git.giacomolli.uk - onechat/blob - model.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         // modelChatGPTURL is the Responses endpoint of the ChatGPT backend.
  22         modelChatGPTURL = "https://chatgpt.com/backend-api/codex/responses"
  23
  24         // modelAPIURL is the Responses endpoint of the OpenAI API.
  25         modelAPIURL = "https://api.openai.com/v1/responses"
  26
  27         // modelQuotaPrefix starts the names of the quota headers of the ChatGPT
  28         // backend, in canonical form.
  29         modelQuotaPrefix = "X-Codex-"
  30
  31         // modelTurnStateHeader is an x-codex-* header that holds an opaque
  32         // state of the backend, not quota.
  33         modelTurnStateHeader = "X-Codex-Turn-State"
  34
  35         // modelErrorBytes is the most bytes of an error response that go into
  36         // the error.
  37         modelErrorBytes = 4096
  38 )
  39
  40 // errModelCallLimit tells that a client sent its limit of calls.
  41 var errModelCallLimit = errors.New("call limit reached")
  42
  43 // modelEndpoint returns the URL and the authorizer of the endpoint in config.
  44 // The ChatGPT backend uses the token file in dir. The OpenAI API uses the key in
  45 // OPENAI_API_KEY, and there is no fallback from one to the other. All clients
  46 // share the authorizer, so one refresh runs at a time.
  47 func modelEndpoint(config Config, dir string) (string, ModelAuthorizer, error) {
  48         if config.Endpoint == configEndpointAPI {
  49                 key := os.Getenv("OPENAI_API_KEY")
  50                 if key == "" {
  51                         return "", nil, errors.New("endpoint api needs OPENAI_API_KEY")
  52                 }
  53
  54                 return modelAPIURL, ModelAPIKey(key), nil
  55         }
  56
  57         return modelChatGPTURL, newChatGPTAuth(filepath.Join(dir, chatgptTokenFileName), chatgptIssuer), nil
  58 }
  59
  60 // newModelClient returns a client for the Responses endpoint at url. It records
  61 // each call in the usage file at usagePath.
  62 func newModelClient(url string, auth ModelAuthorizer, cacheKey, usagePath string) *ModelClient {
  63         return &ModelClient{
  64                 url:      url,
  65                 auth:     auth,
  66                 cacheKey: cacheKey,
  67                 usage: &ModelUsageLog{
  68                         path: usagePath,
  69                 },
  70                 http: &http.Client{},
  71         }
  72 }
  73
  74 // modelUserMessage returns a user input item with one text part for each text.
  75 func modelUserMessage(texts ...string) any {
  76         type part struct {
  77                 Type string `json:"type"`
  78                 Text string `json:"text"`
  79         }
  80
  81         content := make([]part, len(texts))
  82         for n, text := range texts {
  83                 content[n] = part{
  84                         Type: "input_text",
  85                         Text: text,
  86                 }
  87         }
  88
  89         return struct {
  90                 Role    string `json:"role"`
  91                 Content []part `json:"content"`
  92         }{
  93                 Role:    "user",
  94                 Content: content,
  95         }
  96 }
  97
  98 // modelOutputText returns the text of the message items of output.
  99 func modelOutputText(output []json.RawMessage) (string, error) {
 100         var text strings.Builder
 101         for _, raw := range output {
 102                 var item ModelOutputItem
 103                 err := json.Unmarshal(raw, &item)
 104                 if err != nil {
 105                         return "", fmt.Errorf("decode output item: %w", err)
 106                 }
 107
 108                 if item.Type != "message" {
 109                         continue
 110                 }
 111
 112                 for _, part := range item.Content {
 113                         if part.Type == "output_text" {
 114                                 text.WriteString(part.Text)
 115                         }
 116                 }
 117         }
 118
 119         return text.String(), nil
 120 }
 121
 122 // ModelAuthorizer adds the credentials of a model call to the headers of its
 123 // request.
