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