agcodex-core 0.1.0

Core business logic with AST-RAG engine and tree-sitter integration
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
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
use std::io::BufRead;
use std::path::Path;
use std::sync::OnceLock;
use std::time::Duration;

use agcodex_login::AuthMode;
use agcodex_login::CodexAuth;
use bytes::Bytes;
use eventsource_stream::Eventsource;
use futures::prelude::*;
use regex_lite::Regex;
use reqwest::StatusCode;
use serde::Deserialize;
use serde::Serialize;
use serde_json::Value;
use tokio::sync::mpsc;
use tokio::time::timeout;
use tokio_util::io::ReaderStream;
use tracing::debug;
use tracing::trace;
use tracing::warn;
use uuid::Uuid;

use crate::chat_completions::AggregateStreamExt;
use crate::chat_completions::stream_chat_completions;
use crate::client_common::Prompt;
use crate::client_common::ResponseEvent;
use crate::client_common::ResponseStream;
use crate::client_common::ResponsesApiRequest;
use crate::client_common::create_reasoning_param_for_request;
use crate::config::Config;
use crate::error::CodexErr;
use crate::error::Result;
use crate::error::UsageLimitReachedError;
use crate::flags::CODEX_RS_SSE_FIXTURE;
use crate::model_family::ModelFamily;
use crate::model_provider_info::ModelProviderInfo;
use crate::model_provider_info::WireApi;
use crate::models::ResponseItem;
use crate::openai_tools::create_tools_json_for_responses_api;
use crate::protocol::TokenUsage;
use crate::user_agent::get_codex_user_agent;
use crate::util::backoff;
use agcodex_protocol::config_types::ReasoningEffort as ReasoningEffortConfig;
use agcodex_protocol::config_types::ReasoningSummary as ReasoningSummaryConfig;
use std::sync::Arc;

#[derive(Debug, Deserialize)]
struct ErrorResponse {
    error: Error,
}

#[derive(Debug, Deserialize)]
struct Error {
    r#type: Option<String>,
    code: Option<String>,
    message: Option<String>,
}

#[derive(Debug, Clone)]
pub struct ModelClient {
    config: Arc<Config>,
    auth: Option<CodexAuth>,
    client: reqwest::Client,
    provider: ModelProviderInfo,
    session_id: Uuid,
    effort: ReasoningEffortConfig,
    summary: ReasoningSummaryConfig,
}

impl ModelClient {
    pub fn new(
        config: Arc<Config>,
        auth: Option<CodexAuth>,
        provider: ModelProviderInfo,
        effort: ReasoningEffortConfig,
        summary: ReasoningSummaryConfig,
        session_id: Uuid,
    ) -> Self {
        Self {
            config,
            auth,
            client: reqwest::Client::new(),
            provider,
            session_id,
            effort,
            summary,
        }
    }

    /// Dispatches to either the Responses or Chat implementation depending on
    /// the provider config.  Public callers always invoke `stream()` – the
    /// specialised helpers are private to avoid accidental misuse.
    pub async fn stream(&self, prompt: &Prompt) -> Result<ResponseStream> {
        match self.provider.wire_api {
            WireApi::Responses => self.stream_responses(prompt).await,
            WireApi::Chat => {
                // Create the raw streaming connection first.
                let response_stream = stream_chat_completions(
                    prompt,
                    &self.config.model_family,
                    &self.client,
                    &self.provider,
                )
                .await?;

                // Wrap it with the aggregation adapter so callers see *only*
                // the final assistant message per turn (matching the
                // behaviour of the Responses API).
                let mut aggregated = if self.config.show_raw_agent_reasoning {
                    crate::chat_completions::AggregatedChatStream::streaming_mode(response_stream)
                } else {
                    response_stream.aggregate()
                };

                // Bridge the aggregated stream back into a standard
                // `ResponseStream` by forwarding events through a channel.
                let (tx, rx) = mpsc::channel::<Result<ResponseEvent>>(16);

                tokio::spawn(async move {
                    use futures::StreamExt;
                    while let Some(ev) = aggregated.next().await {
                        // Exit early if receiver hung up.
                        if tx.send(ev).await.is_err() {
                            break;
                        }
                    }
                });

