1package provider
2
3import (
4 "bytes"
5 "context"
6 "crypto/rand"
7 "encoding/base64"
8 "encoding/hex"
9 "encoding/json"
10 "fmt"
11 "io"
12 "net/http"
13 "runtime"
14 "strings"
15 "time"
16)
17
18// Codex (ChatGPT subscription) client. Uses OpenAI's Responses API via
19// chatgpt.com/backend-api with the chatgpt-account-id handshake.
20//
21// Wire protocol notes:
22// - Endpoint: POST https://chatgpt.com/backend-api/codex/responses
23// - Headers: Authorization: Bearer <access_token>, chatgpt-account-id: <id>,
24// OpenAI-Beta: responses=experimental, originator: zot
25// - Body: OpenAI Responses API shape (not chat/completions).
26// input: [{role, content: [{type: "input_text" | "input_image" | ... }]}]
27// instructions: <system prompt>
28// tools: [{type:"function", name, description, parameters, strict}]
29// stream: true
30// - SSE events: response.output_item.added, response.output_text.delta,
31// response.function_call_arguments.delta, response.output_item.done,
32// response.completed, response.failed, error
33//
34// The access_token comes from the OpenAI OAuth flow (auth.openai.com).
35// The accountID is parsed from the id_token JWT's `chatgpt_account_id`
36// claim; see auth/oauth.go.
37
38const codexDefaultBaseURL = "https://chatgpt.com/backend-api/codex/responses"
39
40type codexClient struct {
41 token string
42 accountID string
43 baseURL string
44 errorLabel string
45 providerName string
46 modelName func(string) string
47 disableCLIRouting bool
48 http *http.Client
49}
50
51// NewOpenAICodex creates a client that talks to ChatGPT's Codex endpoint
52// using a subscription OAuth access token and the user's ChatGPT
53// account id. baseURL may be empty to use the default.
54func NewOpenAICodex(token, accountID, baseURL string) Client {
55 if baseURL == "" {
56 baseURL = codexDefaultBaseURL
57 }
58 return &codexClient{
59 token: token,
60 accountID: accountID,
61 baseURL: strings.TrimRight(baseURL, "/"),
62 errorLabel: "codex",
63 providerName: "openai-codex",
64 http: &http.Client{Timeout: 0},
65 }
66}
67
68func (c *codexClient) Name() string { return "openai-codex" }
69
70// ---- Responses API wire types (subset needed for zot's surface) ----
71
72type codexInputText struct {
73 Type string `json:"type"` // "input_text"
74 Text string `json:"text"`
75}
76
77type codexInputImage struct {
78 Type string `json:"type"` // "input_image"
79 Detail string `json:"detail"`
80 ImageURL string `json:"image_url"`
81}
82
83type codexOutputText struct {
84 Type string `json:"type"` // "output_text"
85 Text string `json:"text"`
86 Annotations []any `json:"annotations"`
87}
88
89type codexInputMessage struct {
90 Type string `json:"type,omitempty"` // "message" (optional for input)
91 Role string `json:"role"`
92 Content []any `json:"content"`
93}
94
95type codexOutputMessage struct {
96 Type string `json:"type"` // "message"
97 Role string `json:"role"`
98 Status string `json:"status,omitempty"`
99 ID string `json:"id,omitempty"`
100 Content []codexOutputText `json:"content"`
101}
102
103type codexFunctionCall struct {
104 Type string `json:"type"` // "function_call"
105 ID string `json:"id,omitempty"`
106 CallID string `json:"call_id"`
107 Name string `json:"name"`
108 Arguments string `json:"arguments"` // JSON string
109}
110
111type codexFunctionCallOutput struct {
112 Type string `json:"type"` // "function_call_output"
113 CallID string `json:"call_id"`
114 Output string `json:"output"` // string (or ResponseFunctionCallOutputItemList for images; v1 only uses string)
115}
116
117// codexReasoningItem mirrors the Responses API "reasoning" output item.
118// We capture it on incoming streams and replay it verbatim on follow-up
119// requests: the API rejects assistant tool-call replays without it when
120// thinking is enabled.
121type codexReasoningItem struct {
122 Type string `json:"type"` // "reasoning"
123 ID string `json:"id,omitempty"`
124 EncryptedContent string `json:"encrypted_content,omitempty"`
125 // Summary is required by the Responses API even when no summary text
126 // was streamed; encode an empty array rather than omitting the field.
