pi_harness/packages/ai/src/providers/anthropic.ts

import Anthropic from "@anthropic-ai/sdk";
import type {
	CacheControlEphemeral,
	ContentBlockParam,
	MessageCreateParamsStreaming,
	MessageParam,
} from "@anthropic-ai/sdk/resources/messages.js";
import { getEnvApiKey } from "../env-api-keys.js";
import { calculateCost } from "../models.js";
import type {
	Api,
	AssistantMessage,
	CacheRetention,
	Context,
	ImageContent,
	Message,
	Model,
	SimpleStreamOptions,
	StopReason,
	StreamFunction,
	StreamOptions,
	TextContent,
	ThinkingContent,
	Tool,
	ToolCall,
	ToolResultMessage,
} from "../types.js";
import { AssistantMessageEventStream } from "../utils/event-stream.js";
import { headersToRecord } from "../utils/headers.js";
import { parseStreamingJson } from "../utils/json-parse.js";
import { sanitizeSurrogates } from "../utils/sanitize-unicode.js";

import { buildCopilotDynamicHeaders, hasCopilotVisionInput } from "./github-copilot-headers.js";
import { adjustMaxTokensForThinking, buildBaseOptions } from "./simple-options.js";
import { transformMessages } from "./transform-messages.js";

/**
 * Resolve cache retention preference.
 * Defaults to "short" and uses PI_CACHE_RETENTION for backward compatibility.
 */
function resolveCacheRetention(cacheRetention?: CacheRetention): CacheRetention {
	if (cacheRetention) {
		return cacheRetention;
	}
	if (typeof process !== "undefined" && process.env.PI_CACHE_RETENTION === "long") {
		return "long";
	}
	return "short";
}

function getCacheControl(
	baseUrl: string,
	cacheRetention?: CacheRetention,
): { retention: CacheRetention; cacheControl?: CacheControlEphemeral } {
	const retention = resolveCacheRetention(cacheRetention);
	if (retention === "none") {
		return { retention };
	}
	const ttl = retention === "long" && baseUrl.includes("api.anthropic.com") ? "1h" : undefined;
	return {
		retention,
		cacheControl: { type: "ephemeral", ...(ttl && { ttl }) },
	};
}

// Stealth mode: Mimic Claude Code's tool naming exactly
const claudeCodeVersion = "2.1.75";

// Claude Code 2.x tool names (canonical casing)
// Source: https://cchistory.mariozechner.at/data/prompts-2.1.11.md
// To update: https://github.com/badlogic/cchistory
const claudeCodeTools = [
	"Read",
	"Write",
	"Edit",
	"Bash",
	"Grep",
	"Glob",
	"AskUserQuestion",
	"EnterPlanMode",
	"ExitPlanMode",
	"KillShell",
	"NotebookEdit",
	"Skill",
	"Task",
	"TaskOutput",
	"TodoWrite",
	"WebFetch",
	"WebSearch",
];

const ccToolLookup = new Map(claudeCodeTools.map((t) => [t.toLowerCase(), t]));

// Convert tool name to CC canonical casing if it matches (case-insensitive)
const toClaudeCodeName = (name: string) => ccToolLookup.get(name.toLowerCase()) ?? name;
const fromClaudeCodeName = (name: string, tools?: Tool[]) => {
	if (tools && tools.length > 0) {
		const lowerName = name.toLowerCase();
		const matchedTool = tools.find((tool) => tool.name.toLowerCase() === lowerName);
		if (matchedTool) return matchedTool.name;
	}
	return name;
};

/**
 * Convert content blocks to Anthropic API format
 */
function convertContentBlocks(content: (TextContent | ImageContent)[]):
	| string
	| Array<
			| { type: "text"; text: string }
			| {
					type: "image";
					source: {
						type: "base64";
						media_type: "image/jpeg" | "image/png" | "image/gif" | "image/webp";
						data: string;
					};
			  }
	  > {
	// If only text blocks, return as concatenated string for simplicity
	const hasImages = content.some((c) => c.type === "image");
	if (!hasImages) {
		return sanitizeSurrogates(content.map((c) => (c as TextContent).text).join("\n"));
	}

