Merge pull request #148 from transitive-bullshit/feature/browser-improvements

pull/150/head
Travis Fischer 2022-12-16 01:34:53 -06:00 zatwierdzone przez GitHub
commit f5c277f1d4
Nie znaleziono w bazie danych klucza dla tego podpisu
ID klucza GPG: 4AEE18F83AFDEB23
6 zmienionych plików z 691 dodań i 111 usunięć

Wyświetl plik

@ -1,3 +1,4 @@
import delay from 'delay'
import dotenv from 'dotenv-safe'
import { oraPromise } from 'ora'
@ -16,7 +17,12 @@ async function main() {
const email = process.env.OPENAI_EMAIL
const password = process.env.OPENAI_PASSWORD
const api = new ChatGPTAPIBrowser({ email, password, debug: true })
const api = new ChatGPTAPIBrowser({
email,
password,
debug: false,
minimize: true
})
await api.init()
const prompt =

Wyświetl plik

@ -24,7 +24,7 @@ const response = await api.sendMessage('Hello World!')
Note that this solution is not lightweight, but it does work a lot more consistently than the REST API-based versions. I'm currently using this solution to power 10 OpenAI accounts concurrently across 10 minimized Chrome windows for my [Twitter bot](https://github.com/transitive-bullshit/chatgpt-twitter-bot). 😂
If you get a "ChatGPT is at capacity" error when logging in, note that this is also happening quite frequently on the official webapp. Their servers are overloaded, and we're all trying our best to offer access to this amazing technology.
If you get a "ChatGPT is at capacity" error when logging in, note that this can also happen on the official webapp as well. Their servers get overloaded at times, and we're all trying our best to offer access to this amazing technology.
To use the updated version, **make sure you're using the latest version of this package and Node.js >= 18**. Then update your code following the examples below, paying special attention to the sections on [Authentication](#authentication) and [Restrictions](#restrictions).
@ -240,8 +240,6 @@ Pass `sessionToken`, `clearanceToken`, and `userAgent` to the `ChatGPTAPI` const
These restrictions are for the `getOpenAIAuth` + `ChatGPTAPI` solution, which uses the unofficial API. The browser-based solution, `ChatGPTAPIBrowser`, doesn't have many of these restrictions, though you'll still have to manually bypass CAPTCHAs by hand.
Note: currently `ChatGPTAPIBrowser` doesn't support continuing arbitrary conversations based on `conversationId`. You can only continue conversations in the current tab or start new conversations using the `resetThread()` function.
**Please read carefully**
- You must use `node >= 18` at the moment. I'm using `v19.2.0` in my testing.