127 Summary []codexReasoningSummary `json:"summary"`
128}
129
130type codexReasoningSummary struct {
131 Type string `json:"type"` // "summary_text"
132 Text string `json:"text"`
133}
134
135type codexTool struct {
136 Type string `json:"type"` // "function"
137 Name string `json:"name"`
138 Description string `json:"description,omitempty"`
139 Parameters json.RawMessage `json:"parameters"`
140}
141
142type codexReasoningConfig struct {
143 Effort string `json:"effort,omitempty"`
144}
145
146type codexRequest struct {
147 Model string `json:"model"`
148 Store bool `json:"store"`
149 Stream bool `json:"stream"`
150 Instructions string `json:"instructions,omitempty"`
151 Input []any `json:"input"`
152 Tools []codexTool `json:"tools,omitempty"`
153 ToolChoice string `json:"tool_choice,omitempty"`
154 ParallelToolCalls bool `json:"parallel_tool_calls"`
155 Include []string `json:"include,omitempty"`
156 Reasoning *codexReasoningConfig `json:"reasoning,omitempty"`
157 PromptCacheKey string `json:"prompt_cache_key,omitempty"`
158}
159
160// ---- Request building ----
161
162func (c *codexClient) buildRequest(req Request) (*codexRequest, error) {
163 m, err := FindModel("openai-codex", req.Model)
164 if err != nil {
165 m, err = FindModel("openai", req.Model)
166 }
167 if err != nil {
168 return nil, err
169 }
170 _ = m
171
172 body := &codexRequest{
173 Model: req.Model,
174 Store: false,
175 Stream: true,
176 Instructions: req.System,
177 ParallelToolCalls: true,
178 Include: []string{"reasoning.encrypted_content"},
179 }
180 if m.Reasoning {
181 if effort := OpenAICodexReasoningEffort(req.Reasoning, req.Model); effort != "" {
182 body.Reasoning = &codexReasoningConfig{Effort: effort}
183 }
184 }
185 if len(req.Tools) > 0 {
186 body.ToolChoice = "auto"
187 for _, t := range req.Tools {
188 params := t.Schema
189 if len(params) == 0 {
190 params = json.RawMessage(`{"type":"object","properties":{}}`)
191 }
192 body.Tools = append(body.Tools, codexTool{
193 Type: "function",
194 Name: t.Name,
195 Description: t.Description,
196 Parameters: params,
197 })
198 }
199 }
200
201 msgIdx := 0
202 req.Messages = RepairOrphanedToolResults(req.Messages)
203 for _, msg := range req.Messages {
204 switch msg.Role {
205 case RoleUser:
206 content := []any{}
207 for _, c := range msg.Content {
208 switch v := c.(type) {
209 case TextBlock:
210 if v.Text != "" {
211 content = append(content, codexInputText{Type: "input_text", Text: v.Text})
212 }
213 case ImageBlock:
214 url := "data:" + v.MimeType + ";base64," + base64.StdEncoding.EncodeToString(v.Data)
215 content = append(content, codexInputImage{Type: "input_image", Detail: "auto", ImageURL: url})
216 }
217 }
218 if len(content) == 0 {
219 continue
220 }
221 body.Input = append(body.Input, codexInputMessage{Role: "user", Content: content})
222 case RoleAssistant:
223 // Emit one output_message per text block, one function_call per
224 // tool call, and one reasoning item per ReasoningBlock,
225 // preserving the order so the model sees the same interleaving
226 // we captured. The reasoning replay is what keeps OpenAI
227 // Codex from rejecting follow-up tool calls with
228 // "thinking is enabled but reasoning_content is missing".
229 for _, c := range msg.Content {
230 switch v := c.(type) {
231 case ReasoningBlock:
232 item := codexReasoningItem{
233 Type: "reasoning",
234 ID: v.ID,
235 EncryptedContent: v.Encrypted,
236 Summary: []codexReasoningSummary{},
237 }
238 if v.Summary != "" {
239 item.Summary = []codexReasoningSummary{{Type: "summary_text", Text: v.Summary}}
240 }
241 body.Input = append(body.Input, item)
242 case TextBlock:
243 if v.Text == "" {
244 continue
245 }
246 msgIdx++
247 body.Input = append(body.Input, codexOutputMessage{
248 Type: "message",
249 Role: "assistant",
250 Status: "completed",
251 ID: fmt.Sprintf("msg_%d", msgIdx),
252 Content: []codexOutputText{
253 {Type: "output_text", Text: v.Text, Annotations: []any{}},
254 },
255 })
256 case ToolCallBlock:
257 args := string(v.Arguments)
258 if args == "" || !json.Valid([]byte(args)) {
259 args = "{}"
260 }
261 callID, _ := splitCallID(v.ID)
262 body.Input = append(body.Input, codexFunctionCall{
263 Type: "function_call",
264 CallID: callID,
265 Name: v.Name,
266 Arguments: args,
267 })
268 }
269 }
270 case RoleTool:
271 for _, c := range msg.Content {
272 if tr, ok := c.(ToolResultBlock); ok {
273 var text strings.Builder
274 imageCount := 0
275 for _, inner := range tr.Content {
276 switch v := inner.(type) {
277 case TextBlock:
278 if text.Len() > 0 {
279 text.WriteString("\n")
280 }
281 text.WriteString(v.Text)
282 case ImageBlock:
283 imageCount++
284 }
285 }
286 // The Responses API function_call_output only carries a
287 // string, so image bytes cannot ride along here. The agent
288 // loop mirrors any tool-result images into the following
289 // user message (which this client does serialize as
290 // input_image). Leave a short text note so an image-only
291 // result is not an empty output the API may reject, and so
292 // the model knows the image arrives next.
