  eval_count?: number;
+}): Partial<Usage>;
+//#endregion
+export { type AIClient, AIError, AIMappingError, AIProviderError, type AIRequest, AIRequestError, type AIResponse, AIStreamError, type AIStreamEvent, AdapterAuxiliaryState, AdapterBase, AuxiliaryCollector, type AuxiliaryFinalizeOptions, type AuxiliaryFinalizeResult, type AuxiliaryInfo, type BackendAdapter, type BackendTrace, type BillingInfo, type BillingPostprocessHook, type BillingSource, ChatCompletionsAdapter, type ChatCompletionsAdapterOptions, type ContentBlock, type CreateAIClientOptions, type EventFactory, type EventFactoryBackend, type EventFactoryState, type FetchFn, type IncludeSettings, type InputItem, type InstructionBlock, type JsonContentBlock, type LookupResult, type MessageCompletedEvent, type MessageDeltaEvent, type MessageItem, type MessageStartedEvent, MessagesAdapter, type MessagesAdapterOptions, MockAdapter, type MockAdapterOptions, type MockAuxiliaryStep, type MockCompleteStep, type MockErrorStep, type MockHandler, type MockHandlerContext, type MockHistoryRecord, type MockInputExpectation, type MockInterruptStep, type MockMessageStep, type MockOutputStep, type MockReasoningStep, type MockRequestExpectation, type MockStaticHandler, type MockStep, type MockTextStreamOptions, type MockThrowStep, type MockToolCallStep, type MockWarningStep, type NormalizeOptions, type NormalizedRequest, OllamaAdapter, type OllamaAdapterOptions, type OpaqueItem, type OutputItem, type ReasoningCompletedEvent, type ReasoningDeltaEvent, type ReasoningItem, type ReasoningStartedEvent, type ReplayItem, type ResponseAuxiliaryEvent, type ResponseCompletedEvent, type ResponseStartedEvent, type ResponseWarningEvent, ResponsesAdapter, type ResponsesAdapterOptions, type SSEEvent, type StopReason, type StreamEventBase, type StreamResult, type SyntheticStreamOptions, type TextContentBlock, type ToolCallCompletedEvent, type ToolCallDeltaEvent, type ToolCallItem, type ToolCallStartedEvent, type ToolChoice, type ToolDefinition, type ToolResultItem, type Usage, type UsageSource, type ValidationIssue, WarningCode, aggregateEvents, assertMockRequest, assertValidRequest, blockToText, collectStream, contentBlocksToText, createAIClient, createEventFactory, emitMalformedStreamWarning, extractText, imageBlock, instructionsToText, jsonBlock, mapReasoningVisibility, mapStopReason, messageItem, metadataSourceList, normalizeRequest, opaqueBlock, opaqueItem, parseSSEEvents, reasoningItem, replayFromOutput, syntheticStream, textBlock, toolCallItem, toolResultItem, usageFromAnthropicMessages, usageFromChatCompletions, usageFromOllama, usageFromOpenAIResponses, validateRequest, withMockStreaming };
 //# sourceMappingURL=index.d.mts.map

package/dist/index.mjs

@@ -1011,6 +1011,93 @@ function mergeWarnings(...groups) {
 	return merged.length > 0 ? merged : void 0;
 }
 //#endregion
+//#region src/helpers/usage-mapping.ts
+function num(value) {
+	return typeof value === "number" && Number.isFinite(value) ? value : void 0;
+}
+function record(obj) {
+	const out = {};
+	for (const [key, value] of Object.entries(obj)) if (value !== void 0) out[key] = value;
+	return out;
+}
+function billableFromOpenAIStyle(inputTokens, outputTokens, cachedInputTokens, reasoningTokens) {
+	let billableInputTokens;
+	if (inputTokens !== void 0) billableInputTokens = cachedInputTokens !== void 0 ? Math.max(0, inputTokens - cachedInputTokens) : inputTokens;
+	let billableOutputTokens;
+	if (outputTokens !== void 0) billableOutputTokens = reasoningTokens !== void 0 ? Math.max(0, outputTokens - reasoningTokens) : outputTokens;
+	return record({
+		billableInputTokens,
+		billableOutputTokens
+	});
+}
+/** OpenAI Chat Completions `usage` */
+function usageFromChatCompletions(raw) {
+	const inputTokens = num(raw.prompt_tokens);
+	const outputTokens = num(raw.completion_tokens);
+	const cachedInputTokens = num(raw.prompt_tokens_details?.cached_tokens);
+	const reasoningTokens = num(raw.completion_tokens_details?.reasoning_tokens);
+	return record({
+		inputTokens,
+		outputTokens,
+		totalTokens: num(raw.total_tokens) ?? (inputTokens !== void 0 && outputTokens !== void 0 ? inputTokens + outputTokens : void 0),