Wyświetl plik

@ -1,16 +1,23 @@
import delay from 'delay'
import html2md from 'html-to-md'
import pTimeout from 'p-timeout'
import type { Browser, HTTPRequest, HTTPResponse, Page } from 'puppeteer'
import { v4 as uuidv4 } from 'uuid'
import * as types from './types'
import { getBrowser, getOpenAIAuth } from './openai-auth'
import { isRelevantRequest, maximizePage, minimizePage } from './utils'
import {
browserPostEventStream,
isRelevantRequest,
maximizePage,
minimizePage
} from './utils'
export class ChatGPTAPIBrowser {
protected _markdown: boolean
protected _debug: boolean
protected _minimize: boolean
protected _isGoogleLogin: boolean
protected _captchaToken: string
protected _accessToken: string
protected _email: string
protected _password: string
@ -31,7 +38,13 @@ export class ChatGPTAPIBrowser {
/** @defaultValue `false` **/
debug?: boolean
/** @defaultValue `false` **/
isGoogleLogin?: boolean
/** @defaultValue `true` **/
minimize?: boolean
/** @defaultValue `undefined` **/
captchaToken?: string
}) {
const {
@ -40,6 +53,7 @@ export class ChatGPTAPIBrowser {
markdown = true,
debug = false,
isGoogleLogin = false,
minimize = true,
captchaToken
} = opts
@ -49,6 +63,7 @@ export class ChatGPTAPIBrowser {
this._markdown = !!markdown
this._debug = !!debug
this._isGoogleLogin = !!isGoogleLogin
this._minimize = !!minimize
this._captchaToken = captchaToken
}
@ -64,6 +79,9 @@ export class ChatGPTAPIBrowser {
this._page =
(await this._browser.pages())[0] || (await this._browser.newPage())
this._page.on('request', this._onRequest.bind(this))
this._page.on('response', this._onResponse.bind(this))
// bypass cloudflare and login
await getOpenAIAuth({
email: this._email,
@ -114,10 +132,9 @@ export class ChatGPTAPIBrowser {
return false
}
await minimizePage(this._page)
this._page.on('request', this._onRequest.bind(this))
this._page.on('response', this._onResponse.bind(this))
if (this._minimize) {
await minimizePage(this._page)
}
return true
}
@ -197,6 +214,12 @@ export class ChatGPTAPIBrowser {
} else if (url.endsWith('api/auth/session')) {
if (status === 403) {
await this.handle403Error()
} else {
const session: types.SessionResult = body
if (session?.accessToken) {
this._accessToken = session.accessToken
}
}
}
}
@ -209,7 +232,9 @@ export class ChatGPTAPIBrowser {
waitUntil: 'networkidle2',
timeout: 2 * 60 * 1000 // 2 minutes
})
await minimizePage(this._page)
if (this._minimize) {
await minimizePage(this._page)
}
} catch (err) {
console.error(
`ChatGPT "${this._email}" error refreshing session`,
@ -228,121 +253,182 @@ export class ChatGPTAPIBrowser {
}
}
async getLastMessage(): Promise<string | null> {
const messages = await this.getMessages()
// async getLastMessage(): Promise<string | null> {
// const messages = await this.getMessages()
if (messages) {
return messages[messages.length - 1]
} else {
return null
}
}
// if (messages) {
// return messages[messages.length - 1]
// } else {
// return null
// }
// }
async getPrompts(): Promise<string[]> {
// Get all prompts
const messages = await this._page.$$(
'.text-base:has(.whitespace-pre-wrap):not(:has(button:nth-child(2))) .whitespace-pre-wrap'
)
// async getPrompts(): Promise<string[]> {
// // Get all prompts
// const messages = await this._page.$$(
// '.text-base:has(.whitespace-pre-wrap):not(:has(button:nth-child(2))) .whitespace-pre-wrap'
// )
// Prompts are always plaintext
return Promise.all(messages.map((a) => a.evaluate((el) => el.textContent)))
}
// // Prompts are always plaintext
// return Promise.all(messages.map((a) => a.evaluate((el) => el.textContent)))
// }
async getMessages(): Promise<string[]> {
// Get all complete messages
// (in-progress messages that are being streamed back don't contain action buttons)
const messages = await this._page.$$(
'.text-base:has(.whitespace-pre-wrap):has(button:nth-child(2)) .whitespace-pre-wrap'
)
// async getMessages(): Promise<string[]> {
// // Get all complete messages
// // (in-progress messages that are being streamed back don't contain action buttons)
// const messages = await this._page.$$(
// '.text-base:has(.whitespace-pre-wrap):has(button:nth-child(2)) .whitespace-pre-wrap'
// )
if (this._markdown) {
const htmlMessages = await Promise.all(
messages.map((a) => a.evaluate((el) => el.innerHTML))
)
// if (this._markdown) {
// const htmlMessages = await Promise.all(
// messages.map((a) => a.evaluate((el) => el.innerHTML))
// )
const markdownMessages = htmlMessages.map((messageHtml) => {
// parse markdown from message HTML