293 out := text.String()
294 if out == "" && imageCount > 0 {
295 if imageCount == 1 {
296 out = "[image returned; see the following message]"
297 } else {
298 out = fmt.Sprintf("[%d images returned; see the following message]", imageCount)
299 }
300 }
301 callID, _ := splitCallID(tr.CallID)
302 body.Input = append(body.Input, codexFunctionCallOutput{
303 Type: "function_call_output",
304 CallID: callID,
305 Output: out,
306 })
307 }
308 }
309 }
310 }
311
312 return body, nil
313}
314
315func splitCallID(id string) (string, string) {
316 if i := strings.Index(id, "|"); i >= 0 {
317 return id[:i], id[i+1:]
318 }
319 return id, ""
320}
321
322func usesCodexCLIRouting(model string) bool {
323 switch model {
324 case "gpt-5.6-luna", "gpt-5.6-luna-pro", "gpt-5.6-terra", "gpt-5.6-terra-pro":
325 return true
326 default:
327 return false
328 }
329}
330
331func newCodexSessionID() string {
332 var b [16]byte
333 if _, err := rand.Read(b[:]); err != nil {
334 return fmt.Sprintf("zot-%d", time.Now().UnixNano())
335 }
336 return strings.Join([]string{
337 hex.EncodeToString(b[0:4]),
338 hex.EncodeToString(b[4:6]),
339 hex.EncodeToString(b[6:8]),
340 hex.EncodeToString(b[8:10]),
341 hex.EncodeToString(b[10:16]),
342 }, "-")
343}
344
345// ---- Streaming ----
346
347func (c *codexClient) Stream(ctx context.Context, req Request) (<-chan Event, error) {
348 wire, err := c.buildRequest(req)
349 if err != nil {
350 return nil, err
351 }
352 if c.modelName != nil {
353 wire.Model = c.modelName(wire.Model)
354 }
355 var codexCLISessionID string
356 if !c.disableCLIRouting && usesCodexCLIRouting(wire.Model) {
357 codexCLISessionID = newCodexSessionID()
358 wire.PromptCacheKey = codexCLISessionID
359 }
360 body, err := json.Marshal(wire)
361 if err != nil {
362 return nil, err
363 }
364
365 newReq := func() (*http.Request, error) {
366 httpReq, err := http.NewRequestWithContext(ctx, "POST", c.baseURL, bytes.NewReader(body))
367 if err != nil {
368 return nil, err
369 }
370 httpReq.Header.Set("content-type", "application/json")
371 httpReq.Header.Set("accept", "text/event-stream")
372 httpReq.Header.Set("authorization", "Bearer "+c.token)
373 httpReq.Header.Set("chatgpt-account-id", c.accountID)
374 httpReq.Header.Set("openai-beta", "responses=experimental")
375 if codexCLISessionID != "" {
376 // Some preview models are only admitted or reliably served by the
377 // ChatGPT Codex backend when the request follows Codex CLI routing
378 // metadata. Keep this narrow so Sol retains zot's proven shape.