 124 type ModelAuthorizer interface {
 125         Authorize(ctx context.Context, header http.Header) error
 126 }
 127
 128 // ModelAPIKey authorizes calls to the OpenAI API with a key.
 129 type ModelAPIKey string
 130
 131 // Authorize sets the key as the bearer token.
 132 func (k ModelAPIKey) Authorize(ctx context.Context, header http.Header) error {
 133         header.Set("Authorization", "Bearer "+string(k))
 134         return ctx.Err()
 135 }
 136
 137 // ModelClient sends model calls to a Responses endpoint for one agent. The
 138 // ChatGPT backend and the OpenAI API take the same request body and differ only
 139 // in the URL and the credentials. Every call sends the fixed cache key of the
 140 // agent, so the turns and the compactions of the agent share one prompt cache.
 141 // The ChatGPT backend caches only with the key in the session_id header, so a
 142 // call to it sends the key there and in the body. Every call is recorded in the
 143 // usage file.
 144 //
 145 // The request never holds max_output_tokens or metadata, because the ChatGPT
 146 // backend refuses them. It sets store to false, so the client sends the whole
 147 // input on each call, and it asks for the encrypted reasoning, so a later call
 148 // can send the reasoning items back.
 149 type ModelClient struct {
 150         url      string
 151         auth     ModelAuthorizer
 152         cacheKey string
 153         usage    *ModelUsageLog
 154         http     *http.Client
 155         limit    int64
 156         calls    atomic.Int64
 157 }
 158
 159 // LimitCalls makes the client refuse each call after the first limit calls.
 160 func (c *ModelClient) LimitCalls(limit int64) {
 161         c.limit = limit
 162 }
 163
 164 // LimitReached tells if the client refuses further calls.
 165 func (c *ModelClient) LimitReached() bool {
 166         return c.limit > 0 && c.calls.Load() >= c.limit
 167 }
 168
 169 // Create sends the call and reads its event stream to the end. Over the call
 170 // limit it sends nothing and returns errModelCallLimit.
 171 func (c *ModelClient) Create(ctx context.Context, request ModelRequest) (ModelResponse, error) {
 172         if c.calls.Add(1) > c.limit && c.limit > 0 {
 173                 return ModelResponse{}, errModelCallLimit
 174         }
 175
 176         body, err := json.Marshal(ModelRequestBody{
 177                 Model:        request.Model,
 178                 Instructions: request.Instructions,
 179                 Input:        request.Input,
 180                 Tools:        request.Tools,
 181                 Reasoning: ModelReasoning{
 182                         Effort: request.Effort,
 183                 },
 184                 Store:          false,
 185                 Stream:         true,
 186                 Include:        []string{"reasoning.encrypted_content"},
 187                 PromptCacheKey: c.cacheKey,
 188         })
 189         if err != nil {
 190                 return ModelResponse{}, fmt.Errorf("encode request: %w", err)
 191         }
 192
 193         httpRequest, err := http.NewRequestWithContext(ctx, http.MethodPost, c.url, bytes.NewReader(body))
 194         if err != nil {
 195                 return ModelResponse{}, fmt.Errorf("create request: %w", err)
 196         }
 197
 198         httpRequest.Header.Set("Content-Type", "application/json")
 199         httpRequest.Header.Set("Accept", "text/event-stream")
 200         err = c.auth.Authorize(ctx, httpRequest.Header)
 201         if err != nil {
 202                 return ModelResponse{}, fmt.Errorf("authorize model call: %w", err)
 203         }
 204
 205         _, chatgpt := c.auth.(*ChatGPTAuth)
 206         if chatgpt {
 207                 httpRequest.Header.Set("session_id", c.cacheKey)
 208         }
 209
 210         start := time.Now()
 211         record := ModelUsageRecord{
 212                 Date:  start.Format(time.RFC3339Nano),
 213                 Model: request.Model,
 214         }
 215
 216         if request.Node != nil {
 217                 record.Node = &[2]int{request.Node.level, request.Node.index}
 218         }
 219
 220         record.Turn = request.Turn
 221         record.Agent = request.Agent
 222         response, err := c.send(httpRequest, &record)
 223         if err != nil {
 224                 record.Error = err.Error()
 225         }
 226
 227         record.Seconds = time.Since(start).Seconds()
 228         record.InputTokens = response.Usage.Input
 229         record.CachedTokens = response.Usage.Cached
 230         record.OutputTokens = response.Usage.Output
 231         record.TextBytes = len(response.Text)
 232
 233         return response, errors.Join(err, c.usage.Write(record))
 234 }
 235
 236 // send makes the HTTP call and reads its stream. It puts the status and the
 237 // quota headers into record.
