git.giacomolli.uk - onechat/blob - chatgpt.go

   1 package main
   2
   3 import (
   4         "bytes"
   5         "cmp"
   6         "context"
   7         "encoding/base64"
   8         "encoding/json"
   9         "errors"
  10         "fmt"
  11         "io"
  12         "net/http"
  13         "net/url"
  14         "os"
  15         "path/filepath"
  16         "strconv"
  17         "strings"
  18         "sync"
  19         "time"
  20 )
  21
  22 const (
  23         // chatgptIssuer is the OpenAI login service.
  24         chatgptIssuer = "https://auth.openai.com"
  25
  26         // chatgptClientID is the OAuth client ID of the Codex CLI. The ChatGPT
  27         // backend takes the tokens of this client only.
  28         chatgptClientID = "app_EMoamEEZ73f0CkXaXp7hrann"
  29
  30         // chatgptAuthClaim is the JWT claim that holds the ChatGPT account ID.
  31         chatgptAuthClaim = "https://api.openai.com/auth"
  32
  33         // chatgptRefreshMargin is how long before its expiry an access token is
  34         // refreshed.
  35         chatgptRefreshMargin = 60 * time.Second
  36
  37         // chatgptLoginTimeout is how long a device code login waits for the
  38         // user.
  39         chatgptLoginTimeout = 15 * time.Minute
  40
  41         // chatgptTokenFileName is the name of the token file in the data
  42         // directory.
  43         chatgptTokenFileName = "auth.json"
  44
  45         // chatgptDefaultPollInterval is the poll interval of a device code
  46         // login when the login service gives none.
  47         chatgptDefaultPollInterval = 5 * time.Second
  48 )
  49
  50 // errChatGPTLoginNeeded tells that no valid token is available, so the user
  51 // must run "onechat login".
  52 var errChatGPTLoginNeeded = errors.New(`login needed: run "onechat login"`)
  53
  54 // chatgptLogin runs "onechat login": the device code login into the token file
  55 // of the data directory. It holds the lock of the token file, so it does not
  56 // run while the daemon runs.
  57 func chatgptLogin(ctx context.Context, out io.Writer) (err error) {
  58         dir, err := dataDir()
  59         if err != nil {
  60                 return err
  61         }
  62
  63         tokens, err := lockTokens(dir)
  64         if err != nil {
  65                 return err
  66         }
  67
  68         defer func() {
  69                 err = errors.Join(err, tokens.Close())
  70         }()
  71
  72         return newChatGPTAuth(filepath.Join(dir, chatgptTokenFileName), chatgptIssuer).Login(ctx, out)
  73 }
  74
  75 // newChatGPTAuth returns the authorizer for the token file at path, with the
  76 // login service at issuer.
  77 func newChatGPTAuth(path, issuer string) *ChatGPTAuth {
  78         return &ChatGPTAuth{
  79                 path:   path,
  80                 issuer: issuer,
  81                 http:   &http.Client{},
  82         }
  83 }
  84
  85 // ChatGPTAuth authorizes model calls to the ChatGPT backend with the tokens in
  86 // the token file. The tokens belong to OneChat only. The Codex CLI rotates the
  87 // refresh token on each refresh, so two programs with one token file log each
  88 // other out.
  89 //
  90 // Before each call it reads the expiry from the access token. Within
  91 // chatgptRefreshMargin of the expiry it gets new tokens with the refresh token
  92 // and replaces the file atomically. A mutex lets one refresh run at a time, and
  93 // a call that waits for it uses the new tokens.
  94 type ChatGPTAuth struct {
  95         mutex  sync.Mutex
  96         path   string
  97         issuer string
  98         http   *http.Client
  99 }
 100
 101 // Authorize sets the access token and the account ID, after a refresh when the
 102 // token expires within chatgptRefreshMargin.
 103 func (a *ChatGPTAuth) Authorize(ctx context.Context, header http.Header) error {
 104         a.mutex.Lock()
 105         defer a.mutex.Unlock()
 106
 107         tokens, err := a.read()
 108         if err != nil {
 109                 return err
 110         }
 111
 112         expiry, found := chatgptTokenExpiry(tokens.AccessToken)
 113         if !found || !time.Now().Before(expiry.Add(-chatgptRefreshMargin)) {
 114                 tokens, err = a.refresh(ctx, tokens)
 115                 if err != nil {
 116                         return err
 117                 }
 118         }
 119
 120         header.Set("Authorization", "Bearer "+tokens.AccessToken)
 121         header.Set("ChatGPT-Account-Id", tokens.AccountID)
 122
 123         return nil
 124 }
 125
 126 // Login runs the device code login. It writes the link and the code for the
 127 // user to out, waits until the user approves the code, and writes the tokens
 128 // to the token file.
