ultron-ai 1.0.2

Provider layer of packages/ai: message model, streaming events, every API implementation (Anthropic, OpenAI completions/responses/codex, Azure, Google, Vertex, Mistral, Bedrock, pi-messages, Claude Code CLI, Claude subscription), provider auth and every OAuth flow, images API, model catalog, user models.json, env API keys
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
//! OpenAI Codex (ChatGPT backend) Responses provider (`api/openai-codex-responses.ts`): the request body, headers and
//! account id from the ChatGPT JWT, the WebSocket transport ([`websocket`]: `auto`/`websocket`/`websocket-cached`,
//! with its session socket cache, continuation deltas, retries and the fall back to SSE), the SSE transport with its
//! own retry policy and friendly usage-limit errors, the Codex event mapping (`response.done` and friends), and
//! Codex service-tier pricing. The SSE body is sent zstd-compressed, as upstream does with `node:zlib`'s zstd.

mod websocket;

pub use websocket::{
    OpenAICodexWebSocketDebugStats, advance_openai_codex_websocket_clock, close_openai_codex_websocket_sessions,
    get_openai_codex_websocket_debug_stats, reset_openai_codex_websocket_debug_stats,
};

use crate::constrained::grammar_tool_input_properties;
use crate::http::{
    AbortSignal, Body, DEFAULT_MAX_RETRY_DELAY_MS, HeaderBuilder, HttpRequest, HttpResponse, Transport, cancelled,
    is_aborted,
};
use crate::jsutil::{Utf8Stream, latin1};
use crate::models::clamp_thinking_level;
use crate::openai_responses::{
    ConvertOptions, OPENAI_TOOL_CALL_PROVIDERS, ResponsesRun, ResponsesStream, TierPricing, ToolOptions,
    convert_responses_messages, convert_responses_tools,
};
use crate::options::{SimpleStreamOptions, StreamOptions, build_base_options, pi_user_agent};
use crate::provider::{ProviderError, REQUEST_WAS_ABORTED, next_chunk};
use crate::transcript::{
    declared_tools, initial_system_message, resolve_transcript, resolve_transcript_tools, system_message_text,
};
use crate::types::{
    AssistantMessage, AssistantMessageDiagnostic, CacheRetention, DiagnosticErrorInfo, EventSink, Mapped, Model,
    TranscriptContext,
};
use indexmap::IndexMap;
use jsrt::json_parse::parse_json;
use serde::Deserialize;
use serde_json::{Map, Value, json};
use std::time::Duration;

const DEFAULT_CODEX_BASE_URL: &str = "https://chatgpt.com/backend-api";
const JWT_CLAIM_PATH: &str = "https://api.openai.com/auth";
const BASE_DELAY_MS: f64 = 1000.0;
const CODEX_RESPONSE_STATUSES: &[&str] = &[
    "completed",
    "incomplete",
    "failed",
    "cancelled",
    "queued",
    "in_progress",
];

#[derive(Clone, Default, Deserialize)]
#[serde(rename_all = "camelCase")]
pub struct OpenAICodexResponsesOptions {
    #[serde(flatten)]
    pub base: StreamOptions,
    /// `"none" | "minimal" | "low" | "medium" | "high" | "xhigh" | "max"`.
    pub reasoning_effort: Option<String>,
    /// `"auto" | "concise" | "detailed" | "off" | "on"`; `null` means `auto`.
    pub reasoning_summary: Option<String>,
    pub service_tier: Option<String>,
    pub text_verbosity: Option<String>,
    pub tool_choice: Option<String>,
}