 238 func (c *ModelClient) send(httpRequest *http.Request, record *ModelUsageRecord) (response ModelResponse, err error) {
 239         httpResponse, err := c.http.Do(httpRequest)
 240         if err != nil {
 241                 return ModelResponse{}, fmt.Errorf("send model call: %w", err)
 242         }
 243
 244         defer func() {
 245                 err = errors.Join(err, httpResponse.Body.Close())
 246         }()
 247
 248         record.Status = httpResponse.StatusCode
 249         record.Quota = modelQuotaHeaders(httpResponse.Header)
 250         if httpResponse.StatusCode != http.StatusOK {
 251                 text, readErr := io.ReadAll(io.LimitReader(httpResponse.Body, modelErrorBytes))
 252                 err = fmt.Errorf("model call: HTTP %d: %s", httpResponse.StatusCode, strings.TrimSpace(string(text)))
 253
 254                 return ModelResponse{}, errors.Join(err, readErr)
 255         }
 256
 257         return modelReadStream(httpResponse.Body)
 258 }
 259
 260 // ModelUsageLog appends one JSON line for each model call to usage.jsonl. Calls
 261 // can run at once, so a mutex serializes the writes.
 262 type ModelUsageLog struct {
 263         mutex sync.Mutex
 264         path  string
 265 }
 266
 267 // Write appends record to the file with one write.
 268 func (u *ModelUsageLog) Write(record ModelUsageRecord) (err error) {
 269         line, err := json.Marshal(record)
 270         if err != nil {
 271                 return fmt.Errorf("encode usage: %w", err)
 272         }
 273
 274         u.mutex.Lock()
 275         defer u.mutex.Unlock()
 276
 277         file, err := os.OpenFile(u.path, os.O_CREATE|os.O_WRONLY|os.O_APPEND, 0o600)
 278         if err != nil {
 279                 return fmt.Errorf("open usage file: %w", err)
 280         }
 281
 282         defer func() {
 283                 err = errors.Join(err, file.Close())
 284         }()
 285
 286         _, err = file.Write(append(line, '\n'))
 287         if err != nil {
 288                 return fmt.Errorf("write usage: %w", err)
 289         }
 290
 291         return nil
 292 }
 293
 294 // ModelOutputItem holds the fields of an output item that give its text.
 295 type ModelOutputItem struct {
 296         Type    string `json:"type"`
 297         Content []struct {
 298                 Type string `json:"type"`
 299                 Text string `json:"text"`
 300         } `json:"content"`
 301 }
 302
 303 // ModelRequest is one model call. Input holds the input items in order, such as
 304 // modelUserMessage values or output items of an earlier response. Node names
 305 // the node that a compaction builds, or is nil. Turn names the request of a
 306 // turn: first or next. Agent names the subagent that makes the call, or is
 307 // empty for a call of the agent itself. Node, Turn, and Agent go only to the
 308 // usage file, never to the endpoint.
 309 type ModelRequest struct {
 310         Model        string
 311         Effort       string
 312         Instructions string
 313         Tools        []ModelTool
 314         Input        []any
 315         Node         *NodeKey
 316         Turn         string
 317         Agent        string
 318 }
 319
 320 // ModelResponse is the result of a model call. Output holds the output items as
 321 // the endpoint sent them. Text is the text of the message items.