 129 func (a *ChatGPTAuth) Login(ctx context.Context, out io.Writer) error {
 130         a.mutex.Lock()
 131         defer a.mutex.Unlock()
 132
 133         var code ChatGPTDeviceCode
 134         err := a.postJSON(ctx, "/api/accounts/deviceauth/usercode", map[string]string{
 135                 "client_id": chatgptClientID,
 136         }, &code)
 137         if err != nil {
 138                 return fmt.Errorf("request device code (device code login must be allowed in the ChatGPT security settings): %w", err)
 139         }
 140
 141         userCode := cmp.Or(code.UserCode, code.UserCodeAlt)
 142         if code.DeviceAuthID == "" || userCode == "" {
 143                 return errors.New("device code response has no device ID or user code")
 144         }
 145
 146         _, err = fmt.Fprintf(out, "Open %s/codex/device on any device and enter this code:\n\n    %s\n\nWaiting for the approval...\n", a.issuer, userCode)
 147         if err != nil {
 148                 return err
 149         }
 150
 151         grant, err := a.poll(ctx, code.DeviceAuthID, userCode, chatgptPollInterval(code.Interval))
 152         if err != nil {
 153                 return err
 154         }
 155
 156         tokens, err := a.exchange(ctx, grant)
 157         if err != nil {
 158                 return err
 159         }
 160
 161         err = a.write(tokens)
 162         if err != nil {
 163                 return err
 164         }
 165
 166         _, err = fmt.Fprintln(out, "Logged in.")
 167
 168         return err
 169 }
 170
 171 // read reads the token file.
 172 func (a *ChatGPTAuth) read() (ChatGPTTokenFile, error) {
 173         data, err := os.ReadFile(a.path)
 174         if errors.Is(err, os.ErrNotExist) {
 175                 return ChatGPTTokenFile{}, errChatGPTLoginNeeded
 176         }
 177
 178         if err != nil {
 179                 return ChatGPTTokenFile{}, fmt.Errorf("read token file: %w", err)
 180         }
 181
 182         var tokens ChatGPTTokenFile
 183         err = json.Unmarshal(data, &tokens)
 184         if err == nil && (tokens.AccessToken == "" || tokens.RefreshToken == "" || tokens.AccountID == "") {
 185                 err = errors.New("a token or the account ID is missing")
 186         }
 187
 188         if err != nil {
 189                 return ChatGPTTokenFile{}, fmt.Errorf("token file: %w: %w", err, errChatGPTLoginNeeded)
 190         }
 191
 192         return tokens, nil
 193 }
 194
 195 // refresh gets new tokens with the refresh token and replaces the token file.
 196 // It keeps the old refresh token and ID token when the answer has none.
 197 func (a *ChatGPTAuth) refresh(ctx context.Context, tokens ChatGPTTokenFile) (ChatGPTTokenFile, error) {
 198         var response ChatGPTTokenResponse
 199         err := a.postJSON(ctx, "/oauth/token", map[string]string{
 200                 "client_id":     chatgptClientID,
 201                 "grant_type":    "refresh_token",
 202                 "refresh_token": tokens.RefreshToken,
 203                 "scope":         "openid profile email",
 204         }, &response)
 205         if err == nil && response.AccessToken == "" {
 206                 err = errors.New("refresh response has no access token")
 207         }
 208
 209         if err != nil {
 210                 return ChatGPTTokenFile{}, fmt.Errorf("refresh tokens: %w: %w", err, errChatGPTLoginNeeded)
 211         }
 212
 213         tokens.AccessToken = response.AccessToken
 214         tokens.RefreshToken = cmp.Or(response.RefreshToken, tokens.RefreshToken)
 215         tokens.IDToken = cmp.Or(response.IDToken, tokens.IDToken)
 216         tokens.AccountID = cmp.Or(chatgptAccountID(tokens.AccessToken), chatgptAccountID(tokens.IDToken), tokens.AccountID)
 217
 218         err = a.write(tokens)
 219         if err != nil {
 220                 return ChatGPTTokenFile{}, err
 221         }
 222
 223         return tokens, nil
 224 }
 225
 226 // postJSON posts body as JSON to the path of the login service and decodes the
 227 // answer into result.