/// `extractAccountId`: `chatgpt_account_id` from the JWT's `https://api.openai.com/auth` claim.
pub fn extract_account_id(token: &str) -> Result<String, String> {
    let fail = || "Failed to extract accountId from token".to_string();
    let parts: Vec<&str> = token.split('.').collect();
    if parts.len() != 3 {
        return Err(fail());
    }
    let payload = parse_json(&latin1(&jsrt::base64::atob(parts[1]).ok_or_else(fail)?)).map_err(|_| fail())?;
    match payload.get(JWT_CLAIM_PATH).and_then(|c| c.get("chatgpt_account_id")) {
        Some(Value::String(id)) if !id.is_empty() => Ok(id.clone()),
        Some(value) if jsrt::value::truthy(Some(value)) => Ok(jsrt::value::string(Some(value))),
        _ => Err(fail()),
    }
}

pub fn resolve_codex_url(base_url: &str) -> String {
    let raw = if jsrt::string::trim(base_url).is_empty() {
        DEFAULT_CODEX_BASE_URL
    } else {
        base_url
    };
    let normalized = raw.trim_end_matches('/');
    if normalized.ends_with("/codex/responses") {
        normalized.to_string()
    } else if normalized.ends_with("/codex") {
        format!("{normalized}/responses")
    } else {
        format!("{normalized}/codex/responses")
    }
}

fn build_request_body(
    model: &Model,
    context: &TranscriptContext,
    options: &OpenAICodexResponsesOptions,
    cache_session_id: Option<&str>,
    grammar: &IndexMap<String, String>,
    now: i64,
) -> Result<Map<String, Value>, String> {
    let flag = |key: &str, default: bool| model.compat_bool(key).unwrap_or(default);
    let supports_additional_tools = flag("supportsAdditionalTools", false);
    let supports_tool_search = flag("supportsToolSearch", false);
    let tool_options = ToolOptions {
        null_strict: true,
        supports_strict_mode: flag("supportsStrictMode", true),
        supports_openai_grammar_tools: flag("supportsOpenAIGrammarTools", false),
        tool_search_result: false,
    };
    let tools = resolve_transcript_tools(&context.messages, supports_additional_tools || supports_tool_search);
    let input = convert_responses_messages(
        model,
        context,
        &ConvertOptions {
            tool_call_providers: OPENAI_TOOL_CALL_PROVIDERS,
            include_system_prompt: false,
            grammar,
            supports_mid_convo_system_messages: flag("supportsMidConvoSystemMessages", false),
            supports_additional_tools,
            supports_tool_search,
            tool_options,
        },
        now,
    )?;
    let instructions = initial_system_message(&context.messages)
        .map(system_message_text)
        .filter(|text| !text.is_empty())
        .unwrap_or_else(|| "You are a helpful assistant.".into());
    let mut body = Map::new();
    body.insert("model".into(), json!(model.id));
    body.insert("store".into(), json!(false));
    body.insert("stream".into(), json!(true));
    body.insert("instructions".into(), json!(instructions));
    body.insert("input".into(), Value::Array(input));
    body.insert(
        "text".into(),
        json!({"verbosity": options.text_verbosity.as_deref().filter(|v| !v.is_empty()).unwrap_or("low")}),
    );
    body.insert("include".into(), json!(["reasoning.encrypted_content"]));
    if let Some(session) = cache_session_id {
        body.insert("prompt_cache_key".into(), json!(session));
    }
    body.insert(
        "tool_choice".into(),
        json!(options.tool_choice.as_deref().unwrap_or("auto")),
    );
    body.insert("parallel_tool_calls".into(), json!(true));
    if let Some(temperature) = options.base.temperature {
        body.insert("temperature".into(), json!(temperature));
    }
    if let Some(tier) = &options.service_tier {
        body.insert("service_tier".into(), json!(tier));
    }
    if !tools.request_tools.is_empty() {
        body.insert(
            "tools".into(),
            Value::Array(convert_responses_tools(&tools.request_tools, tool_options)?),
        );
    }
    let summary = options.reasoning_summary.as_deref().unwrap_or("auto");
    if let Some(effort) = &options.reasoning_effort {
        let effort: Option<&str> = if effort == "none" {
            match model.thinking_level("off") {
                Mapped::Missing => Some("none"),
                Mapped::Unsupported => None,
                Mapped::Value(value) => Some(value),
            }
        } else {
            Some(model.mapped_or(effort))
        };
        if let Some(effort) = effort {
            body.insert("reasoning".into(), json!({"effort": effort, "summary": summary}));
        }
    } else if model.reasoning && model.thinking_level("off") != Mapped::Unsupported {
        let off = match model.thinking_level("off") {
            Mapped::Value(value) => value,
            _ => "none",
        };
        body.insert("reasoning".into(), json!({"effort": off}));
    }
    Ok(body)
}