messageHtml = messageHtml
.replaceAll('Copy code</button>', '</button>')
.replace(/Copy code\s*<\/button>/gim, '</button>')
// const markdownMessages = htmlMessages.map((messageHtml) => {
// // parse markdown from message HTML
// messageHtml = messageHtml
// .replaceAll('Copy code</button>', '</button>')
// .replace(/Copy code\s*<\/button>/gim, '</button>')
return html2md(messageHtml, {
ignoreTags: [
'button',
'svg',
'style',
'form',
'noscript',
'script',
'meta',
'head'
],
skipTags: ['button', 'svg']
})
})
// return html2md(messageHtml, {
// ignoreTags: [
// 'button',
// 'svg',
// 'style',
// 'form',
// 'noscript',
// 'script',
// 'meta',
// 'head'
// ],
// skipTags: ['button', 'svg']
// })
// })
return markdownMessages
} else {
// plaintext
const plaintextMessages = await Promise.all(
messages.map((a) => a.evaluate((el) => el.textContent))
)
return plaintextMessages
}
}
// return markdownMessages
// } else {
// // plaintext
// const plaintextMessages = await Promise.all(
// messages.map((a) => a.evaluate((el) => el.textContent))
// )
// return plaintextMessages
// }
// }
async sendMessage(
message: string,
opts: {
timeoutMs?: number
} = {}
opts: types.SendMessageOptions = {}
): Promise<string> {
const { timeoutMs } = opts
const {
conversationId,
parentMessageId = uuidv4(),
messageId = uuidv4(),
action = 'next',
// TODO
timeoutMs,
// onProgress,
onConversationResponse
} = opts
const inputBox = await this._getInputBox()
if (!inputBox) throw new Error('not signed in')
const lastMessage = await this.getLastMessage()
await inputBox.focus()
const paragraphs = message.split('\n')
for (let i = 0; i < paragraphs.length; i++) {
await inputBox.type(paragraphs[i], { delay: 0 })
if (i < paragraphs.length - 1) {
await this._page.keyboard.down('Shift')
await inputBox.press('Enter')
await this._page.keyboard.up('Shift')
} else {
await inputBox.press('Enter')
}
if (!inputBox || !this._accessToken) {
const error = new types.ChatGPTError('Not signed in')
error.statusCode = 401
throw error
}
const responseP = new Promise<string>(async (resolve, reject) => {
try {
do {
await delay(1000)
// TODO: this logic needs some work because we can have repeat messages...
const newLastMessage = await this.getLastMessage()
if (
newLastMessage &&
lastMessage?.toLowerCase() !== newLastMessage?.toLowerCase()
) {
return resolve(newLastMessage)
const url = `https://chat.openai.com/backend-api/conversation`
const body: types.ConversationJSONBody = {
action,
messages: [
{
id: messageId,
role: 'user',
content: {
content_type: 'text',
parts: [message]
}
} while (true)
} catch (err) {
return reject(err)
}
})
if (timeoutMs) {
return pTimeout(responseP, {
milliseconds: timeoutMs
})
} else {
return responseP
}
],
model: 'text-davinci-002-render',
parent_message_id: parentMessageId
}
if (conversationId) {
body.conversation_id = conversationId
}
// console.log('>>> EVALUATE', url, this._accessToken, body)
const result = await this._page.evaluate(
browserPostEventStream,
url,
this._accessToken,
body,
timeoutMs
)
// console.log('<<< EVALUATE', result)
if (result.error) {
const error = new types.ChatGPTError(result.error.message)
error.statusCode = result.error.statusCode
error.statusText = result.error.statusText
if (error.statusCode === 403) {
await this.handle403Error()
}
throw error
}
// TODO: support sending partial response events
if (onConversationResponse) {
onConversationResponse(result.conversationResponse)
}
return result.response
// const lastMessage = await this.getLastMessage()
// await inputBox.focus()
// const paragraphs = message.split('\n')
// for (let i = 0; i < paragraphs.length; i++) {
// await inputBox.type(paragraphs[i], { delay: 0 })
// if (i < paragraphs.length - 1) {
// await this._page.keyboard.down('Shift')
// await inputBox.press('Enter')
// await this._page.keyboard.up('Shift')
// } else {
// await inputBox.press('Enter')
// }
// }
// const responseP = new Promise<string>(async (resolve, reject) => {
// try {
// do {
// await delay(1000)
// // TODO: this logic needs some work because we can have repeat messages...
// const newLastMessage = await this.getLastMessage()
// if (
// newLastMessage &&
// lastMessage?.toLowerCase() !== newLastMessage?.toLowerCase()
// ) {
// return resolve(newLastMessage)
// }
// } while (true)
// } catch (err) {
// return reject(err)
// }
// })
// if (timeoutMs) {
// return pTimeout(responseP, {
// milliseconds: timeoutMs
// })
// } else {
// return responseP
// }
}
async resetThread() {