	// If we have images, convert to content block array
	const blocks = content.map((block) => {
		if (block.type === "text") {
			return {
				type: "text" as const,
				text: sanitizeSurrogates(block.text),
			};
		}
		return {
			type: "image" as const,
			source: {
				type: "base64" as const,
				media_type: block.mimeType as "image/jpeg" | "image/png" | "image/gif" | "image/webp",
				data: block.data,
			},
		};
	});

	// If only images (no text), add placeholder text block
	const hasText = blocks.some((b) => b.type === "text");
	if (!hasText) {
		blocks.unshift({
			type: "text" as const,
			text: "(see attached image)",
		});
	}

	return blocks;
}

export type AnthropicEffort = "low" | "medium" | "high" | "xhigh" | "max";

export type AnthropicThinkingDisplay = "summarized" | "omitted";

export interface AnthropicOptions extends StreamOptions {
	/**
	 * Enable extended thinking.
	 * For Opus 4.6 and Sonnet 4.6: uses adaptive thinking (model decides when/how much to think).
	 * For older models: uses budget-based thinking with thinkingBudgetTokens.
	 */
	thinkingEnabled?: boolean;
	/**
	 * Token budget for extended thinking (older models only).
	 * Ignored for Opus 4.6 and Sonnet 4.6, which use adaptive thinking.
	 */
	thinkingBudgetTokens?: number;
	/**
	 * Effort level for adaptive thinking (Opus 4.6+ and Sonnet 4.6).
	 * Controls how much thinking Claude allocates:
	 * - "max": Always thinks with no constraints (Opus 4.6 only)
	 * - "xhigh": Highest reasoning level (Opus 4.7)
	 * - "high": Always thinks, deep reasoning (default)
	 * - "medium": Moderate thinking, may skip for simple queries
	 * - "low": Minimal thinking, skips for simple tasks
	 * Ignored for older models.
	 */
	effort?: AnthropicEffort;
	/**
	 * Controls how thinking content is returned in API responses.
	 * - "summarized": Thinking blocks contain summarized thinking text (default here).
	 * - "omitted": Thinking blocks return an empty thinking field; the encrypted
	 *   signature still travels back for multi-turn continuity. Use for faster
	 *   time-to-first-text-token when your UI does not surface thinking.
	 *
	 * Note: Anthropic's API default for Claude Opus 4.7 and Claude Mythos Preview
	 * is "omitted". We default to "summarized" here to keep behavior consistent
	 * with older Claude 4 models. Set this explicitly to "omitted" to opt in.
	 */
	thinkingDisplay?: AnthropicThinkingDisplay;
	interleavedThinking?: boolean;
	toolChoice?: "auto" | "any" | "none" | { type: "tool"; name: string };
	/**
	 * Pre-built Anthropic client instance. When provided, skips internal client
	 * construction entirely. Use this to inject alternative SDK clients such as
	 * `AnthropicVertex` that shares the same messaging API.
	 */
	client?: Anthropic;
}

function mergeHeaders(...headerSources: (Record<string, string> | undefined)[]): Record<string, string> {
	const merged: Record<string, string> = {};
	for (const headers of headerSources) {
		if (headers) {
			Object.assign(merged, headers);
		}
	}
	return merged;
}

export const streamAnthropic: StreamFunction<"anthropic-messages", AnthropicOptions> = (
	model: Model<"anthropic-messages">,
	context: Context,
	options?: AnthropicOptions,
): AssistantMessageEventStream => {
	const stream = new AssistantMessageEventStream();

