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

   1 package main
   2
   3 import (
   4         "bufio"
   5         "bytes"
   6         "context"
   7         "encoding/json"
   8         "errors"
   9         "fmt"
  10         "io"
  11         "net/http"
  12         "os"
  13         "path/filepath"
  14         "strings"
  15         "sync"
  16         "sync/atomic"
  17         "time"
  18 )
  19
  20 const (
  21         // chatgptURL is the Responses endpoint of the ChatGPT backend.
  22         chatgptURL = "https://chatgpt.com/backend-api/codex/responses"
  23
  24         // apiURL is the Responses endpoint of the OpenAI API.
  25         apiURL = "https://api.openai.com/v1/responses"
  26
  27         // quotaPrefix starts the names of the quota headers of the ChatGPT
  28         // backend, in canonical form.
  29         quotaPrefix = "X-Codex-"
  30
  31         // turnStateHeader is an x-codex-* header that holds an opaque state of
  32         // the backend, not quota.
  33         turnStateHeader = "X-Codex-Turn-State"
  34
  35         // errorBytes is the most bytes of an error response that go into the
  36         // error.
  37         errorBytes = 4096
  38 )
  39
  40 // errCallLimit tells that a client sent its limit of calls.
  41 var errCallLimit = errors.New("call limit reached")
  42
  43 // authorizer adds the credentials of a model call to the headers of its
  44 // request.
  45 type authorizer interface {
  46         Authorize(ctx context.Context, header http.Header) error
  47 }
  48
  49 // apiKey authorizes calls to the OpenAI API with a key.
  50 type apiKey string
  51
  52 // Client sends model calls to a Responses endpoint for one agent. The ChatGPT
  53 // backend and the OpenAI API take the same request body and differ only in the
  54 // URL and the credentials. Every call sends the fixed cache key of the agent,
  55 // so the turns and the compactions of the agent share one prompt cache. The
  56 // ChatGPT backend caches only with the key in the session_id header, so a call
  57 // to it sends the key there and in the body. Every call is recorded in the
  58 // usage file.
  59 //
  60 // The request never holds max_output_tokens or metadata, because the ChatGPT
  61 // backend refuses them. It sets store to false, so the client sends the whole
  62 // input on each call, and it asks for the encrypted reasoning, so a later call
  63 // can send the reasoning items back.
  64 type Client struct {
  65         url      string
  66         auth     authorizer
  67         cacheKey string
  68         usage    *usageLog
  69         http     *http.Client
  70         limit    int64
  71         calls    atomic.Int64
  72 }
  73
  74 // Request is one model call. Input holds the input items in order, such as
  75 // userMessage values or output items of an earlier response. Node names the
  76 // node that a compaction builds, or is nil. Turn names the request of a turn:
  77 // first or next. Node and Turn go only to the usage file, never to the
  78 // endpoint.
  79 type Request struct {
  80         Model        string
  81         Effort       string
  82         Instructions string
  83         Tools        []Tool
  84         Input        []any
  85         Node         *nodeKey
  86         Turn         string
  87 }
  88
  89 // Tool is the definition of a function tool.
  90 type Tool struct {
  91         Type        string `json:"type"`
  92         Name        string `json:"name"`
  93         Description string `json:"description"`
  94         Parameters  any    `json:"parameters"`
  95         Strict      bool   `json:"strict"`
  96 }
  97
  98 // Response is the result of a model call. Output holds the output items as the
  99 // endpoint sent them. Text is the text of the message items.
 100 type Response struct {
 101         Output []json.RawMessage
 102         Text   string
 103         Usage  Usage
 104 }
 105
 106 // Usage gives the token counts of a call. Cached is the part of the input
 107 // tokens that came from the prompt cache.
 108 type Usage struct {
 109         Input  int
 110         Cached int
 111         Output int
 112 }
 113
 114 // requestBody is the JSON body of a call.
 115 type requestBody struct {
 116         Model          string    `json:"model"`
 117         Instructions   string    `json:"instructions"`
 118         Input          []any     `json:"input"`
 119         Tools          []Tool    `json:"tools"`
 120         Reasoning      reasoning `json:"reasoning"`
 121         Store          bool      `json:"store"`
 122         Stream         bool      `json:"stream"`
 123         Include        []string  `json:"include"`
 124         PromptCacheKey string    `json:"prompt_cache_key"`
 125 }
 126
 127 type reasoning struct {
 128         Effort string `json:"effort"`
 129 }
 130
 131 // streamEvent holds the fields of the stream events that the client reads.