/// `buildBaseCodexHeaders`.
fn base_headers(model: &Model, options: &StreamOptions, account_id: &str, token: &str) -> HeaderBuilder {
    let mut headers = HeaderBuilder::default();
    headers.extend_set(model.headers.iter().flatten());
    headers.extend(options.headers.iter().flatten());
    headers
        .set("Authorization", format!("Bearer {token}"))
        .set("chatgpt-account-id", account_id)
        .set("originator", "pi")
        .set("User-Agent", pi_user_agent());
    headers
}

/// `buildWebSocketHeaders`: the base headers with the WebSocket beta, and `request_id` as the session and request id.
fn websocket_headers(
    model: &Model,
    options: &StreamOptions,
    account_id: &str,
    token: &str,
    request_id: &str,
) -> HeaderBuilder {
    let mut headers = base_headers(model, options, account_id, token);
    for name in ["accept", "content-type", "openai-beta"] {
        headers.remove(name);
    }
    headers
        .set("OpenAI-Beta", websocket::OPENAI_BETA_RESPONSES_WEBSOCKETS)
        .set("x-client-request-id", request_id)
        .set("session-id", request_id);
    headers
}

fn sse_headers(
    model: &Model,
    options: &StreamOptions,
    account_id: &str,
    token: &str,
    session_id: Option<&str>,
) -> HeaderBuilder {
    let mut headers = base_headers(model, options, account_id, token);
    headers
        .set("OpenAI-Beta", "responses=experimental")
        .set("accept", "text/event-stream")
        .set("content-type", "application/json");
    if let Some(session_id) = session_id {
        headers
            .set("session-id", session_id)
            .set("x-client-request-id", session_id);
    }
    headers
}

fn normalize_timeout_ms(value: Option<f64>) -> Result<Option<u64>, String> {
    match value {
        None => Ok(None),
        Some(v) if !v.is_finite() || v < 0.0 => Err(format!("Invalid timeoutMs: {}", jsrt::number::to_string(v))),
        Some(v) => Ok(Some(v.floor() as u64)),
    }
}

/// Why a Codex request failed: the kind decides retries and the WebSocket fallback.
#[derive(Debug)]
enum CodexError {
    Plain(String),
    /// `RetryDelayExceededError`: never retried.
    RetryDelay(String),
    /// `CodexApiError`: an `error` or `response.failed` event, with its code.
    Api {
        message: String,
        code: Option<String>,
    },
    /// `CodexProtocolError`: an event that is not JSON.
    Protocol(String),
    /// `WebSocketCloseError`: the socket closed before the response completed.
    Closed {
        message: String,
        code: u16,
    },
    /// A non-2xx SSE response, until the retry loop decides on it.
    Status {
        status: u16,
        headers: crate::http::HeaderMap,
        text: String,
    },
}

impl ProviderError for CodexError {
    fn plain(message: String) -> Self {
        CodexError::Plain(message)
    }

    fn aborted() -> Self {
        aborted_error()
    }
}

impl CodexError {
    fn message(&self) -> &str {
        match self {
            CodexError::Plain(m) | CodexError::RetryDelay(m) | CodexError::Protocol(m) => m,
            CodexError::Status { text, .. } => text,
            CodexError::Api { message, .. } | CodexError::Closed { message, .. } => message,
        }
    }

    /// `isCodexNonTransportError`.
    fn is_non_transport(&self) -> bool {
        matches!(self, CodexError::Api { .. } | CodexError::Protocol(_))
    }

    fn api_code(&self) -> Option<&str> {
        match self {
            CodexError::Api { code, .. } => code.as_deref(),
            _ => None,
        }
    }