Wyświetl plik

@ -127,7 +127,7 @@ export async function getOpenAIAuth({
}
await page.click('button[type="submit"]')
await page.waitForSelector('#password')
await page.waitForSelector('#password', { timeout: timeoutMs })
await page.type('#password', password, { delay: 10 })
submitP = () => page.click('button[type="submit"]')
}

Wyświetl plik

@ -297,3 +297,19 @@ export class ChatGPTError extends Error {
response?: Response
originalError?: Error
}
export type ChatError = {
error: { message: string; statusCode?: number; statusText?: string }
response: null
conversationId?: string
messageId?: string
conversationResponse?: ConversationResponseEvent
}
export type ChatResponse = {
error: null
response: string
conversationId: string
messageId: string
conversationResponse?: ConversationResponseEvent
}

Wyświetl plik

@ -1,7 +1,14 @@
import type * as PTimeoutTypes from 'p-timeout'
import type {
EventSourceParseCallback,
EventSourceParser
} from 'eventsource-parser'
import type { Page } from 'puppeteer'
import { remark } from 'remark'
import stripMarkdown from 'strip-markdown'
import * as types from './types'
export function markdownToText(markdown?: string): string {
return remark()
.use(stripMarkdown)
@ -44,7 +51,10 @@ export function isRelevantRequest(url: string): boolean {
return false
}
if (pathname.startsWith('/_next')) {
if (
!pathname.startsWith('/backend-api/') &&
!pathname.startsWith('/api/auth/session')
) {
return false
}
@ -54,3 +64,467 @@ export function isRelevantRequest(url: string): boolean {
return true
}
/**
* This function is injected into the ChatGPT webapp page using puppeteer. It
* has to be fully self-contained, so we copied a few third-party sources and
* included them in here.
*/
export async function browserPostEventStream(
url: string,
accessToken: string,
body: types.ConversationJSONBody,
timeoutMs?: number
): Promise<types.ChatError | types.ChatResponse> {
// Workaround for https://github.com/esbuild-kit/tsx/issues/113
globalThis.__name = () => undefined
class TimeoutError extends Error {
readonly name: 'TimeoutError'
constructor(message) {
super(message)
this.name = 'TimeoutError'
}
}
/**
An error to be thrown when the request is aborted by AbortController.
DOMException is thrown instead of this Error when DOMException is available.
*/
class AbortError extends Error {
constructor(message) {
super()
this.name = 'AbortError'
this.message = message
}
}
const BOM = [239, 187, 191]
let conversationResponse: types.ConversationResponseEvent
let conversationId: string = body?.conversation_id
let messageId: string = body?.messages?.[0]?.id
let response = ''
try {
console.log('browserPostEventStream', url, accessToken, body)
let abortController: AbortController = null
if (timeoutMs) {
abortController = new AbortController()
}
const res = await fetch(url, {
method: 'POST',
body: JSON.stringify(body),
signal: abortController?.signal,
headers: {
accept: 'text/event-stream',
'x-openai-assistant-app-id': '',
authorization: `Bearer ${accessToken}`,
'content-type': 'application/json'
}
})
console.log('browserPostEventStream response', res)
if (!res.ok) {
return {
error: {
message: `ChatGPTAPI error ${res.status || res.statusText}`,
statusCode: res.status,
statusText: res.statusText
},
response: null,
conversationId,
messageId
}
}
const responseP = new Promise<types.ChatResponse>(
async (resolve, reject) => {
function onMessage(data: string) {
if (data === '[DONE]') {
return resolve({
error: null,
response,
conversationId,
messageId,
conversationResponse
})
}
try {
const convoResponseEvent: types.ConversationResponseEvent =
JSON.parse(data)
conversationResponse = convoResponseEvent
if (convoResponseEvent.conversation_id) {
conversationId = convoResponseEvent.conversation_id
}
if (convoResponseEvent.message?.id) {
messageId = convoResponseEvent.message.id
}