	(async () => {
		const output: AssistantMessage = {
			role: "assistant",
			content: [],
			api: model.api as Api,
			provider: model.provider,
			model: model.id,
			usage: {
				input: 0,
				output: 0,
				cacheRead: 0,
				cacheWrite: 0,
				totalTokens: 0,
				cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
			},
			stopReason: "stop",
			timestamp: Date.now(),
		};

		try {
			let client: Anthropic;
			let isOAuth: boolean;

			if (options?.client) {
				client = options.client;
				isOAuth = false;
			} else {
				const apiKey = options?.apiKey ?? getEnvApiKey(model.provider) ?? "";

				let copilotDynamicHeaders: Record<string, string> | undefined;
				if (model.provider === "github-copilot") {
					const hasImages = hasCopilotVisionInput(context.messages);
					copilotDynamicHeaders = buildCopilotDynamicHeaders({
						messages: context.messages,
						hasImages,
					});
				}

				const created = createClient(
					model,
					apiKey,
					options?.interleavedThinking ?? true,
					options?.headers,
					copilotDynamicHeaders,
				);
				client = created.client;
				isOAuth = created.isOAuthToken;
			}
			let params = buildParams(model, context, isOAuth, options);
			const nextParams = await options?.onPayload?.(params, model);
			if (nextParams !== undefined) {
				params = nextParams as MessageCreateParamsStreaming;
			}
			const { data: anthropicStream, response } = await client.messages
				.stream({ ...params, stream: true }, { signal: options?.signal })
				.withResponse();
			await options?.onResponse?.({ status: response.status, headers: headersToRecord(response.headers) }, model);
			stream.push({ type: "start", partial: output });

			type Block = (ThinkingContent | TextContent | (ToolCall & { partialJson: string })) & { index: number };
			const blocks = output.content as Block[];