    /// `extractDiagnosticError`.
    fn diagnostic(&self) -> DiagnosticErrorInfo {
        let (name, code) = match self {
            CodexError::Plain(_) | CodexError::RetryDelay(_) | CodexError::Status { .. } => ("Error", None),
            CodexError::Api { code, .. } => ("CodexApiError", code.as_ref().map(|c| json!(c))),
            CodexError::Protocol(_) => ("CodexProtocolError", None),
            CodexError::Closed { code, .. } => ("WebSocketCloseError", Some(json!(code))),
        };
        DiagnosticErrorInfo {
            name: Some(name.into()),
            message: self.message().to_string(),
            stack: None,
            code,
        }
    }
}

fn aborted_error() -> CodexError {
    CodexError::Plain(REQUEST_WAS_ABORTED.into())
}

fn is_terminal_rate_limit_error(text: &str) -> bool {
    let lower = text.to_lowercase();
    [
        "gousagelimiterror",
        "freeusagelimiterror",
        "monthly usage limit reached",
        "available balance",
        "insufficient_quota",
        "out of budget",
        "quota exceeded",
        "billing",
    ]
    .iter()
    .any(|needle| lower.contains(needle))
}

fn is_retryable_error(status: u16, text: &str) -> bool {
    if status == 429 && is_terminal_rate_limit_error(text) {
        return false;
    }
    if matches!(status, 429 | 500 | 502 | 503 | 504) {
        return true;
    }
    crate::jsutil::loose_match(text, "rate", "limit")
        || text.to_lowercase().contains("overloaded")
        || crate::jsutil::loose_match(text, "service", "unavailable")
        || crate::jsutil::loose_match(text, "upstream", "connect")
        || crate::jsutil::loose_match(text, "connection", "refused")
}

/// `getRetryAfterDelayMs`: `retry-after-ms`, else `retry-after` in seconds, else as an HTTP date (the time until
/// then, at least 0).
fn retry_after_delay_ms(headers: &crate::http::HeaderMap) -> Option<f64> {
    if let Some(value) = headers.get("retry-after-ms")
        && let Some(millis) = Some(jsrt::number::parse(value)).filter(|m| m.is_finite())
    {
        return Some(millis.max(0.0));
    }
    let retry_after = headers.get("retry-after").filter(|v| !v.is_empty())?;
    if let Some(seconds) = Some(jsrt::number::parse(retry_after)).filter(|s| s.is_finite()) {
        return Some((seconds * 1000.0).max(0.0));
    }
    crate::http::until_http_date_ms(retry_after).map(|ms| ms.max(0.0))
}

fn validate_retry_delay_ms(delay_ms: f64, max: Option<f64>) -> Result<f64, CodexError> {
    let max = max.unwrap_or(DEFAULT_MAX_RETRY_DELAY_MS);
    if max > 0.0 && delay_ms > max {
        return Err(CodexError::RetryDelay(format!(
            "Server requested {}s retry delay (max: {}s)",
            (delay_ms / 1000.0).ceil(),
            (max / 1000.0).ceil()
        )));
    }
    Ok(delay_ms)
}

fn status_text(status: u16) -> String {
    reqwest::StatusCode::from_u16(status)
        .ok()
        .and_then(|s| s.canonical_reason())
        .unwrap_or_default()
        .to_string()
}

/// `parseErrorResponse`: the error message, and a friendly one for ChatGPT usage limits.
pub fn parse_error_response(status: u16, raw: &str, now: i64) -> (String, Option<String>) {
    let mut message = if !raw.is_empty() {
        raw.to_string()
    } else {
        let text = status_text(status);
        if text.is_empty() { "Request failed".into() } else { text }
    };
    let mut friendly = None;
    if let Ok(parsed) = parse_json(raw)
        && let Some(err) = parsed.get("error").filter(|e| jsrt::value::truthy(Some(e)))
    {
        let code = [err.get("code"), err.get("type")]
            .into_iter()
            .find(|v| jsrt::value::truthy(*v))
            .flatten()
            .map(|v| jsrt::value::string(Some(v)))
            .unwrap_or_default();
        let lower = code.to_lowercase();
        if lower.contains("usage_limit_reached")
            || lower.contains("usage_not_included")
            || lower.contains("rate_limit_exceeded")
            || status == 429
        {
            let plan = match err.get("plan_type").filter(|p| jsrt::value::truthy(Some(p))) {
                Some(plan) => format!(" ({} plan)", jsrt::value::string(Some(plan)).to_lowercase()),
                None => String::new(),
            };
            let when = match err
                .get("resets_at")
                .filter(|r| jsrt::value::truthy(Some(r)))
                .and_then(Value::as_f64)
            {
                Some(resets_at) => {
                    let minutes = ((resets_at * 1000.0 - now as f64) / 60000.0 + 0.5).floor().max(0.0);
                    format!(" Try again in ~{} min.", jsrt::number::to_string(minutes))
                }
                None => String::new(),
            };
            friendly =
                Some(jsrt::string::trim(&format!("You have hit your ChatGPT usage limit{plan}.{when}")).to_string());
        }
        if let Some(text) = err.get("message").filter(|m| jsrt::value::truthy(Some(m))) {
            message = jsrt::value::string(Some(text));
        } else if let Some(friendly) = &friendly {
            message = friendly.clone();
        }
    }
    (message, friendly)
}