const partialResponse =
convoResponseEvent.message?.content?.parts?.[0]
if (partialResponse) {
response = partialResponse
}
} catch (err) {
console.warn('fetchSSE onMessage unexpected error', err)
reject(err)
}
}
const parser = createParser((event) => {
if (event.type === 'event') {
onMessage(event.data)
}
})
for await (const chunk of streamAsyncIterable(res.body)) {
const str = new TextDecoder().decode(chunk)
parser.feed(str)
}
}
)
if (timeoutMs) {
if (abortController) {
// This will be called when a timeout occurs in order for us to forcibly
// ensure that the underlying HTTP request is aborted.
;(responseP as any).cancel = () => {
abortController.abort()
}
}
return await pTimeout(responseP, {
milliseconds: timeoutMs,
message: 'ChatGPT timed out waiting for response'
})
} else {
return await responseP
}
} catch (err) {
const errMessageL = err.toString().toLowerCase()
if (
response &&
(errMessageL === 'error: typeerror: terminated' ||
errMessageL === 'typeerror: terminated')
) {
// OpenAI sometimes forcefully terminates the socket from their end before
// the HTTP request has resolved cleanly. In my testing, these cases tend to
// happen when OpenAI has already send the last `response`, so we can ignore
// the `fetch` error in this case.
return {
error: null,
response,
conversationId,
messageId,
conversationResponse
}
}
return {
error: {
message: err.toString(),
statusCode: err.statusCode || err.status || err.response?.statusCode,
statusText: err.statusText || err.response?.statusText
},
response: null,
conversationId,
messageId,
conversationResponse
}
}
async function* streamAsyncIterable<T>(stream: ReadableStream<T>) {
const reader = stream.getReader()
try {
while (true) {
const { done, value } = await reader.read()
if (done) {
return
}
yield value
}
} finally {
reader.releaseLock()
}
}
// @see https://github.com/rexxars/eventsource-parser
function createParser(onParse: EventSourceParseCallback): EventSourceParser {
// Processing state
let isFirstChunk: boolean
let buffer: string
let startingPosition: number
let startingFieldLength: number
// Event state
let eventId: string | undefined
let eventName: string | undefined
let data: string
reset()
return { feed, reset }
function reset(): void {
isFirstChunk = true
buffer = ''
startingPosition = 0
startingFieldLength = -1
eventId = undefined
eventName = undefined
data = ''
}
function feed(chunk: string): void {
buffer = buffer ? buffer + chunk : chunk
// Strip any UTF8 byte order mark (BOM) at the start of the stream.
// Note that we do not strip any non - UTF8 BOM, as eventsource streams are
// always decoded as UTF8 as per the specification.
if (isFirstChunk && hasBom(buffer)) {
buffer = buffer.slice(BOM.length)
}
isFirstChunk = false
// Set up chunk-specific processing state
const length = buffer.length
let position = 0
let discardTrailingNewline = false
// Read the current buffer byte by byte
while (position < length) {
// EventSource allows for carriage return + line feed, which means we
// need to ignore a linefeed character if the previous character was a
// carriage return
// @todo refactor to reduce nesting, consider checking previous byte?
// @todo but consider multiple chunks etc
if (discardTrailingNewline) {
if (buffer[position] === '\n') {
++position
}
discardTrailingNewline = false
}
let lineLength = -1
let fieldLength = startingFieldLength
let character: string
for (
let index = startingPosition;
lineLength < 0 && index < length;
++index
) {
character = buffer[index]
if (character === ':' && fieldLength < 0) {
fieldLength = index - position
} else if (character === '\r') {
discardTrailingNewline = true
lineLength = index - position
} else if (character === '\n') {
lineLength = index - position
}
}
if (lineLength < 0) {
startingPosition = length - position
startingFieldLength = fieldLength
break
} else {