			for await (const event of anthropicStream) {
				if (event.type === "message_start") {
					output.responseId = event.message.id;
					// Capture initial token usage from message_start event
					// This ensures we have input token counts even if the stream is aborted early
					output.usage.input = event.message.usage.input_tokens || 0;
					output.usage.output = event.message.usage.output_tokens || 0;
					output.usage.cacheRead = event.message.usage.cache_read_input_tokens || 0;
					output.usage.cacheWrite = event.message.usage.cache_creation_input_tokens || 0;
					// Anthropic doesn't provide total_tokens, compute from components
					output.usage.totalTokens =
						output.usage.input + output.usage.output + output.usage.cacheRead + output.usage.cacheWrite;
					calculateCost(model, output.usage);
				} else if (event.type === "content_block_start") {
					if (event.content_block.type === "text") {
						const block: Block = {
							type: "text",
							text: "",
							index: event.index,
						};
						output.content.push(block);
						stream.push({ type: "text_start", contentIndex: output.content.length - 1, partial: output });
					} else if (event.content_block.type === "thinking") {
						const block: Block = {
							type: "thinking",
							thinking: "",
							thinkingSignature: "",
							index: event.index,
						};
						output.content.push(block);
						stream.push({ type: "thinking_start", contentIndex: output.content.length - 1, partial: output });
					} else if (event.content_block.type === "redacted_thinking") {
						const block: Block = {
							type: "thinking",
							thinking: "[Reasoning redacted]",
							thinkingSignature: event.content_block.data,
							redacted: true,
							index: event.index,
						};
						output.content.push(block);
						stream.push({ type: "thinking_start", contentIndex: output.content.length - 1, partial: output });
					} else if (event.content_block.type === "tool_use") {
						const block: Block = {
							type: "toolCall",
							id: event.content_block.id,
							name: isOAuth
								? fromClaudeCodeName(event.content_block.name, context.tools)
								: event.content_block.name,
							arguments: (event.content_block.input as Record<string, any>) ?? {},
							partialJson: "",
							index: event.index,
						};
						output.content.push(block);
						stream.push({ type: "toolcall_start", contentIndex: output.content.length - 1, partial: output });
					}
				} else if (event.type === "content_block_delta") {
					if (event.delta.type === "text_delta") {
						const index = blocks.findIndex((b) => b.index === event.index);
						const block = blocks[index];
						if (block && block.type === "text") {
							block.text += event.delta.text;
							stream.push({
								type: "text_delta",
								contentIndex: index,
								delta: event.delta.text,
								partial: output,
							});
						}
					} else if (event.delta.type === "thinking_delta") {
						const index = blocks.findIndex((b) => b.index === event.index);
						const block = blocks[index];
						if (block && block.type === "thinking") {
							block.thinking += event.delta.thinking;
							stream.push({
								type: "thinking_delta",
								contentIndex: index,
								delta: event.delta.thinking,
								partial: output,
							});
						}
					} else if (event.delta.type === "input_json_delta") {
						const index = blocks.findIndex((b) => b.index === event.index);
						const block = blocks[index];
						if (block && block.type === "toolCall") {
							block.partialJson += event.delta.partial_json;
							block.arguments = parseStreamingJson(block.partialJson);
							stream.push({
								type: "toolcall_delta",
								contentIndex: index,
								delta: event.delta.partial_json,
								partial: output,
							});
						}
					} else if (event.delta.type === "signature_delta") {
						const index = blocks.findIndex((b) => b.index === event.index);
						const block = blocks[index];
						if (block && block.type === "thinking") {
							block.thinkingSignature = block.thinkingSignature || "";
							block.thinkingSignature += event.delta.signature;
						}
					}
				} else if (event.type === "content_block_stop") {
					const index = blocks.findIndex((b) => b.index === event.index);
					const block = blocks[index];
					if (block) {
						delete (block as any).index;
						if (block.type === "text") {
							stream.push({
								type: "text_end",
								contentIndex: index,
								content: block.text,
								partial: output,
							});
						} else if (block.type === "thinking") {
							stream.push({
								type: "thinking_end",
								contentIndex: index,
								content: block.thinking,
								partial: output,
							});
						} else if (block.type === "toolCall") {
							block.arguments = parseStreamingJson(block.partialJson);
							// Finalize in-place and strip the scratch buffer so replay only
							// carries parsed arguments.
							delete (block as { partialJson?: string }).partialJson;
							stream.push({
								type: "toolcall_end",
								contentIndex: index,
								toolCall: block,
								partial: output,
							});
						}
					}
				} else if (event.type === "message_delta") {
					if (event.delta.stop_reason) {
						output.stopReason = mapStopReason(event.delta.stop_reason);
					}
					// Only update usage fields if present (not null).
					// Preserves input_tokens from message_start when proxies omit it in message_delta.
					if (event.usage.input_tokens != null) {
						output.usage.input = event.usage.input_tokens;
					}
					if (event.usage.output_tokens != null) {
						output.usage.output = event.usage.output_tokens;
					}
					if (event.usage.cache_read_input_tokens != null) {
						output.usage.cacheRead = event.usage.cache_read_input_tokens;
					}
					if (event.usage.cache_creation_input_tokens != null) {
						output.usage.cacheWrite = event.usage.cache_creation_input_tokens;
					}
					// Anthropic doesn't provide total_tokens, compute from components
					output.usage.totalTokens =
						output.usage.input + output.usage.output + output.usage.cacheRead + output.usage.cacheWrite;
					calculateCost(model, output.usage);
				}
			}

			if (options?.signal?.aborted) {
				throw new Error("Request was aborted");
			}

			if (output.stopReason === "aborted" || output.stopReason === "error") {
				throw new Error("An unknown error occurred");
			}

			stream.push({ type: "done", reason: output.stopReason, message: output });
			stream.end();
		} catch (error) {
			for (const block of output.content) {
				delete (block as { index?: number }).index;
				// partialJson is only a streaming scratch buffer; never persist it.
				delete (block as { partialJson?: string }).partialJson;
			}
			output.stopReason = options?.signal?.aborted ? "aborted" : "error";
			output.errorMessage = error instanceof Error ? error.message : JSON.stringify(error);
			stream.push({ type: "error", reason: output.stopReason, error: output });
			stream.end();
		}
	})();

	return stream;
};