/// `parseSSE`: frames end at a blank line (`\n\n`); `data:` lines are trimmed and joined; `[DONE]` is skipped.
#[derive(Default)]
struct CodexSse {
    decoder: Utf8Stream,
    buffer: String,
}

impl CodexSse {
    fn frames(&mut self, text: &str, done: bool) -> Result<Vec<Value>, CodexError> {
        self.buffer.push_str(text);
        if done && !jsrt::string::trim(&self.buffer).is_empty() {
            self.buffer.push_str("\n\n");
        }
        let mut out = Vec::new();
        while let Some(index) = self.buffer.find("\n\n") {
            let chunk: String = self.buffer[..index].to_string();
            self.buffer.drain(..index + 2);
            let lines: Vec<&str> = chunk
                .split('\n')
                .filter(|l| l.starts_with("data:"))
                .map(|l| jsrt::string::trim(&l[5..]))
                .collect();
            if lines.is_empty() {
                continue;
            }
            let data = lines.join("\n");
            let data = jsrt::string::trim(&data);
            if !data.is_empty() && data != "[DONE]" {
                let value = parse_json(data).map_err(|e| {
                    let message = jsrt::json::parse(data).err().unwrap_or_else(|| e.to_string());
                    CodexError::Protocol(format!("Invalid Codex SSE JSON: {message}"))
                })?;
                out.push(value);
            }
        }
        Ok(out)
    }
}

/// `mapCodexEvents` for one event: `Ok(None)` skips it, a terminal event is returned as `response.completed`.
fn map_codex_event(event: Value, output: &mut AssistantMessage) -> Result<Option<(Value, bool)>, CodexError> {
    let Some(kind) = event.get("type").and_then(Value::as_str).filter(|t| !t.is_empty()) else {
        return Ok(None);
    };
    let field = |value: &Value, key: &str| value.get(key).and_then(Value::as_str).map(str::to_string);
    match kind {
        "error" => {
            let nested = event.get("error").filter(|e| e.is_object());
            let code = field(&event, "code").or_else(|| nested.and_then(|n| field(n, "code")));
            let message = field(&event, "message").or_else(|| nested.and_then(|n| field(n, "message")));
            let text = message
                .filter(|m| !m.is_empty())
                .or(code.clone().filter(|c| !c.is_empty()))
                .unwrap_or_else(|| jsrt::json::stringify(&event));
            Err(CodexError::Api {
                message: format!("Codex error: {text}"),
                code,
            })
        }
        "response.failed" => {
            let message = event
                .pointer("/response/error/message")
                .and_then(Value::as_str)
                .filter(|m| !m.is_empty())
                .unwrap_or("Codex response failed");
            Err(CodexError::Api {
                message: message.to_string(),
                code: event
                    .pointer("/response/error/code")
                    .and_then(Value::as_str)
                    .map(str::to_string),
            })
        }
        "response.done" | "response.completed" | "response.incomplete" => {
            let mut event = event;
            if let Some(end_turn) = event.pointer("/response/end_turn").and_then(Value::as_bool) {
                output.end_turn = Some(end_turn);
            }
            if let Some(response) = event.get_mut("response").and_then(Value::as_object_mut) {
                let status = response
                    .get("status")
                    .and_then(Value::as_str)
                    .filter(|s| CODEX_RESPONSE_STATUSES.contains(s))
                    .map(str::to_string);
                match status {
                    Some(status) => response.insert("status".into(), json!(status)),
                    None => response.shift_remove("status"),
                };
            }
            event["type"] = json!("response.completed");
            Ok(Some((event, true)))
        }
        _ => Ok(Some((event, false))),
    }
}