 132 type streamEvent struct {
 133         Type     string          `json:"type"`
 134         Item     json.RawMessage `json:"item"`
 135         Response *eventResponse  `json:"response"`
 136         Code     string          `json:"code"`
 137         Message  string          `json:"message"`
 138 }
 139
 140 type eventResponse struct {
 141         Output            []json.RawMessage `json:"output"`
 142         Usage             *wireUsage        `json:"usage"`
 143         Error             *wireError        `json:"error"`
 144         IncompleteDetails *struct {
 145                 Reason string `json:"reason"`
 146         } `json:"incomplete_details"`
 147 }
 148
 149 type wireUsage struct {
 150         InputTokens        int `json:"input_tokens"`
 151         OutputTokens       int `json:"output_tokens"`
 152         InputTokensDetails struct {
 153                 CachedTokens int `json:"cached_tokens"`
 154         } `json:"input_tokens_details"`
 155 }
 156
 157 type wireError struct {
 158         Code    string `json:"code"`
 159         Message string `json:"message"`
 160 }
 161
 162 // outputItem holds the fields of an output item that give its text.
 163 type outputItem struct {
 164         Type    string `json:"type"`
 165         Content []struct {
 166                 Type string `json:"type"`
 167                 Text string `json:"text"`
 168         } `json:"content"`
 169 }
 170
 171 // modelEndpoint returns the URL and the authorizer of the endpoint in config.
 172 // The ChatGPT backend uses the token file in dir. The OpenAI API uses the key in
 173 // OPENAI_API_KEY, and there is no fallback from one to the other. All clients
 174 // share the authorizer, so one refresh runs at a time.
 175 func modelEndpoint(config Config, dir string) (string, authorizer, error) {
 176         if config.Endpoint == endpointAPI {
 177                 key := os.Getenv("OPENAI_API_KEY")
 178                 if key == "" {
 179                         return "", nil, errors.New("endpoint api needs OPENAI_API_KEY")
 180                 }
 181
 182                 return apiURL, apiKey(key), nil
 183         }
 184
 185         return chatgptURL, newChatGPTAuth(filepath.Join(dir, tokenFileName), authIssuer), nil
 186 }
 187
 188 // newClient returns a client for the Responses endpoint at url. It records
 189 // each call in the usage file at usagePath.
 190 func newClient(url string, auth authorizer, cacheKey, usagePath string) *Client {
 191         return &Client{
 192                 url:      url,
 193                 auth:     auth,
 194                 cacheKey: cacheKey,
 195                 usage: &usageLog{
 196                         path: usagePath,
 197                 },
 198                 http: &http.Client{},
 199         }
 200 }
 201
 202 // LimitCalls makes the client refuse each call after the first limit calls.
 203 func (c *Client) LimitCalls(limit int64) {
 204         c.limit = limit
 205 }
 206
 207 // LimitReached tells if the client refuses further calls.
 208 func (c *Client) LimitReached() bool {
 209         return c.limit > 0 && c.calls.Load() >= c.limit
 210 }
 211
 212 // Create sends the call and reads its event stream to the end. Over the call
 213 // limit it sends nothing and returns errCallLimit.
 214 func (c *Client) Create(ctx context.Context, request Request) (Response, error) {
 215         if c.calls.Add(1) > c.limit && c.limit > 0 {
 216                 return Response{}, errCallLimit
 217         }
 218
 219         body, err := json.Marshal(requestBody{
 220                 Model:        request.Model,
 221                 Instructions: request.Instructions,
 222                 Input:        request.Input,
 223                 Tools:        request.Tools,
 224                 Reasoning: reasoning{
 225                         Effort: request.Effort,
 226                 },
 227                 Store:          false,
 228                 Stream:         true,
 229                 Include:        []string{"reasoning.encrypted_content"},
 230                 PromptCacheKey: c.cacheKey,
 231         })
 232         if err != nil {
 233                 return Response{}, fmt.Errorf("encode request: %w", err)
 234         }
 235
 236         httpRequest, err := http.NewRequestWithContext(ctx, http.MethodPost, c.url, bytes.NewReader(body))
 237         if err != nil {
 238                 return Response{}, fmt.Errorf("create request: %w", err)
 239         }
 240
 241         httpRequest.Header.Set("Content-Type", "application/json")
 242         httpRequest.Header.Set("Accept", "text/event-stream")
 243         err = c.auth.Authorize(ctx, httpRequest.Header)
 244         if err != nil {
 245                 return Response{}, fmt.Errorf("authorize model call: %w", err)
 246         }
 247
 248         _, chatgpt := c.auth.(*chatgptAuth)
 249         if chatgpt {
 250                 httpRequest.Header.Set("session_id", c.cacheKey)
 251         }
 252
 253         start := time.Now()
 254         record := usageRecord{
 255                 Date:  start.Format(time.RFC3339Nano),
 256                 Model: request.Model,
 257         }
 258
 259         if request.Node != nil {
 260                 record.Node = &[2]int{request.Node.level, request.Node.index}
 261         }
 262
 263         record.Turn = request.Turn
 264         response, err := c.send(httpRequest, &record)
 265         if err != nil {
 266                 record.Error = err.Error()
 267         }
 268
 269         record.Seconds = time.Since(start).Seconds()
 270         record.InputTokens = response.Usage.Input
 271         record.CachedTokens = response.Usage.Cached
 272         record.OutputTokens = response.Usage.Output
 273         record.TextBytes = len(response.Text)
 274
 275         return response, errors.Join(err, c.usage.Write(record))
 276 }
 277
 278 // send makes the HTTP call and reads its stream. It puts the status and the
 279 // quota headers into record.