startingPosition = 0
startingFieldLength = -1
}
parseEventStreamLine(buffer, position, fieldLength, lineLength)
position += lineLength + 1
}
if (position === length) {
// If we consumed the entire buffer to read the event, reset the buffer
buffer = ''
} else if (position > 0) {
// If there are bytes left to process, set the buffer to the unprocessed
// portion of the buffer only
buffer = buffer.slice(position)
}
}
function parseEventStreamLine(
lineBuffer: string,
index: number,
fieldLength: number,
lineLength: number
) {
if (lineLength === 0) {
// We reached the last line of this event
if (data.length > 0) {
onParse({
type: 'event',
id: eventId,
event: eventName || undefined,
data: data.slice(0, -1) // remove trailing newline
})
data = ''
eventId = undefined
}
eventName = undefined
return
}
const noValue = fieldLength < 0
const field = lineBuffer.slice(
index,
index + (noValue ? lineLength : fieldLength)
)
let step = 0
if (noValue) {
step = lineLength
} else if (lineBuffer[index + fieldLength + 1] === ' ') {
step = fieldLength + 2
} else {
step = fieldLength + 1
}
const position = index + step
const valueLength = lineLength - step
const value = lineBuffer
.slice(position, position + valueLength)
.toString()
if (field === 'data') {
data += value ? `${value}\n` : '\n'
} else if (field === 'event') {
eventName = value
} else if (field === 'id' && !value.includes('\u0000')) {
eventId = value
} else if (field === 'retry') {
const retry = parseInt(value, 10)
if (!Number.isNaN(retry)) {
onParse({ type: 'reconnect-interval', value: retry })
}
}
}
}
function hasBom(buffer: string) {
return BOM.every(
(charCode: number, index: number) => buffer.charCodeAt(index) === charCode
)
}
/**
TODO: Remove AbortError and just throw DOMException when targeting Node 18.
*/
function getDOMException(errorMessage) {
return globalThis.DOMException === undefined
? new AbortError(errorMessage)
: new DOMException(errorMessage)
}
/**
TODO: Remove below function and just 'reject(signal.reason)' when targeting Node 18.
*/
function getAbortedReason(signal) {
const reason =
signal.reason === undefined
? getDOMException('This operation was aborted.')
: signal.reason
return reason instanceof Error ? reason : getDOMException(reason)
}
// @see https://github.com/sindresorhus/p-timeout
function pTimeout<ValueType, ReturnType = ValueType>(
promise: PromiseLike<ValueType>,
options: PTimeoutTypes.Options<ReturnType>
): PTimeoutTypes.ClearablePromise<ValueType | ReturnType> {
const {
milliseconds,
fallback,
message,
customTimers = { setTimeout, clearTimeout }
} = options
let timer
const cancelablePromise = new Promise((resolve, reject) => {
if (typeof milliseconds !== 'number' || Math.sign(milliseconds) !== 1) {
throw new TypeError(
`Expected \`milliseconds\` to be a positive number, got \`${milliseconds}\``
)
}
if (milliseconds === Number.POSITIVE_INFINITY) {
resolve(promise)
return
}
if (options.signal) {
const { signal } = options
if (signal.aborted) {
reject(getAbortedReason(signal))
}
signal.addEventListener('abort', () => {
reject(getAbortedReason(signal))
})
}
timer = customTimers.setTimeout.call(
undefined,
() => {
if (fallback) {
try {
resolve(fallback())
} catch (error) {
reject(error)
}
return
}
const errorMessage =
typeof message === 'string'
? message
: `Promise timed out after ${milliseconds} milliseconds`
const timeoutError =
message instanceof Error ? message : new TimeoutError(errorMessage)
if (typeof (promise as any).cancel === 'function') {
;(promise as any).cancel()
}
reject(timeoutError)
},
milliseconds
)
;(async () => {
try {
resolve(await promise)
} catch (error) {
reject(error)
} finally {
customTimers.clearTimeout.call(undefined, timer)
}
})()
})
;(cancelablePromise as any).clear = () => {
customTimers.clearTimeout.call(undefined, timer)
timer = undefined
}
return cancelablePromise as any
}
}