/**
 * Check if a model supports adaptive thinking (Opus 4.6+, Sonnet 4.6)
 */
function supportsAdaptiveThinking(modelId: string): boolean {
	// Adaptive-thinking model IDs (with or without date suffix)
	return (
		modelId.includes("opus-4-6") ||
		modelId.includes("opus-4.6") ||
		modelId.includes("opus-4-7") ||
		modelId.includes("opus-4.7") ||
		modelId.includes("sonnet-4-6") ||
		modelId.includes("sonnet-4.6")
	);
}

/**
 * Map ThinkingLevel to Anthropic effort levels for adaptive thinking.
 * Note: effort "max" is only valid on Opus 4.6, while Opus 4.7 supports "xhigh".
 */
function mapThinkingLevelToEffort(level: SimpleStreamOptions["reasoning"], modelId: string): AnthropicEffort {
	switch (level) {
		case "minimal":
			return "low";
		case "low":
			return "low";
		case "medium":
			return "medium";
		case "high":
			return "high";
		case "xhigh":
			if (modelId.includes("opus-4-6") || modelId.includes("opus-4.6")) {
				return "max";
			}
			if (modelId.includes("opus-4-7") || modelId.includes("opus-4.7")) {
				return "xhigh";
			}
			return "high";
		default:
			return "high";
	}
}

export const streamSimpleAnthropic: StreamFunction<"anthropic-messages", SimpleStreamOptions> = (
	model: Model<"anthropic-messages">,
	context: Context,
	options?: SimpleStreamOptions,
): AssistantMessageEventStream => {
	const apiKey = options?.apiKey || getEnvApiKey(model.provider);
	if (!apiKey) {
		throw new Error(`No API key for provider: ${model.provider}`);
	}

	const base = buildBaseOptions(model, options, apiKey);
	if (!options?.reasoning) {
		return streamAnthropic(model, context, { ...base, thinkingEnabled: false } satisfies AnthropicOptions);
	}

	// For Opus 4.6 and Sonnet 4.6: use adaptive thinking with effort level
	// For older models: use budget-based thinking
	if (supportsAdaptiveThinking(model.id)) {
		const effort = mapThinkingLevelToEffort(options.reasoning, model.id);
		return streamAnthropic(model, context, {
			...base,
			thinkingEnabled: true,
			effort,
		} satisfies AnthropicOptions);
	}

	const adjusted = adjustMaxTokensForThinking(
		base.maxTokens || 0,
		model.maxTokens,
		options.reasoning,
		options.thinkingBudgets,
	);

	return streamAnthropic(model, context, {
		...base,
		maxTokens: adjusted.maxTokens,
		thinkingEnabled: true,
		thinkingBudgetTokens: adjusted.thinkingBudget,
	} satisfies AnthropicOptions);
};

function isOAuthToken(apiKey: string): boolean {
	return apiKey.includes("sk-ant-oat");
}

function createClient(
	model: Model<"anthropic-messages">,
	apiKey: string,
	interleavedThinking: boolean,
	optionsHeaders?: Record<string, string>,
	dynamicHeaders?: Record<string, string>,
): { client: Anthropic; isOAuthToken: boolean } {
	// Adaptive thinking models (Opus 4.6, Sonnet 4.6) have interleaved thinking built-in.
	// The beta header is deprecated on Opus 4.6 and redundant on Sonnet 4.6, so skip it.
	const needsInterleavedBeta = interleavedThinking && !supportsAdaptiveThinking(model.id);

