my-project-public

repository

loading code, commits, and activity

repositories

loading repo index

#1	/**
#2	* Conway Inference Client
#3	*
#4	* Wraps Conway's /v1/chat/completions endpoint (OpenAI-compatible).
#5	* The automaton pays for its own thinking through Conway credits.
#6	*/
#7
#8	import type {
#9	InferenceClient,
#10	ChatMessage,
#11	InferenceOptions,
#12	InferenceResponse,
#13	InferenceToolCall,
#14	TokenUsage,
#15	InferenceToolDefinition,
#16	} from "../types.js";
#17
#18	interface InferenceClientOptions {
#19	apiUrl: string;
#20	apiKey: string;
#21	defaultModel: string;
#22	maxTokens: number;
#23	lowComputeModel?: string;
#24	}
#25
#26	export function createInferenceClient(
#27	options: InferenceClientOptions,
#28	): InferenceClient {
#29	const { apiUrl, apiKey } = options;
#30	let currentModel = options.defaultModel;
#31	let maxTokens = options.maxTokens;
#32
#33	const chat = async (
#34	messages: ChatMessage[],
#35	opts?: InferenceOptions,
#36	): Promise<InferenceResponse> => {
#37	const model = opts?.model \|\| currentModel;
#38	const tools = opts?.tools;
#39
#40	// Newer models (o-series, gpt-5.x) require max_completion_tokens
#41	// GPT-4.1 and similar models also use max_completion_tokens
#42	const usesCompletionTokens =
#43	/^(o[1-9]\|gpt-5)/.test(model) \|\|
#44	model.includes("4.1") \|\|
#45	model.includes("4o-mini");
#46	const tokenLimit = opts?.maxTokens \|\| maxTokens;
#47
#48	const body: Record<string, unknown> = {
#49	model,
#50	messages: messages.map(formatMessage),
#51	stream: false,
#52	};
#53
#54	if (usesCompletionTokens) {
#55	body.max_completion_tokens = tokenLimit;
#56	} else {
#57	body.max_tokens = tokenLimit;
#58	}
#59
#60	if (opts?.temperature !== undefined) {
#61	body.temperature = opts.temperature;
#62	}
#63
#64	if (tools && tools.length > 0) {
#65	body.tools = tools;
#66	body.tool_choice = "auto";
#67	}
#68
#69	const resp = await fetch(`${apiUrl}/v1/chat/completions`, {
#70	method: "POST",
#71	headers: {
#72	"Content-Type": "application/json",
#73	Authorization: apiKey,
#74	},
#75	body: JSON.stringify(body),
#76	});
#77
#78	if (!resp.ok) {
#79	const text = await resp.text();
#80	throw new Error(
#81	`Inference error: ${resp.status}: ${text}`,
#82	);
#83	}
#84
#85	const data = await resp.json() as any;
#86	const choice = data.choices?.[0];
#87
#88	if (!choice) {
#89	throw new Error("No completion choice returned from inference");
#90	}
#91
#92	const message = choice.message;
#93	const usage: TokenUsage = {
#94	promptTokens: data.usage?.prompt_tokens \|\| 0,
#95	completionTokens: data.usage?.completion_tokens \|\| 0,
#96	totalTokens: data.usage?.total_tokens \|\| 0,
#97	};
#98
#99	const toolCalls: InferenceToolCall[] \| undefined =
#100	message.tool_calls?.map((tc: any) => ({
#101	id: tc.id,
#102	type: "function" as const,
#103	function: {
#104	name: tc.function.name,
#105	arguments: tc.function.arguments,
#106	},
#107	}));
#108
#109	return {
#110	id: data.id \|\| "",
#111	model: data.model \|\| model,
#112	message: {
#113	role: message.role,
#114	content: message.content \|\| "",
#115	tool_calls: toolCalls,
#116	},
#117	toolCalls,
#118	usage,
#119	finishReason: choice.finish_reason \|\| "stop",
#120	};
#121	};
#122
#123	const setLowComputeMode = (enabled: boolean): void => {
#124	if (enabled) {
#125	currentModel = "gpt-4o-mini";
#126	maxTokens = 4096;
#127	} else {
#128	currentModel = options.defaultModel;
#129	maxTokens = options.maxTokens;
#130	}
#131	};
#132
#133	const getDefaultModel = (): string => {
#134	return currentModel;
#135	};
#136
#137	return {
#138	chat,
#139	setLowComputeMode,
#140	getDefaultModel,
#141	};
#142	}
#143
#144	function formatMessage(
#145	msg: ChatMessage,
#146	): Record<string, unknown> {
#147	const formatted: Record<string, unknown> = {
#148	role: msg.role,
#149	content: msg.content,
#150	};
#151
#152	if (msg.name) formatted.name = msg.name;
#153	if (msg.tool_calls) formatted.tool_calls = msg.tool_calls;
#154	if (msg.tool_call_id) formatted.tool_call_id = msg.tool_call_id;
#155
#156	return formatted;
#157	}
#158

z6Mkq5mY3JWtxoxUobWcfNHm7AkRubgSWEZTkBVqZXJviFZ5/my-project-public