async fn process_stream(
    run: &mut ResponsesRun<'_, '_>,
    response: &mut HttpResponse,
    model: &Model,
    grammar: &IndexMap<String, String>,
    service_tier: Option<&str>,
    signal: Option<&AbortSignal>,
) -> Result<(), CodexError> {
    let mut sse = CodexSse::default();
    let body: &mut dyn Body = response.body.as_mut();
    'read: loop {
        let (text, done) = match next_chunk::<CodexError>(body, signal).await? {
            Some(bytes) => (sse.decoder.push(&bytes), false),
            None => (sse.decoder.finish(), true),
        };
        if is_aborted(signal) {
            return Err(aborted_error());
        }
        for event in sse.frames(&text, done)? {
            let Some((event, terminal)) = map_codex_event(event, &mut run.output)? else {
                continue;
            };
            run.handle(&event, model, grammar, TierPricing::Codex(service_tier))
                .map_err(CodexError::Plain)?;
            if terminal {
                break 'read;
            }
        }
        if done {
            break;
        }
    }
    run.terminated().map_err(CodexError::Plain)
}

/// Stream one Codex response; failures end with an `error` event.
pub async fn stream(
    model: &Model,
    context: &TranscriptContext,
    options: &OpenAICodexResponsesOptions,
    transport: &dyn Transport,
    sink: &mut EventSink<'_>,
) -> AssistantMessage {
    let context = resolve_transcript(
        context,
        model.compat_bool("supportsMidConvoSystemMessages").unwrap_or(false),
    );
    let now = jsrt::date::now_ms();
    let mut output = AssistantMessage::pending(model, now);
    output.api = "openai-codex-responses".into();
    let mut run = ResponsesRun::new(output, sink, ResponsesStream::default());
    let result = drive(&mut run, model, &context, options, transport, now).await;
    run.end(
        result.map_err(|error| error.message().to_string()),
        options.base.signal(),
    )
}