	// Copilot: Bearer auth, selective betas (no fine-grained-tool-streaming)
	if (model.provider === "github-copilot") {
		const betaFeatures: string[] = [];
		if (needsInterleavedBeta) {
			betaFeatures.push("interleaved-thinking-2025-05-14");
		}

		const client = new Anthropic({
			apiKey: null,
			authToken: apiKey,
			baseURL: model.baseUrl,
			dangerouslyAllowBrowser: true,
			defaultHeaders: mergeHeaders(
				{
					accept: "application/json",
					"anthropic-dangerous-direct-browser-access": "true",
					...(betaFeatures.length > 0 ? { "anthropic-beta": betaFeatures.join(",") } : {}),
				},
				model.headers,
				dynamicHeaders,
				optionsHeaders,
			),
		});

		return { client, isOAuthToken: false };
	}

	const betaFeatures = ["fine-grained-tool-streaming-2025-05-14"];
	if (needsInterleavedBeta) {
		betaFeatures.push("interleaved-thinking-2025-05-14");
	}

	// OAuth: Bearer auth, Claude Code identity headers
	if (isOAuthToken(apiKey)) {
		const client = new Anthropic({
			apiKey: null,
			authToken: apiKey,
			baseURL: model.baseUrl,
			dangerouslyAllowBrowser: true,
			defaultHeaders: mergeHeaders(
				{
					accept: "application/json",
					"anthropic-dangerous-direct-browser-access": "true",
					"anthropic-beta": `claude-code-20250219,oauth-2025-04-20,${betaFeatures.join(",")}`,
					"user-agent": `claude-cli/${claudeCodeVersion}`,
					"x-app": "cli",
				},
				model.headers,
				optionsHeaders,
			),
		});

		return { client, isOAuthToken: true };
	}

	// API key auth
	const client = new Anthropic({
		apiKey,
		baseURL: model.baseUrl,
		dangerouslyAllowBrowser: true,
		defaultHeaders: mergeHeaders(
			{
				accept: "application/json",
				"anthropic-dangerous-direct-browser-access": "true",
				"anthropic-beta": betaFeatures.join(","),
			},
			model.headers,
			optionsHeaders,
		),
	});

	return { client, isOAuthToken: false };
}

function buildParams(
	model: Model<"anthropic-messages">,
	context: Context,
	isOAuthToken: boolean,
	options?: AnthropicOptions,
): MessageCreateParamsStreaming {
	const { cacheControl } = getCacheControl(model.baseUrl, options?.cacheRetention);
	const params: MessageCreateParamsStreaming = {
		model: model.id,
		messages: convertMessages(context.messages, model, isOAuthToken, cacheControl),
		max_tokens: options?.maxTokens || (model.maxTokens / 3) | 0,
		stream: true,
	};

	// For OAuth tokens, we MUST include Claude Code identity
	if (isOAuthToken) {
		params.system = [
			{
				type: "text",
				text: "You are Claude Code, Anthropic's official CLI for Claude.",
				...(cacheControl ? { cache_control: cacheControl } : {}),
			},
		];
		if (context.systemPrompt) {
			params.system.push({
				type: "text",
				text: sanitizeSurrogates(context.systemPrompt),
				...(cacheControl ? { cache_control: cacheControl } : {}),
			});
		}
	} else if (context.systemPrompt) {
		// Add cache control to system prompt for non-OAuth tokens
		params.system = [
			{
				type: "text",
				text: sanitizeSurrogates(context.systemPrompt),
				...(cacheControl ? { cache_control: cacheControl } : {}),
			},
		];
	}

	// Temperature is incompatible with extended thinking (adaptive or budget-based).
	if (options?.temperature !== undefined && !options?.thinkingEnabled) {
		params.temperature = options.temperature;
	}

	if (context.tools) {
		params.tools = convertTools(context.tools, isOAuthToken, cacheControl);
	}