 280 func (c *Client) send(httpRequest *http.Request, record *usageRecord) (response Response, err error) {
 281         httpResponse, err := c.http.Do(httpRequest)
 282         if err != nil {
 283                 return Response{}, fmt.Errorf("send model call: %w", err)
 284         }
 285
 286         defer func() {
 287                 err = errors.Join(err, httpResponse.Body.Close())
 288         }()
 289
 290         record.Status = httpResponse.StatusCode
 291         record.Quota = quotaHeaders(httpResponse.Header)
 292         if httpResponse.StatusCode != http.StatusOK {
 293                 text, readErr := io.ReadAll(io.LimitReader(httpResponse.Body, errorBytes))
 294                 err = fmt.Errorf("model call: HTTP %d: %s", httpResponse.StatusCode, strings.TrimSpace(string(text)))
 295
 296                 return Response{}, errors.Join(err, readErr)
 297         }
 298
 299         return readStream(httpResponse.Body)
 300 }
 301
 302 // usageLog appends one JSON line for each model call to usage.jsonl. Calls can
 303 // run at once, so a mutex serializes the writes.
 304 type usageLog struct {
 305         mutex sync.Mutex
 306         path  string
 307 }
 308
 309 // usageRecord is one line of usage.jsonl. Date is the start of the call, and
 310 // Seconds its length. Node is the node [l, i] of a compaction. Status is the
 311 // HTTP status, or 0 when no response came. TextBytes is the size of the text of
 312 // the reply. Turn names the request of a turn. Quota holds the x-codex-* headers of the response, without the
 313 // turn state.
 314 type usageRecord struct {
 315         Date         string            `json:"date"`
 316         Seconds      float64           `json:"seconds"`
 317         Model        string            `json:"model"`
 318         Node         *[2]int           `json:"node,omitempty"`
 319         Turn         string            `json:"turn,omitempty"`
 320         Status       int               `json:"status"`
 321         InputTokens  int               `json:"input_tokens"`
 322         CachedTokens int               `json:"cached_tokens"`
 323         OutputTokens int               `json:"output_tokens"`
 324         TextBytes    int               `json:"text_bytes"`
 325         Quota        map[string]string `json:"quota,omitempty"`
 326         Error        string            `json:"error,omitempty"`
 327 }
 328
 329 // Write appends record to the file with one write.
 330 func (u *usageLog) Write(record usageRecord) (err error) {
 331         line, err := json.Marshal(record)
 332         if err != nil {
 333                 return fmt.Errorf("encode usage: %w", err)
 334         }
 335
 336         u.mutex.Lock()
 337         defer u.mutex.Unlock()
 338
 339         file, err := os.OpenFile(u.path, os.O_CREATE|os.O_WRONLY|os.O_APPEND, 0o600)
 340         if err != nil {
 341                 return fmt.Errorf("open usage file: %w", err)
 342         }
 343
 344         defer func() {
 345                 err = errors.Join(err, file.Close())
 346         }()
 347
 348         _, err = file.Write(append(line, '\n'))
 349         if err != nil {
 350                 return fmt.Errorf("write usage: %w", err)
 351         }
 352
 353         return nil
 354 }
 355
 356 // Authorize sets the key as the bearer token.
 357 func (k apiKey) Authorize(ctx context.Context, header http.Header) error {
 358         header.Set("Authorization", "Bearer "+string(k))
 359         return ctx.Err()
 360 }
 361
 362 // userMessage returns a user input item with one text part for each text.
 363 func userMessage(texts ...string) any {
 364         type part struct {
 365                 Type string `json:"type"`
 366                 Text string `json:"text"`
 367         }
 368
 369         content := make([]part, len(texts))
 370         for n, text := range texts {
 371                 content[n] = part{
 372                         Type: "input_text",
 373                         Text: text,
 374                 }
 375         }
 376
 377         return struct {
 378                 Role    string `json:"role"`
 379                 Content []part `json:"content"`
 380         }{
 381                 Role:    "user",
 382                 Content: content,
 383         }
 384 }
 385
 386 // quotaHeaders returns the x-codex-* headers, with lowercase names, without the
 387 // turn state.