+		cachedInputTokens,
+		reasoningTokens,
+		...billableFromOpenAIStyle(inputTokens, outputTokens, cachedInputTokens, reasoningTokens)
+	});
+}
+/** OpenAI Responses API `usage` */
+function usageFromOpenAIResponses(raw) {
+	const inputTokens = num(raw.input_tokens);
+	const outputTokens = num(raw.output_tokens);
+	const cachedInputTokens = num(raw.input_tokens_details?.cached_tokens);
+	const reasoningTokens = num(raw.output_tokens_details?.reasoning_tokens);
+	return record({
+		inputTokens,
+		outputTokens,
+		totalTokens: num(raw.total_tokens) ?? (inputTokens !== void 0 && outputTokens !== void 0 ? inputTokens + outputTokens : void 0),
+		cachedInputTokens,
+		reasoningTokens,
+		...billableFromOpenAIStyle(inputTokens, outputTokens, cachedInputTokens, reasoningTokens)
+	});
+}
+/** Anthropic Messages `usage`（message_start / message_delta） */
+function usageFromAnthropicMessages(raw) {
+	const inputTokens = num(raw.input_tokens);
+	const outputTokens = num(raw.output_tokens);
+	const cacheWriteInputTokens = num(raw.cache_creation_input_tokens);
+	const cachedInputTokens = num(raw.cache_read_input_tokens);
+	const inputParts = [
+		inputTokens,
+		cacheWriteInputTokens,
+		cachedInputTokens
+	].filter((n) => n !== void 0);
+	const summedInput = inputParts.length > 0 ? inputParts.reduce((sum, n) => sum + n, 0) : void 0;
+	const totalTokens = summedInput !== void 0 && outputTokens !== void 0 ? summedInput + outputTokens : void 0;
+	let billableInputTokens;
+	if (inputTokens !== void 0 || cacheWriteInputTokens !== void 0) billableInputTokens = (inputTokens ?? 0) + (cacheWriteInputTokens ?? 0);
+	return record({
+		inputTokens,
+		outputTokens,
+		totalTokens,
+		cachedInputTokens,
+		cacheWriteInputTokens,
+		billableInputTokens,
+		billableOutputTokens: outputTokens
+	});
+}
+/** Ollama 流式 chunk（无 cache / reasoning 细分时仅填基础与 billable 镜像） */
+function usageFromOllama(raw) {
+	const inputTokens = num(raw.prompt_eval_count);
+	const outputTokens = num(raw.eval_count);
+	return record({
+		inputTokens,
+		outputTokens,
+		totalTokens: inputTokens !== void 0 && outputTokens !== void 0 ? inputTokens + outputTokens : void 0,
+		billableInputTokens: inputTokens,
+		billableOutputTokens: outputTokens
+	});
+}
+//#endregion
 //#region src/helpers/sse-parser.ts
 /**
 * 将 SSE 文本块解析为事件数组。
@@ -1313,11 +1400,7 @@ var ResponsesAdapter = class extends AdapterBase {
 		let rawResponseId;
 		if (completedResponse) {
 			rawResponseId = completedResponse.id;
-			if (completedResponse.usage) auxiliary.recordUsage({
-				inputTokens: completedResponse.usage.input_tokens,
-				outputTokens: completedResponse.usage.output_tokens,
-				totalTokens: completedResponse.usage.total_tokens
-			}, "final", completedResponse.usage);
+			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 }));
@@ -1456,78 +1539,86 @@ var MessagesAdapter = class extends AdapterBase {
 	buildRequest(request) {
 		const messages = [];
 		let systemPrompt;
+		let pendingToolResultMessage;
 		if (request.instructions) systemPrompt = instructionsToMessagesText(request.instructions);
-		for (const item of request.input) switch (item.type) {
-			case "message": {
-				const role = item.role === "user" ? "user" : "assistant";
-				const supportedContent = ensureMessagesTextBlocks(item.content, `input message (${item.role}) content`);
-				if (supportedContent.length === 1 && supportedContent[0]?.type === "text") messages.push({
-					role,
-					content: supportedContent[0].text
-				});
-				else messages.push({
-					role,
-					content: supportedContent.map(canonicalToMessagesBlock)
-				});
-				break;
-			}
-			case "tool_call": {
-				const lastMsg = messages[messages.length - 1];
-				const toolBlock = {
-					type: "tool_use",
-					id: item.id,
-					name: item.name,
-					input: item.argumentsJson ?? parseToolUseInput(item.argumentsText)
-				};
-				if (lastMsg && lastMsg.role === "assistant" && typeof lastMsg.content !== "string") lastMsg.content.push(toolBlock);
-				else messages.push({
-					role: "assistant",
-					content: [toolBlock]
-				});
-				break;
-			}
-			case "tool_result": {