	// Configure thinking mode: adaptive (Opus 4.6+ and Sonnet 4.6),
	// budget-based (older models), or explicitly disabled.
	if (model.reasoning) {
		if (options?.thinkingEnabled) {
			// Default to "summarized" so Opus 4.7 and Mythos Preview behave like
			// older Claude 4 models (whose API default is also "summarized").
			const display: AnthropicThinkingDisplay = options.thinkingDisplay ?? "summarized";
			if (supportsAdaptiveThinking(model.id)) {
				// Adaptive thinking: Claude decides when and how much to think.
				params.thinking = { type: "adaptive", display };
				if (options.effort) {
					// The Anthropic SDK types can lag newly supported effort values such as "xhigh".
					params.output_config =
						options.effort === "xhigh"
							? ({ effort: options.effort } as unknown as NonNullable<
									MessageCreateParamsStreaming["output_config"]
								>)
							: { effort: options.effort };
				}
			} else {
				// Budget-based thinking for older models
				params.thinking = {
					type: "enabled",
					budget_tokens: options.thinkingBudgetTokens || 1024,
					display,
				};
			}
		} else if (options?.thinkingEnabled === false) {
			params.thinking = { type: "disabled" };
		}
	}

	if (options?.metadata) {
		const userId = options.metadata.user_id;
		if (typeof userId === "string") {
			params.metadata = { user_id: userId };
		}
	}

	if (options?.toolChoice) {
		if (typeof options.toolChoice === "string") {
			params.tool_choice = { type: options.toolChoice };
		} else {
			params.tool_choice = options.toolChoice;
		}
	}

	return params;
}

// Normalize tool call IDs to match Anthropic's required pattern and length
function normalizeToolCallId(id: string): string {
	return id.replace(/[^a-zA-Z0-9_-]/g, "_").slice(0, 64);
}

function convertMessages(
	messages: Message[],
	model: Model<"anthropic-messages">,
	isOAuthToken: boolean,
	cacheControl?: CacheControlEphemeral,
): MessageParam[] {
	const params: MessageParam[] = [];

	// Transform messages for cross-provider compatibility
	const transformedMessages = transformMessages(messages, model, normalizeToolCallId);

	for (let i = 0; i < transformedMessages.length; i++) {
		const msg = transformedMessages[i];

		if (msg.role === "user") {
			if (typeof msg.content === "string") {
				if (msg.content.trim().length > 0) {
					params.push({
						role: "user",
						content: sanitizeSurrogates(msg.content),
					});
				}
			} else {
				const blocks: ContentBlockParam[] = msg.content.map((item) => {
					if (item.type === "text") {
						return {
							type: "text",
							text: sanitizeSurrogates(item.text),
						};
					} else {
						return {
							type: "image",
							source: {
								type: "base64",
								media_type: item.mimeType as "image/jpeg" | "image/png" | "image/gif" | "image/webp",
								data: item.data,
							},
						};
					}
				});
				let filteredBlocks = !model?.input.includes("image") ? blocks.filter((b) => b.type !== "image") : blocks;
				filteredBlocks = filteredBlocks.filter((b) => {
					if (b.type === "text") {
						return b.text.trim().length > 0;
					}
					return true;
				});
				if (filteredBlocks.length === 0) continue;
				params.push({
					role: "user",
					content: filteredBlocks,
				});
			}
		} else if (msg.role === "assistant") {
			const blocks: ContentBlockParam[] = [];