379 httpReq.Header.Set("originator", "codex_cli_rs")
380 httpReq.Header.Set("session-id", codexCLISessionID)
381 httpReq.Header.Set("user-agent", "codex_cli_rs/0.0.0")
382 } else {
383 httpReq.Header.Set("originator", "zot")
384 httpReq.Header.Set("user-agent", fmt.Sprintf("zot (%s %s)", runtime.GOOS, runtime.GOARCH))
385 }
386 return httpReq, nil
387 }
388
389 resp, err := doStreamWithRetry(ctx, c.http, newReq)
390 if err != nil {
391 return nil, fmt.Errorf("openai-codex: %w", err)
392 }
393 if resp.StatusCode != http.StatusOK {
394 b, _ := io.ReadAll(resp.Body)
395 resp.Body.Close()
396 return nil, fmt.Errorf("openai-codex: http %d: %s", resp.StatusCode, strings.TrimSpace(string(b)))
397 }
398
399 out := make(chan Event, 16)
400 go c.runStream(ctx, resp, req, out)
401 return out, nil
402}
403
404func (c *codexClient) runStream(ctx context.Context, resp *http.Response, req Request, out chan<- Event) {
405 defer close(out)
406 defer resp.Body.Close()
407
408 model, _ := FindModel("openai-codex", req.Model)
409 if model.ID == "" {
410 model, _ = FindModel("openai", req.Model)
411 }
412 providerName := c.providerName
413 if providerName == "" {
414 providerName = "openai-codex"
415 }
416 out <- EventStart{Model: req.Model, Provider: providerName}
417
418 raw := make(chan sseEvent, 16)
419 go readSSE(resp.Body, raw)
420
421 // Accumulators. The Responses API emits output_items in order; each
422 // item is either a "message" (text) or a "function_call". We track
423 // the in-flight item by its index.
424 type itemState struct {
425 kind string // "message" | "function_call" | "reasoning"
426 callID string
427 name string
428 argsBuf strings.Builder
429 textBuf strings.Builder
430 summary strings.Builder
431 rawID string
432 encrypted string
433 announced bool
434 }
435 var (
436 items = map[int]*itemState{}
437 order []int
438 usage Usage
439 stop StopReason = StopEnd
440 finalErr error
441 )
442
443 assemble := func() Message {
444 content := []Content{}
445 for _, idx := range order {
446 it := items[idx]
447 switch it.kind {
448 case "message":
449 if it.textBuf.Len() > 0 {
450 content = append(content, TextBlock{Text: it.textBuf.String()})
451 }
452 case "function_call":
453 args := it.argsBuf.String()
454 if args == "" || !json.Valid([]byte(args)) {
455 args = "{}"
456 }
457 content = append(content, ToolCallBlock{
458 ID: it.callID, Name: it.name, Arguments: json.RawMessage(args),
459 })
460 case "reasoning":
461 if it.encrypted == "" && it.summary.Len() == 0 && it.rawID == "" {
462 continue
463 }
464 content = append(content, ReasoningBlock{
465 ID: it.rawID,
466 Summary: it.summary.String(),
467 Encrypted: it.encrypted,
468 })
469 }
470 }
471 return Message{Role: RoleAssistant, Content: content, Time: time.Now()}
472 }
473
474 sendDone := func() {
475 usage.CostUSD = ComputeCost(model, usage)
476 out <- EventUsage{Usage: usage}
477 out <- EventDone{Stop: stop, Err: finalErr, Message: assemble()}
478 }
479
480 for {
481 select {
482 case <-ctx.Done():
483 stop = StopAborted
484 finalErr = ctx.Err()
485 sendDone()
486 return
487 case ev, ok := <-raw:
488 if !ok {
489 sendDone()
490 return
491 }
492 var head struct {
493 Type string `json:"type"`
494 }
495 if err := json.Unmarshal([]byte(ev.Data), &head); err != nil {
496 continue
497 }
498
499 switch head.Type {
500 case "response.output_item.added":
501 var p struct {
502 OutputIndex int `json:"output_index"`
503 Item struct {
504 Type string `json:"type"` // "message" | "function_call" | "reasoning"
505 ID string `json:"id"`
506 CallID string `json:"call_id"`
507 Name string `json:"name"`
508 EncryptedContent string `json:"encrypted_content"`
509 } `json:"item"`
510 }
511 _ = json.Unmarshal([]byte(ev.Data), &p)
512 it := &itemState{}
513 switch p.Item.Type {
514 case "message":
515 it.kind = "message"
516 case "function_call":
517 it.kind = "function_call"
518 it.callID = p.Item.CallID
519 it.name = p.Item.Name
520 if !it.announced {
521 it.announced = true
522 out <- EventToolStart{ID: it.callID, Name: it.name}
523 }
524 case "reasoning":
525 it.kind = "reasoning"
526 it.rawID = p.Item.ID
527 it.encrypted = p.Item.EncryptedContent