 388 func quotaHeaders(header http.Header) map[string]string {
 389         var quota map[string]string
 390         for name, values := range header {
 391                 if !strings.HasPrefix(name, quotaPrefix) || name == turnStateHeader || len(values) == 0 {
 392                         continue
 393                 }
 394
 395                 if quota == nil {
 396                         quota = make(map[string]string)
 397                 }
 398
 399                 quota[strings.ToLower(name)] = values[0]
 400         }
 401
 402         return quota
 403 }
 404
 405 // readStream reads the server-sent events of a call until the response
 406 // completes. The output items come from the output_item.done events, or from
 407 // the completed response when the stream sent none.
 408 func readStream(stream io.Reader) (Response, error) {
 409         var response Response
 410         reader := bufio.NewReader(stream)
 411         var data bytes.Buffer
 412         for {
 413                 line, err := reader.ReadBytes('\n')
 414                 if err != nil && !errors.Is(err, io.EOF) {
 415                         return response, fmt.Errorf("read event stream: %w", err)
 416                 }
 417
 418                 line = bytes.TrimRight(line, "\r\n")
 419                 if len(line) > 0 {
 420                         value, found := bytes.CutPrefix(line, []byte("data:"))
 421                         if found {
 422                                 data.Write(bytes.TrimPrefix(value, []byte(" ")))
 423                                 data.WriteByte('\n')
 424                         }
 425                 } else if data.Len() > 0 {
 426                         done, eventErr := readEvent(&response, data.Bytes())
 427                         data.Reset()
 428                         if eventErr != nil || done {
 429                                 return response, eventErr
 430                         }
 431                 }
 432
 433                 if errors.Is(err, io.EOF) {
 434                         return response, errors.New("event stream ended before the response completed")
 435                 }
 436         }
 437 }
 438
 439 // readEvent adds one stream event to response, and tells if it ends the
 440 // stream.
 441 func readEvent(response *Response, data []byte) (bool, error) {
 442         var event streamEvent
 443         err := json.Unmarshal(data, &event)
 444         if err != nil {
 445                 return true, fmt.Errorf("decode stream event: %w", err)
 446         }
 447
 448         switch event.Type {
 449         case "response.output_item.done":
 450                 response.Output = append(response.Output, event.Item)
 451         case "response.completed":
 452                 if event.Response == nil {
 453                         return true, errors.New("completed event without a response")
 454                 }
 455
 456                 if len(response.Output) == 0 {
 457                         response.Output = event.Response.Output
 458                 }
 459
 460                 if event.Response.Usage != nil {
 461                         response.Usage = Usage{
 462                                 Input:  event.Response.Usage.InputTokens,
 463                                 Cached: event.Response.Usage.InputTokensDetails.CachedTokens,
 464                                 Output: event.Response.Usage.OutputTokens,
 465                         }
 466                 }
 467
 468                 response.Text, err = outputText(response.Output)
 469
 470                 return true, err
 471         case "response.incomplete":
 472                 reason := "unknown"
 473                 if event.Response != nil && event.Response.IncompleteDetails != nil {
 474                         reason = event.Response.IncompleteDetails.Reason
 475                 }
 476
 477                 return true, fmt.Errorf("response incomplete: %s", reason)
 478         case "response.failed":
 479                 message := "no details"
 480                 if event.Response != nil && event.Response.Error != nil {
 481                         message = event.Response.Error.Code + ": " + event.Response.Error.Message
 482                 }
 483
 484                 return true, fmt.Errorf("response failed: %s", message)
 485         case "error":
 486                 return true, fmt.Errorf("stream error: %s: %s", event.Code, event.Message)
 487         }
 488
 489         return false, nil
 490 }
 491
 492 // outputText returns the text of the message items of output.
 493 func outputText(output []json.RawMessage) (string, error) {
 494         var text strings.Builder
 495         for _, raw := range output {
 496                 var item outputItem
 497                 err := json.Unmarshal(raw, &item)
 498                 if err != nil {
 499                         return "", fmt.Errorf("decode output item: %w", err)
 500                 }
 501
 502                 if item.Type != "message" {
 503                         continue
 504                 }
 505
 506                 for _, part := range item.Content {
 507                         if part.Type == "output_text" {
 508                                 text.WriteString(part.Text)
 509                         }
 510                 }
 511         }
 512
 513         return text.String(), nil
 514 }