 228 func (a *ChatGPTAuth) postJSON(ctx context.Context, path string, body, result any) error {
 229         data, err := json.Marshal(body)
 230         if err != nil {
 231                 return fmt.Errorf("encode request: %w", err)
 232         }
 233
 234         _, err = a.post(ctx, path, "application/json", data, result)
 235
 236         return err
 237 }
 238
 239 // poll asks for the grant of the device code until the user approves it or the
 240 // login times out.
 241 func (a *ChatGPTAuth) poll(ctx context.Context, deviceID, userCode string, interval time.Duration) (ChatGPTDeviceGrant, error) {
 242         ctx, cancel := context.WithTimeout(ctx, chatgptLoginTimeout)
 243         defer cancel()
 244
 245         body, err := json.Marshal(map[string]string{
 246                 "device_auth_id": deviceID,
 247                 "user_code":      userCode,
 248         })
 249         if err != nil {
 250                 return ChatGPTDeviceGrant{}, fmt.Errorf("encode poll request: %w", err)
 251         }
 252
 253         for {
 254                 var grant ChatGPTDeviceGrant
 255                 status, err := a.post(ctx, "/api/accounts/deviceauth/token", "application/json", body, &grant)
 256                 if err == nil && (grant.AuthorizationCode == "" || grant.CodeVerifier == "") {
 257                         err = errors.New("device grant has no authorization code or code verifier")
 258                 }
 259
 260                 if err == nil {
 261                         return grant, nil
 262                 }
 263
 264                 if status != http.StatusForbidden && status != http.StatusNotFound {
 265                         return ChatGPTDeviceGrant{}, fmt.Errorf("poll device code: %w", err)
 266                 }
 267
 268                 timer := time.NewTimer(interval)
 269                 select {
 270                 case <-ctx.Done():
 271                         timer.Stop()
 272                         return ChatGPTDeviceGrant{}, fmt.Errorf("wait for approval: %w", ctx.Err())
 273                 case <-timer.C:
 274                 }
 275         }
 276 }
 277
 278 // exchange trades the grant of the device code for tokens.
 279 func (a *ChatGPTAuth) exchange(ctx context.Context, grant ChatGPTDeviceGrant) (ChatGPTTokenFile, error) {
 280         form := url.Values{
 281                 "grant_type":    {"authorization_code"},
 282                 "code":          {grant.AuthorizationCode},
 283                 "redirect_uri":  {a.issuer + "/deviceauth/callback"},
 284                 "client_id":     {chatgptClientID},
 285                 "code_verifier": {grant.CodeVerifier},
 286         }
 287
 288         var response ChatGPTTokenResponse
 289         _, err := a.post(ctx, "/oauth/token", "application/x-www-form-urlencoded", []byte(form.Encode()), &response)
 290         if err != nil {
 291                 return ChatGPTTokenFile{}, fmt.Errorf("exchange code for tokens: %w", err)
 292         }
 293
 294         if response.AccessToken == "" || response.RefreshToken == "" || response.IDToken == "" {
 295                 return ChatGPTTokenFile{}, errors.New("token response has no access, refresh, or ID token")
 296         }
 297
 298         tokens := ChatGPTTokenFile{
 299                 AccessToken:  response.AccessToken,
 300                 RefreshToken: response.RefreshToken,
 301                 IDToken:      response.IDToken,
 302         }
 303
 304         tokens.AccountID = cmp.Or(chatgptAccountID(tokens.AccessToken), chatgptAccountID(tokens.IDToken))
 305         if tokens.AccountID == "" {
 306                 return ChatGPTTokenFile{}, errors.New("tokens have no ChatGPT account ID")
 307         }
 308
 309         return tokens, nil
 310 }
 311
 312 // write replaces the token file atomically. The file has mode 0600.
 313 func (a *ChatGPTAuth) write(tokens ChatGPTTokenFile) error {
 314         data, err := json.Marshal(tokens)
 315         if err != nil {
 316                 return fmt.Errorf("encode tokens: %w", err)
 317         }
 318
 319         err = writeFileAtomic(a.path, append(data, '\n'))
 320         if err != nil {
 321                 return fmt.Errorf("write token file: %w", err)
 322         }
 323
 324         return nil
 325 }
 326
 327 // post posts body to the path of the login service and decodes a 200 answer
 328 // into result. It returns the HTTP status, or 0 when no answer came. The error
 329 // holds the status, but not the body, which can hold a token.