                Ok(ResponseStream { rx_event: rx })
            }
        }
    }

    /// Implementation for the OpenAI *Responses* experimental API.
    async fn stream_responses(&self, prompt: &Prompt) -> Result<ResponseStream> {
        if let Some(path) = &*CODEX_RS_SSE_FIXTURE {
            // short circuit for tests
            warn!(path, "Streaming from fixture");
            return stream_from_fixture(path, self.provider.clone()).await;
        }

        let auth = self.auth.clone();

        let auth_mode = auth.as_ref().map(|a| a.mode);

        let store = prompt.store && auth_mode != Some(AuthMode::ChatGPT);

        let full_instructions = prompt.get_full_instructions(&self.config.model_family);
        let tools_json = create_tools_json_for_responses_api(&prompt.tools)?;
        let reasoning = create_reasoning_param_for_request(
            &self.config.model_family,
            self.effort,
            self.summary,
        );

        // Request encrypted COT if we are not storing responses,
        // otherwise reasoning items will be referenced by ID
        let include: Vec<String> = if !store && reasoning.is_some() {
            vec!["reasoning.encrypted_content".to_string()]
        } else {
            vec![]
        };

        let input_with_instructions = prompt.get_formatted_input();

        let payload = ResponsesApiRequest {
            model: &self.config.model,
            instructions: &full_instructions,
            input: &input_with_instructions,
            tools: &tools_json,
            tool_choice: "auto",
            parallel_tool_calls: false,
            reasoning,
            store,
            stream: true,
            include,
            prompt_cache_key: Some(self.session_id.to_string()),
        };

        let mut attempt = 0;
        let max_retries = self.provider.request_max_retries();

        trace!(
            "POST to {}: {}",
            self.provider.get_full_url(&auth),
            serde_json::to_string(&payload)?
        );

        loop {
            attempt += 1;

            let mut req_builder = self
                .provider
                .create_request_builder(&self.client, &auth)
                .await?;

            req_builder = req_builder
                .header("OpenAI-Beta", "responses=experimental")
                .header("session_id", self.session_id.to_string())
                .header(reqwest::header::ACCEPT, "text/event-stream")
                .json(&payload);

            if let Some(auth) = auth.as_ref()
                && auth.mode == AuthMode::ChatGPT
                && let Some(account_id) = auth.get_account_id()
            {
                req_builder = req_builder.header("chatgpt-account-id", account_id);
            }

            let originator = &self.config.responses_originator_header;
            req_builder = req_builder.header("originator", originator);
            req_builder = req_builder.header("User-Agent", get_codex_user_agent(Some(originator)));

            let res = req_builder.send().await;
            if let Ok(resp) = &res {
                trace!(
                    "Response status: {}, request-id: {}",
                    resp.status(),
                    resp.headers()
                        .get("x-request-id")
                        .map(|v| v.to_str().unwrap_or_default())
                        .unwrap_or_default()
                );
            }

            match res {
                Ok(resp) if resp.status().is_success() => {
                    let (tx_event, rx_event) = mpsc::channel::<Result<ResponseEvent>>(1600);

                    // spawn task to process SSE
                    let stream = resp.bytes_stream().map_err(CodexErr::Reqwest);
                    tokio::spawn(process_sse(
                        stream,
                        tx_event,
                        self.provider.stream_idle_timeout(),
                    ));

                    return Ok(ResponseStream { rx_event });
                }
                Ok(res) => {
                    let status = res.status();

                    // Pull out Retry‑After header if present.
                    let retry_after_secs = res
                        .headers()
                        .get(reqwest::header::RETRY_AFTER)
                        .and_then(|v| v.to_str().ok())
                        .and_then(|s| s.parse::<u64>().ok());

                    if status == StatusCode::UNAUTHORIZED {
                        if let Some(a) = auth.as_ref() {
                            let _ = a.refresh_token().await;
                        }
                        // Retry immediately with refreshed credentials.
                        continue;
                    }