async fn drive(
    run: &mut ResponsesRun<'_, '_>,
    model: &Model,
    context: &TranscriptContext,
    options: &OpenAICodexResponsesOptions,
    transport: &dyn Transport,
    now: i64,
) -> Result<(), CodexError> {
    let base = &options.base;
    let signal = base.signal();
    let api_key = base
        .api_key
        .clone()
        .filter(|k| !k.is_empty())
        .ok_or_else(|| CodexError::Plain(format!("No API key for provider: {}", model.provider)))?;
    let account_id = extract_account_id(&api_key).map_err(CodexError::Plain)?;
    let grammar = grammar_tool_input_properties(
        &declared_tools(&context.messages),
        model.compat_bool("supportsOpenAIGrammarTools").unwrap_or(false),
    )
    .map_err(CodexError::Plain)?;
    let cache_session_id = if base.cache_retention == Some(CacheRetention::None) {
        None
    } else {
        base.session_id.as_deref()
    };
    let codex_session_id = cache_session_id.map(crate::openai_responses::clamp_prompt_cache_key);
    let mut body = Value::Object(
        build_request_body(model, context, options, codex_session_id.as_deref(), &grammar, now)
            .map_err(CodexError::Plain)?,
    );
    if let Some(next) = base.on_payload.as_ref().and_then(|hook| hook(&body, model)) {
        body = next;
    }
    let websocket_request_id = codex_session_id
        .clone()
        .filter(|id| !id.is_empty())
        .unwrap_or_else(|| jsrt::random::uuid_v7(jsrt::date::now_ms()));
    let headers = sse_headers(model, base, &account_id, &api_key, codex_session_id.as_deref()).build();
    let ws_headers = websocket_headers(model, base, &account_id, &api_key, &websocket_request_id).build();
    let body_json = serde_json::to_string(&body).unwrap_or_default();
    let http_timeout_ms = normalize_timeout_ms(base.timeout_ms.map(|t| t as f64)).map_err(CodexError::Plain)?;
    let connect_timeout_ms = normalize_timeout_ms(base.websocket_connect_timeout_ms).map_err(CodexError::Plain)?;
    let configured = base.transport.as_deref().filter(|t| !t.is_empty()).unwrap_or("auto");
    let mut start_emitted = false;
    let websocket_disabled = configured != "sse" && websocket::is_websocket_sse_fallback_active(cache_session_id);
    if websocket_disabled {
        websocket::record_websocket_sse_fallback(cache_session_id);
    }
    if configured != "sse" && !websocket_disabled {
        let url = websocket::resolve_codex_websocket_url(&model.base_url);
        let request = websocket::Request {
            url: &url,
            body: &body,
            headers: &ws_headers,
            model,
            grammar: &grammar,
            service_tier: options.service_tier.as_deref(),
            idle_timeout_ms: http_timeout_ms,
            connect_timeout_ms,
            cache_session_id,
            account_id: &account_id,
            // The continuation is used only where the caller asked for `auto` or `websocket-cached` (upstream reads
            // `options.transport`, not its default).
            transport_kind: base.transport.as_deref().unwrap_or_default(),
            signal,
        };
        let mut retried_connection_limit = false;
        let mut retried_missing_continuation = false;
        loop {
            let mut started = false;
            let attempt =
                websocket::process_websocket_stream(run, transport, &request, &mut started, &mut start_emitted).await;
            let error = match attempt {
                Ok(()) => return run.stopped(signal, CODEX_NO_STOP_REASON).map_err(CodexError::Plain),
                Err(error) => error,
            };
            let aborted = is_aborted(signal);
            let connection_limit_before_start =
                !started && error.api_code() == Some(websocket::WEBSOCKET_CONNECTION_LIMIT_REACHED_CODE);
            let previous_response_not_found = error.api_code() == Some(websocket::PREVIOUS_RESPONSE_NOT_FOUND_CODE);
            if !aborted && previous_response_not_found && !retried_missing_continuation {
                retried_missing_continuation = true;
                continue;
            }
            if !aborted && connection_limit_before_start && !retried_connection_limit {
                retried_connection_limit = true;
                continue;
            }
            if aborted || (error.is_non_transport() && !connection_limit_before_start) {
                return Err(error);
            }
            let mut details = Map::new();
            details.insert("configuredTransport".into(), json!(configured));
            if !started {
                details.insert("fallbackTransport".into(), json!("sse"));
            }
            details.insert("eventsEmitted".into(), json!(started));
            details.insert(
                "phase".into(),
                json!(if started {
                    "after_message_stream_start"
                } else {
                    "before_message_stream_start"
                }),
            );
            details.insert("requestBytes".into(), json!(body_json.len()));
            run.output
                .diagnostics
                .get_or_insert_with(Vec::new)
                .push(AssistantMessageDiagnostic {
                    kind: "provider_transport_failure".into(),
                    timestamp: jsrt::date::now_ms(),
                    error: Some(error.diagnostic()),
                    details: Some(details),
                });
            websocket::record_websocket_failure(cache_session_id, error.message());
            if started {
                return Err(error);
            }
            websocket::record_websocket_sse_fallback(cache_session_id);
            break;
        }
    }
    // The SSE body goes out zstd-compressed (`compressRequestBodyZstd`; the backend decodes `Content-Encoding:
    // zstd`); the WebSocket frame above stays uncompressed, as the official Codex client sends it. The request
    // carries the JSON and names the encoding; the transport encodes the bytes it sends (crate::http::wire_body).
    let mut headers = headers;
    headers.insert("content-encoding".into(), "zstd".into());
    let request = HttpRequest {
        method: "POST".into(),
        url: resolve_codex_url(&model.base_url),
        headers,
        body: body_json,
    };
    let mut response = send_with_codex_retry(transport, &request, base, http_timeout_ms, signal).await?;
    if !start_emitted {
        run.start();
    }
    process_stream(
        run,
        &mut response,
        model,
        &grammar,
        options.service_tier.as_deref(),
        signal,
    )
    .await?;
    run.stopped(signal, CODEX_NO_STOP_REASON).map_err(CodexError::Plain)
}