			for (const block of msg.content) {
				if (block.type === "text") {
					if (block.text.trim().length === 0) continue;
					blocks.push({
						type: "text",
						text: sanitizeSurrogates(block.text),
					});
				} else if (block.type === "thinking") {
					// Redacted thinking: pass the opaque payload back as redacted_thinking
					if (block.redacted) {
						blocks.push({
							type: "redacted_thinking",
							data: block.thinkingSignature!,
						});
						continue;
					}
					if (block.thinking.trim().length === 0) continue;
					// If thinking signature is missing/empty (e.g., from aborted stream),
					// convert to plain text block without <thinking> tags to avoid API rejection
					// and prevent Claude from mimicking the tags in responses
					if (!block.thinkingSignature || block.thinkingSignature.trim().length === 0) {
						blocks.push({
							type: "text",
							text: sanitizeSurrogates(block.thinking),
						});
					} else {
						blocks.push({
							type: "thinking",
							thinking: sanitizeSurrogates(block.thinking),
							signature: block.thinkingSignature,
						});
					}
				} else if (block.type === "toolCall") {
					blocks.push({
						type: "tool_use",
						id: block.id,
						name: isOAuthToken ? toClaudeCodeName(block.name) : block.name,
						input: block.arguments ?? {},
					});
				}
			}
			if (blocks.length === 0) continue;
			params.push({
				role: "assistant",
				content: blocks,
			});
		} else if (msg.role === "toolResult") {
			// Collect all consecutive toolResult messages, needed for z.ai Anthropic endpoint
			const toolResults: ContentBlockParam[] = [];

			// Add the current tool result
			toolResults.push({
				type: "tool_result",
				tool_use_id: msg.toolCallId,
				content: convertContentBlocks(msg.content),
				is_error: msg.isError,
			});

			// Look ahead for consecutive toolResult messages
			let j = i + 1;
			while (j < transformedMessages.length && transformedMessages[j].role === "toolResult") {
				const nextMsg = transformedMessages[j] as ToolResultMessage; // We know it's a toolResult
				toolResults.push({
					type: "tool_result",
					tool_use_id: nextMsg.toolCallId,
					content: convertContentBlocks(nextMsg.content),
					is_error: nextMsg.isError,
				});
				j++;
			}

			// Skip the messages we've already processed
			i = j - 1;

			// Add a single user message with all tool results
			params.push({
				role: "user",
				content: toolResults,
			});
		}
	}

	// Add cache_control to the last user message to cache conversation history
	if (cacheControl && params.length > 0) {
		const lastMessage = params[params.length - 1];
		if (lastMessage.role === "user") {
			if (Array.isArray(lastMessage.content)) {
				const lastBlock = lastMessage.content[lastMessage.content.length - 1];
				if (
					lastBlock &&
					(lastBlock.type === "text" || lastBlock.type === "image" || lastBlock.type === "tool_result")
				) {
					(lastBlock as any).cache_control = cacheControl;
				}
			} else if (typeof lastMessage.content === "string") {
				lastMessage.content = [
					{
						type: "text",
						text: lastMessage.content,
						cache_control: cacheControl,
					},
				] as any;
			}
		}
	}

	return params;
}

function convertTools(
	tools: Tool[],
	isOAuthToken: boolean,
	cacheControl?: CacheControlEphemeral,
): Anthropic.Messages.Tool[] {
	if (!tools) return [];

	return tools.map((tool, index) => {
		const schema = tool.parameters as { properties?: unknown; required?: string[] };

		return {
			name: isOAuthToken ? toClaudeCodeName(tool.name) : tool.name,
			description: tool.description,
			input_schema: {
				type: "object",
				properties: schema.properties ?? {},
				required: schema.required ?? [],
			},
			...(cacheControl && index === tools.length - 1 ? { cache_control: cacheControl } : {}),
		};
	});
}

function mapStopReason(reason: Anthropic.Messages.StopReason | string): StopReason {
	switch (reason) {
		case "end_turn":
			return "stop";
		case "max_tokens":
			return "length";
		case "tool_use":
			return "toolUse";
		case "refusal":
			return "error";
		case "pause_turn": // Stop is good enough -> resubmit
			return "stop";
		case "stop_sequence":
			return "stop"; // We don't supply stop sequences, so this should never happen
		case "sensitive": // Content flagged by safety filters (not yet in SDK types)
			return "error";
		default:
			// Handle unknown stop reasons gracefully (API may add new values)
			throw new Error(`Unhandled stop reason: ${reason}`);
	}
}