-				assertMessagesToolResultOutcome(item.outcome);
-				const content = ensureMessagesTextBlocks(item.content, `tool_result ${item.callId} content`).map(blockToText).join("\n");
-				const block = {
-					type: "tool_result",
-					tool_use_id: item.callId,
-					content,
-					is_error: item.outcome === "error"
-				};
-				messages.push({
-					role: "user",
-					content: [block]
-				});
-				break;
-			}
-			case "reasoning": {
-				const block = {
-					type: "thinking",
-					thinking: contentBlocksToText(ensureMessagesReasoningBlocks(item.content, "reasoning content"))
-				};
-				const lastMsg = messages[messages.length - 1];
-				if (lastMsg && lastMsg.role === "assistant" && typeof lastMsg.content !== "string") lastMsg.content.push(block);
-				else messages.push({
-					role: "assistant",
-					content: [block]
-				});
-				break;
-			}
-			case "opaque":
-				if (item.purpose === "replay" && typeof item.payload === "object" && item.payload !== null) {
-					const payload = item.payload;
-					if (payload.role === "assistant" && Array.isArray(payload.content)) {
-						if (payload.content.every((b) => typeof b === "object" && b !== null && "type" in b && (b.type === "text" || b.type === "thinking" || b.type === "redacted_thinking" || b.type === "tool_use" || b.type === "tool_result"))) {
-							rollbackTrailingAssistantMessages$2(messages);
-							messages.push({
-								role: "assistant",
-								content: payload.content
-							});
-						}
+		for (const item of request.input) {
+			if (item.type !== "tool_result") pendingToolResultMessage = void 0;
+			switch (item.type) {
+				case "message": {
+					const role = item.role === "user" ? "user" : "assistant";
+					const supportedContent = ensureMessagesTextBlocks(item.content, `input message (${item.role}) content`);
+					if (supportedContent.length === 1 && supportedContent[0]?.type === "text") messages.push({
+						role,
+						content: supportedContent[0].text
+					});
+					else messages.push({
+						role,
+						content: supportedContent.map(canonicalToMessagesBlock)
+					});
+					break;
+				}
+				case "tool_call": {
+					const lastMsg = messages[messages.length - 1];
+					const toolBlock = {
+						type: "tool_use",
+						id: item.id,
+						name: item.name,
+						input: item.argumentsJson ?? parseToolUseInput(item.argumentsText)
+					};
+					if (lastMsg && lastMsg.role === "assistant" && typeof lastMsg.content !== "string") lastMsg.content.push(toolBlock);
+					else messages.push({
+						role: "assistant",
+						content: [toolBlock]
+					});
+					break;
+				}
+				case "tool_result": {
+					assertMessagesToolResultOutcome(item.outcome);
+					const content = ensureMessagesTextBlocks(item.content, `tool_result ${item.callId} content`).map(blockToText).join("\n");
+					const block = {
+						type: "tool_result",
+						tool_use_id: item.callId,
+						content,
+						is_error: item.outcome === "error"
+					};
+					if (pendingToolResultMessage && typeof pendingToolResultMessage.content !== "string") pendingToolResultMessage.content.push(block);
+					else {
+						pendingToolResultMessage = {
+							role: "user",
+							content: [block]
+						};
+						messages.push(pendingToolResultMessage);
 					}
+					break;
 				}
-				break;
+				case "reasoning": {
+					const block = {
+						type: "thinking",
+						thinking: contentBlocksToText(ensureMessagesReasoningBlocks(item.content, "reasoning content"))
+					};
+					const lastMsg = messages[messages.length - 1];
+					if (lastMsg && lastMsg.role === "assistant" && typeof lastMsg.content !== "string") lastMsg.content.push(block);
+					else messages.push({
+						role: "assistant",
+						content: [block]
+					});
+					break;
+				}
+				case "opaque":
+					if (item.purpose === "replay" && typeof item.payload === "object" && item.payload !== null) {
+						const payload = item.payload;
+						if (payload.role === "assistant" && Array.isArray(payload.content)) {
+							if (payload.content.every((b) => typeof b === "object" && b !== null && "type" in b && (b.type === "text" || b.type === "thinking" || b.type === "redacted_thinking" || b.type === "tool_use" || b.type === "tool_result"))) {