 322 type ModelResponse struct {
 323         Output []json.RawMessage
 324         Text   string
 325         Usage  ModelUsage
 326 }
 327
 328 // ModelRequestBody is the JSON body of a call.
 329 type ModelRequestBody struct {
 330         Model          string         `json:"model"`
 331         Instructions   string         `json:"instructions"`
 332         Input          []any          `json:"input"`
 333         Tools          []ModelTool    `json:"tools"`
 334         Reasoning      ModelReasoning `json:"reasoning"`
 335         Store          bool           `json:"store"`
 336         Stream         bool           `json:"stream"`
 337         Include        []string       `json:"include"`
 338         PromptCacheKey string         `json:"prompt_cache_key"`
 339 }
 340
 341 type ModelReasoning struct {
 342         Effort string `json:"effort"`
 343 }
 344
 345 // ModelUsageRecord is one line of usage.jsonl. Date is the start of the call,
 346 // and Seconds its length. Node is the node [l, i] of a compaction. Status is
 347 // the HTTP status, or 0 when no response came. TextBytes is the size of the
 348 // text of the reply. Turn names the request of a turn. Agent names the subagent
 349 // of the call. Quota holds the x-codex-* headers of the response, without the
 350 // turn state.
 351 type ModelUsageRecord struct {
 352         Date         string            `json:"date"`
 353         Seconds      float64           `json:"seconds"`
 354         Model        string            `json:"model"`
 355         Node         *[2]int           `json:"node,omitempty"`
 356         Turn         string            `json:"turn,omitempty"`
 357         Agent        string            `json:"agent,omitempty"`
 358         Status       int               `json:"status"`
 359         InputTokens  int               `json:"input_tokens"`
 360         CachedTokens int               `json:"cached_tokens"`
 361         OutputTokens int               `json:"output_tokens"`
 362         TextBytes    int               `json:"text_bytes"`
 363         Quota        map[string]string `json:"quota,omitempty"`
 364         Error        string            `json:"error,omitempty"`
 365 }
 366
 367 // ModelTool is the definition of a function tool.
 368 type ModelTool struct {
 369         Type        string `json:"type"`
 370         Name        string `json:"name"`
 371         Description string `json:"description"`
 372         Parameters  any    `json:"parameters"`
 373         Strict      bool   `json:"strict"`
 374 }
 375
 376 // ModelUsage gives the token counts of a call. Cached is the part of the input
 377 // tokens that came from the prompt cache.
 378 type ModelUsage struct {
 379         Input  int
 380         Cached int
 381         Output int
 382 }
 383
 384 // ModelStreamEvent holds the fields of the stream events that the client reads.
 385 type ModelStreamEvent struct {
 386         Type     string              `json:"type"`
 387         Item     json.RawMessage     `json:"item"`
 388         Response *ModelEventResponse `json:"response"`
 389         Code     string              `json:"code"`
 390         Message  string              `json:"message"`
 391 }
 392
 393 type ModelEventResponse struct {
 394         Output            []json.RawMessage `json:"output"`
 395         Usage             *ModelWireUsage   `json:"usage"`
 396         Error             *ModelWireError   `json:"error"`
 397         IncompleteDetails *struct {
 398                 Reason string `json:"reason"`
 399         } `json:"incomplete_details"`
 400 }
 401
 402 type ModelWireUsage struct {
 403         InputTokens        int `json:"input_tokens"`
 404         OutputTokens       int `json:"output_tokens"`
 405         InputTokensDetails struct {
 406                 CachedTokens int `json:"cached_tokens"`
 407         } `json:"input_tokens_details"`
 408 }
 409
 410 type ModelWireError struct {
 411         Code    string `json:"code"`
 412         Message string `json:"message"`
 413 }
 414
 415 // modelQuotaHeaders returns the x-codex-* headers, with lowercase names,
 416 // without the turn state.