 330 func (a *ChatGPTAuth) post(ctx context.Context, path, contentType string, body []byte, result any) (status int, err error) {
 331         request, err := http.NewRequestWithContext(ctx, http.MethodPost, a.issuer+path, bytes.NewReader(body))
 332         if err != nil {
 333                 return 0, fmt.Errorf("create request: %w", err)
 334         }
 335
 336         request.Header.Set("Content-Type", contentType)
 337         response, err := a.http.Do(request)
 338         if err != nil {
 339                 return 0, fmt.Errorf("send request: %w", err)
 340         }
 341
 342         defer func() {
 343                 err = errors.Join(err, response.Body.Close())
 344         }()
 345
 346         if response.StatusCode != http.StatusOK {
 347                 return response.StatusCode, fmt.Errorf("HTTP %d", response.StatusCode)
 348         }
 349
 350         err = json.NewDecoder(response.Body).Decode(result)
 351         if err != nil {
 352                 return response.StatusCode, fmt.Errorf("decode answer: %w", err)
 353         }
 354
 355         return response.StatusCode, nil
 356 }
 357
 358 // ChatGPTDeviceCode is the answer of the user code endpoint. The interval is a
 359 // number or a string of seconds.
 360 type ChatGPTDeviceCode struct {
 361         DeviceAuthID string          `json:"device_auth_id"`
 362         UserCode     string          `json:"user_code"`
 363         UserCodeAlt  string          `json:"usercode"`
 364         Interval     json.RawMessage `json:"interval"`
 365 }
 366
 367 // ChatGPTTokenFile is the content of the token file.
 368 type ChatGPTTokenFile struct {
 369         AccessToken  string `json:"access_token"`
 370         RefreshToken string `json:"refresh_token"`
 371         IDToken      string `json:"id_token"`
 372         AccountID    string `json:"account_id"`
 373 }
 374
 375 // ChatGPTTokenResponse is the answer of the token endpoint.
 376 type ChatGPTTokenResponse struct {
 377         AccessToken  string `json:"access_token"`
 378         RefreshToken string `json:"refresh_token"`
 379         IDToken      string `json:"id_token"`
 380 }
 381
 382 // ChatGPTDeviceGrant is the answer of the device token endpoint after the user
 383 // approves the code.
 384 type ChatGPTDeviceGrant struct {
 385         AuthorizationCode string `json:"authorization_code"`
 386         CodeVerifier      string `json:"code_verifier"`
 387 }
 388
 389 // chatgptTokenExpiry returns the expiry of a JWT from its exp claim.
 390 func chatgptTokenExpiry(token string) (time.Time, bool) {
 391         var claims struct {
 392                 Exp *float64 `json:"exp"`
 393         }
 394
 395         if !chatgptDecodeClaims(token, &claims) || claims.Exp == nil {
 396                 return time.Time{}, false
 397         }
 398
 399         return time.Unix(int64(*claims.Exp), 0), true
 400 }
 401
 402 // chatgptPollInterval returns the poll interval of a device code, given in
 403 // seconds as a number or a string.
 404 func chatgptPollInterval(raw json.RawMessage) time.Duration {
 405         seconds, err := strconv.ParseFloat(strings.Trim(string(raw), `"`), 64)
 406         if err != nil || seconds < 0 {
 407                 return chatgptDefaultPollInterval
 408         }
 409
 410         return time.Duration(seconds * float64(time.Second))
 411 }
 412
 413 // chatgptAccountID returns the ChatGPT account ID of a JWT, or "" when it has
 414 // none.
 415 func chatgptAccountID(token string) string {
 416         var claims map[string]json.RawMessage
 417         if !chatgptDecodeClaims(token, &claims) {
 418                 return ""
 419         }
 420
 421         var auth struct {
 422                 AccountID string `json:"chatgpt_account_id"`
 423         }
 424
 425         err := json.Unmarshal(claims[chatgptAuthClaim], &auth)
 426         if err != nil {
 427                 return ""
 428         }
 429
 430         return auth.AccountID
 431 }
 432
 433 // chatgptDecodeClaims decodes the payload of a JWT into claims, without a check
 434 // of the signature. The token comes from the login service over HTTPS, and the
 435 // client reads only the expiry and the account ID from it.
 436 func chatgptDecodeClaims(token string, claims any) bool {
 437         parts := strings.Split(token, ".")
 438         if len(parts) != 3 {
 439                 return false
 440         }
 441
 442         payload, err := base64.RawURLEncoding.DecodeString(strings.TrimRight(parts[1], "="))
 443         if err != nil {
 444                 return false
 445         }
 446
 447         return json.Unmarshal(payload, claims) == nil
 448 }