+								rollbackTrailingAssistantMessages$2(messages);
+								messages.push({
+									role: "assistant",
+									content: payload.content
+								});
+							}
+						}
+					}
+					break;
+			}
 		}
 		const body = {
 			model: request.model,
@@ -1728,11 +1819,7 @@ var MessagesAdapter = class extends AdapterBase {
 						stopReason = sseEvent.data.delta.stop_reason;
 						stopSequence = sseEvent.data.delta.stop_sequence;
 						const u = sseEvent.data.usage;
-						if (u) auxiliary.recordUsage({
-							inputTokens: u.input_tokens,
-							outputTokens: u.output_tokens,
-							totalTokens: u.input_tokens + u.output_tokens
-						}, "stream", u);
+						if (u) auxiliary.recordUsage(usageFromAnthropicMessages(u), "stream", u);
 						continue;
 					}
 					case "message_stop": break;
@@ -2067,11 +2154,7 @@ var ChatCompletionsAdapter = class extends AdapterBase {
 				if (malformedWarning) yield malformedWarning;
 				for (const chunk of chunks) {
 					responseId = chunk.id;
-					if (chunk.usage) auxiliary.recordUsage({
-						inputTokens: chunk.usage.prompt_tokens,
-						outputTokens: chunk.usage.completion_tokens,
-						totalTokens: chunk.usage.total_tokens
-					}, "final", chunk.usage);
+					if (chunk.usage) auxiliary.recordUsage(usageFromChatCompletions(chunk.usage), "final", chunk.usage);
 					for (const choice of chunk.choices) {
 						if (choice.index !== 0) continue;
 						const delta = choice.delta;
@@ -2441,11 +2524,10 @@ var OllamaAdapter = class extends AdapterBase {
 							yield factory.toolCallCompleted(toolCall);
 							output.push(toolCall);
 						}
-						if (request.include?.usage !== "off" && (chunk.prompt_eval_count !== void 0 || chunk.eval_count !== void 0)) auxiliary.recordUsage({
-							inputTokens: chunk.prompt_eval_count,
-							outputTokens: chunk.eval_count,
-							totalTokens: chunk.prompt_eval_count !== void 0 && chunk.eval_count !== void 0 ? chunk.prompt_eval_count + chunk.eval_count : void 0
-						}, "final", {
+						if (request.include?.usage !== "off" && (chunk.prompt_eval_count !== void 0 || chunk.eval_count !== void 0)) auxiliary.recordUsage(usageFromOllama({
+							prompt_eval_count: chunk.prompt_eval_count,
+							eval_count: chunk.eval_count
+						}), "final", {
 							prompt_eval_count: chunk.prompt_eval_count,
 							eval_count: chunk.eval_count
 						});
@@ -3010,6 +3092,6 @@ function* emitToolCallEvents(item, factory) {
 	yield factory.toolCallCompleted(item);
 }
 //#endregion
-export { AIError, AIMappingError, AIProviderError, AIRequestError, AIStreamError, AdapterAuxiliaryState, AdapterBase, AuxiliaryCollector, ChatCompletionsAdapter, MessagesAdapter, MockAdapter, OllamaAdapter, ResponsesAdapter, WarningCode, aggregateEvents, assertMockRequest, assertValidRequest, blockToText, collectStream, contentBlocksToText, createAIClient, createEventFactory, emitMalformedStreamWarning, extractText, imageBlock, instructionsToText, jsonBlock, mapReasoningVisibility, mapStopReason, messageItem, metadataSourceList, normalizeRequest, opaqueBlock, opaqueItem, parseSSEEvents, reasoningItem, replayFromOutput, syntheticStream, textBlock, toolCallItem, toolResultItem, validateRequest, withMockStreaming };
+export { AIError, AIMappingError, AIProviderError, AIRequestError, AIStreamError, AdapterAuxiliaryState, AdapterBase, AuxiliaryCollector, ChatCompletionsAdapter, MessagesAdapter, MockAdapter, OllamaAdapter, ResponsesAdapter, WarningCode, aggregateEvents, assertMockRequest, assertValidRequest, blockToText, collectStream, contentBlocksToText, createAIClient, createEventFactory, emitMalformedStreamWarning, extractText, imageBlock, instructionsToText, jsonBlock, mapReasoningVisibility, mapStopReason, messageItem, metadataSourceList, normalizeRequest, opaqueBlock, opaqueItem, parseSSEEvents, reasoningItem, replayFromOutput, syntheticStream, textBlock, toolCallItem, toolResultItem, usageFromAnthropicMessages, usageFromChatCompletions, usageFromOllama, usageFromOpenAIResponses, validateRequest, withMockStreaming };
 
 //# sourceMappingURL=index.mjs.map