 417 func modelQuotaHeaders(header http.Header) map[string]string {
 418         var quota map[string]string
 419         for name, values := range header {
 420                 if !strings.HasPrefix(name, modelQuotaPrefix) || name == modelTurnStateHeader || len(values) == 0 {
 421                         continue
 422                 }
 423
 424                 if quota == nil {
 425                         quota = make(map[string]string)
 426                 }
 427
 428                 quota[strings.ToLower(name)] = values[0]
 429         }
 430
 431         return quota
 432 }
 433
 434 // modelReadStream reads the server-sent events of a call until the response
 435 // completes. The output items come from the output_item.done events, or from
 436 // the completed response when the stream sent none.
 437 func modelReadStream(stream io.Reader) (ModelResponse, error) {
 438         var response ModelResponse
 439         reader := bufio.NewReader(stream)
 440         var data bytes.Buffer
 441         for {
 442                 line, err := reader.ReadBytes('\n')
 443                 if err != nil && !errors.Is(err, io.EOF) {
 444                         return response, fmt.Errorf("read event stream: %w", err)
 445                 }
 446
 447                 line = bytes.TrimRight(line, "\r\n")
 448                 if len(line) > 0 {
 449                         value, found := bytes.CutPrefix(line, []byte("data:"))
 450                         if found {
 451                                 data.Write(bytes.TrimPrefix(value, []byte(" ")))
 452                                 data.WriteByte('\n')
 453                         }
 454                 } else if data.Len() > 0 {
 455                         done, eventErr := modelReadEvent(&response, data.Bytes())
 456                         data.Reset()
 457                         if eventErr != nil || done {
 458                                 return response, eventErr
 459                         }
 460                 }
 461
 462                 if errors.Is(err, io.EOF) {
 463                         return response, errors.New("event stream ended before the response completed")
 464                 }
 465         }
 466 }
 467
 468 // modelReadEvent adds one stream event to response, and tells if it ends the
 469 // stream.
 470 func modelReadEvent(response *ModelResponse, data []byte) (bool, error) {
 471         var event ModelStreamEvent
 472         err := json.Unmarshal(data, &event)
 473         if err != nil {
 474                 return true, fmt.Errorf("decode stream event: %w", err)
 475         }
 476
 477         switch event.Type {
 478         case "response.output_item.done":
 479                 response.Output = append(response.Output, event.Item)
 480         case "response.completed":
 481                 if event.Response == nil {
 482                         return true, errors.New("completed event without a response")
 483                 }
 484
 485                 if len(response.Output) == 0 {
 486                         response.Output = event.Response.Output
 487                 }
 488
 489                 if event.Response.Usage != nil {
 490                         response.Usage = ModelUsage{
 491                                 Input:  event.Response.Usage.InputTokens,
 492                                 Cached: event.Response.Usage.InputTokensDetails.CachedTokens,
 493                                 Output: event.Response.Usage.OutputTokens,
 494                         }
 495                 }
 496
 497                 response.Text, err = modelOutputText(response.Output)
 498
 499                 return true, err
 500         case "response.incomplete":
 501                 reason := "unknown"
 502                 if event.Response != nil && event.Response.IncompleteDetails != nil {
 503                         reason = event.Response.IncompleteDetails.Reason
 504                 }
 505
 506                 return true, fmt.Errorf("response incomplete: %s", reason)
 507         case "response.failed":
 508                 message := "no details"
 509                 if event.Response != nil && event.Response.Error != nil {
 510                         message = event.Response.Error.Code + ": " + event.Response.Error.Message
 511                 }
 512
 513                 return true, fmt.Errorf("response failed: %s", message)
 514         case "error":
 515                 return true, fmt.Errorf("stream error: %s: %s", event.Code, event.Message)
 516         }
 517
 518         return false, nil
 519 }