528 default:
529 continue
530 }
531 items[p.OutputIndex] = it
532 order = append(order, p.OutputIndex)
533 case "response.output_text.delta":
534 var p struct {
535 OutputIndex int `json:"output_index"`
536 Delta string `json:"delta"`
537 }
538 _ = json.Unmarshal([]byte(ev.Data), &p)
539 if it, ok := items[p.OutputIndex]; ok && it.kind == "message" {
540 it.textBuf.WriteString(p.Delta)
541 out <- EventTextDelta{Delta: p.Delta}
542 }
543 case "response.reasoning_summary_text.delta":
544 var p struct {
545 OutputIndex int `json:"output_index"`
546 Delta string `json:"delta"`
547 }
548 _ = json.Unmarshal([]byte(ev.Data), &p)
549 if it, ok := items[p.OutputIndex]; ok && it.kind == "reasoning" {
550 it.summary.WriteString(p.Delta)
551 }
552 case "response.reasoning_summary_text.done":
553 // summary text already accumulated via deltas
554 case "response.function_call_arguments.delta":
555 var p struct {
556 OutputIndex int `json:"output_index"`
557 Delta string `json:"delta"`
558 }
559 _ = json.Unmarshal([]byte(ev.Data), &p)
560 if it, ok := items[p.OutputIndex]; ok && it.kind == "function_call" {
561 it.argsBuf.WriteString(p.Delta)
562 out <- EventToolArgs{ID: it.callID, Delta: p.Delta}
563 }
564 case "response.output_item.done":
565 var p struct {
566 OutputIndex int `json:"output_index"`
567 Item struct {
568 Type string `json:"type"`
569 ID string `json:"id"`
570 EncryptedContent string `json:"encrypted_content"`
571 Summary []struct {
572 Type string `json:"type"`
573 Text string `json:"text"`
574 } `json:"summary"`
575 } `json:"item"`
576 }
577 _ = json.Unmarshal([]byte(ev.Data), &p)
578 if it, ok := items[p.OutputIndex]; ok {
579 switch it.kind {
580 case "function_call":
581 out <- EventToolEnd{ID: it.callID}
582 case "reasoning":
583 if p.Item.EncryptedContent != "" {
584 it.encrypted = p.Item.EncryptedContent
585 }
586 if it.rawID == "" && p.Item.ID != "" {
587 it.rawID = p.Item.ID
588 }
589 for _, s := range p.Item.Summary {
590 if s.Text == "" {
591 continue
592 }
593 if it.summary.Len() > 0 {
594 it.summary.WriteString("\n")
595 }
596 it.summary.WriteString(s.Text)
597 }
598 }
599 }
600 case "response.completed", "response.done":
601 var p struct {
602 Response struct {
603 Usage struct {
604 InputTokens int `json:"input_tokens"`
605 OutputTokens int `json:"output_tokens"`
606 InputTokensDetails struct {
607 CachedTokens int `json:"cached_tokens"`
608 } `json:"input_tokens_details"`
609 } `json:"usage"`
610 Status string `json:"status"`
611 } `json:"response"`
612 }
613 _ = json.Unmarshal([]byte(ev.Data), &p)
614 usage.InputTokens = p.Response.Usage.InputTokens - p.Response.Usage.InputTokensDetails.CachedTokens
615 if usage.InputTokens < 0 {
616 usage.InputTokens = p.Response.Usage.InputTokens
617 }
618 usage.OutputTokens = p.Response.Usage.OutputTokens
619 usage.CacheReadTokens = p.Response.Usage.InputTokensDetails.CachedTokens
620
621 hadTool := false
622 for _, it := range items {
623 if it.kind == "function_call" {
624 hadTool = true
625 break
626 }
627 }
628 if hadTool {
629 stop = StopToolUse
630 } else {
631 stop = StopEnd
632 }
633 sendDone()
634 return
635 case "response.failed":
636 var p struct {
637 Response struct {
638 Error struct {
639 Message string `json:"message"`
640 } `json:"error"`
641 } `json:"response"`
642 }
643 _ = json.Unmarshal([]byte(ev.Data), &p)
644 stop = StopError
645 finalErr = fmt.Errorf("codex: %s", p.Response.Error.Message)
646 sendDone()
647 return
648 case "error":
649 var p struct {
650 Message string `json:"message"`
651 Code string `json:"code"`
652 Error struct {
653 Message string `json:"message"`
654 Code string `json:"code"`
655 } `json:"error"`
656 }
657 _ = json.Unmarshal([]byte(ev.Data), &p)
658 msg := p.Message
659 if msg == "" {
660 msg = p.Error.Message
661 }
662 if msg == "" {
663 msg = p.Code
664 }
665 if msg == "" {
666 msg = p.Error.Code
667 }
668 if msg == "" {
669 msg = strings.TrimSpace(ev.Data)
670 }
671 stop = StopError
672 label := c.errorLabel
673 if label == "" {
674 label = "codex"
675 }
676 finalErr = fmt.Errorf("%s error: %s", label, msg)
677 sendDone()
678 return
679 }
680 }
681 }
682}