/// `assertSuccessfulOutput`'s message for a stream that set no stop reason.
const CODEX_NO_STOP_REASON: &str = "Codex stream ended without a stop reason";

async fn read_text(response: &mut HttpResponse) -> String {
    let mut decoder = Utf8Stream::default();
    let mut text = String::new();
    while let Ok(Some(chunk)) = response.body.next_chunk().await {
        text.push_str(&decoder.push(&chunk));
    }
    text.push_str(&decoder.finish());
    text
}

/// One `fetch` with the response-header timeout.
async fn send_once(
    transport: &dyn Transport,
    request: &HttpRequest,
    timeout_ms: Option<u64>,
    signal: Option<&AbortSignal>,
) -> Result<HttpResponse, CodexError> {
    let timeout = async {
        match timeout_ms.filter(|t| *t > 0) {
            Some(ms) => tokio::time::sleep(Duration::from_millis(ms)).await,
            None => std::future::pending().await,
        }
    };
    tokio::select! {
        sent = transport.send(request.clone()) => sent.map_err(CodexError::Plain),
        () = cancelled(signal) => Err(aborted_error()),
        () = timeout => Err(CodexError::Plain(format!(
            "Codex SSE response headers timed out after {}ms",
            timeout_ms.unwrap_or_default()
        ))),
    }
}

/// The SSE request with Codex's retry loop (`maxRetries` default 0, 1 s doubling, `retry-after` honored).
async fn send_with_codex_retry(
    transport: &dyn Transport,
    request: &HttpRequest,
    base: &StreamOptions,
    timeout_ms: Option<u64>,
    signal: Option<&AbortSignal>,
) -> Result<HttpResponse, CodexError> {
    let max_retries = base.max_retries.unwrap_or(0);
    let send = |_| async {
        if is_aborted(signal) {
            return Err(aborted_error());
        }
        let mut response = send_once(transport, request, timeout_ms, signal).await?;
        if (200..300).contains(&response.status) {
            return Ok(response);
        }
        let text = read_text(&mut response).await;
        Err(CodexError::Status {
            status: response.status,
            headers: response.headers,
            text,
        })
    };
    crate::provider::retry(signal, aborted_error, send, |error, attempt| {
        let error = match error {
            CodexError::Status { status, headers, text } => {
                if attempt < max_retries && is_retryable_error(status, &text) {
                    return match retry_after_delay_ms(&headers) {
                        None => Ok(BASE_DELAY_MS * 2f64.powi(attempt as i32)),
                        Some(delay) => validate_retry_delay_ms(delay, base.max_retry_delay_ms),
                    };
                }
                let (message, friendly) = parse_error_response(status, &text, jsrt::date::now_ms());
                CodexError::Plain(friendly.unwrap_or(message))
            }
            error => error,
        };
        if error.message() == REQUEST_WAS_ABORTED {
            return Err(aborted_error());
        }
        let retryable = !matches!(error, CodexError::RetryDelay(_)) && !error.message().contains("usage limit");
        if attempt < max_retries && retryable {
            Ok(BASE_DELAY_MS * 2f64.powi(attempt as i32))
        } else {
            Err(error)
        }
    })
    .await
}

/// `streamSimple`: reasoning clamped to the model's levels and sent as `reasoningEffort`.
pub async fn stream_simple(
    model: &Model,
    context: &TranscriptContext,
    options: &SimpleStreamOptions,
    transport: &dyn Transport,
    sink: &mut EventSink<'_>,
) -> Result<AssistantMessage, String> {
    if options.base.api_key.as_deref().is_none_or(str::is_empty) {
        return Err(format!("No API key for provider: {}", model.provider));
    }
    let base = build_base_options(model, context, options);
    let effort = options
        .reasoning
        .as_deref()
        .map(|r| clamp_thinking_level(model, r))
        .filter(|e| *e != "off");
    let codex = OpenAICodexResponsesOptions {
        base,
        reasoning_effort: effort.map(str::to_string),
        tool_choice: options.tool_choice.clone(),
        ..Default::default()
    };
    Ok(stream(model, context, &codex, transport, sink).await)
}