                    // The OpenAI Responses endpoint returns structured JSON bodies even for 4xx/5xx
                    // errors. When we bubble early with only the HTTP status the caller sees an opaque
                    // "unexpected status 400 Bad Request" which makes debugging nearly impossible.
                    // Instead, read (and include) the response text so higher layers and users see the
                    // exact error message (e.g. "Unknown parameter: 'input[0].metadata'"). The body is
                    // small and this branch only runs on error paths so the extra allocation is
                    // negligible.
                    if !(status == StatusCode::TOO_MANY_REQUESTS || status.is_server_error()) {
                        // Surface the error body to callers. Use `unwrap_or_default` per Clippy.
                        let body = res.text().await.unwrap_or_default();
                        return Err(CodexErr::UnexpectedStatus(status, body));
                    }

                    if status == StatusCode::TOO_MANY_REQUESTS {
                        let body = res.json::<ErrorResponse>().await.ok();
                        if let Some(ErrorResponse {
                            error:
                                Error {
                                    r#type: Some(error_type),
                                    ..
                                },
                        }) = body
                        {
                            if error_type == "usage_limit_reached" {
                                return Err(CodexErr::UsageLimitReached(UsageLimitReachedError {
                                    plan_type: auth.and_then(|a| a.get_plan_type()),
                                }));
                            } else if error_type == "usage_not_included" {
                                return Err(CodexErr::UsageNotIncluded);
                            }
                        }
                    }

                    if attempt > max_retries {
                        if status == StatusCode::INTERNAL_SERVER_ERROR {
                            return Err(CodexErr::InternalServerError);
                        }

                        return Err(CodexErr::RetryLimit(status));
                    }

                    let delay = retry_after_secs
                        .map(|s| Duration::from_millis(s * 1_000))
                        .unwrap_or_else(|| backoff(attempt));
                    tokio::time::sleep(delay).await;
                }
                Err(e) => {
                    if attempt > max_retries {
                        return Err(e.into());
                    }
                    let delay = backoff(attempt);
                    tokio::time::sleep(delay).await;
                }
            }
        }
    }

    pub fn get_provider(&self) -> ModelProviderInfo {
        self.provider.clone()
    }

    /// Returns the currently configured model slug.
    pub fn get_model(&self) -> String {
        self.config.model.clone()
    }

    /// Returns the currently configured model family.
    pub fn get_model_family(&self) -> ModelFamily {
        self.config.model_family.clone()
    }

    /// Returns the current reasoning effort setting.
    pub const fn get_reasoning_effort(&self) -> ReasoningEffortConfig {
        self.effort
    }

    /// Returns the current reasoning summary setting.
    pub const fn get_reasoning_summary(&self) -> ReasoningSummaryConfig {
        self.summary
    }

    pub fn get_auth(&self) -> Option<CodexAuth> {
        self.auth.clone()
    }
}

#[derive(Debug, Deserialize, Serialize)]
struct SseEvent {
    #[serde(rename = "type")]
    kind: String,
    response: Option<Value>,
    item: Option<Value>,
    delta: Option<String>,
}

#[derive(Debug, Deserialize)]
struct ResponseCreated {}

#[derive(Debug, Deserialize)]
struct ResponseCompleted {
    id: String,
    usage: Option<ResponseCompletedUsage>,
}

#[derive(Debug, Deserialize)]
struct ResponseCompletedUsage {
    input_tokens: u64,
    input_tokens_details: Option<ResponseCompletedInputTokensDetails>,
    output_tokens: u64,
    output_tokens_details: Option<ResponseCompletedOutputTokensDetails>,
    total_tokens: u64,
}

impl From<ResponseCompletedUsage> for TokenUsage {
    fn from(val: ResponseCompletedUsage) -> Self {
        TokenUsage {
            input_tokens: val.input_tokens,
            cached_input_tokens: val.input_tokens_details.map(|d| d.cached_tokens),
            output_tokens: val.output_tokens,
            reasoning_output_tokens: val.output_tokens_details.map(|d| d.reasoning_tokens),
            total_tokens: val.total_tokens,
        }
    }
}

#[derive(Debug, Deserialize)]
struct ResponseCompletedInputTokensDetails {
    cached_tokens: u64,
}

#[derive(Debug, Deserialize)]
struct ResponseCompletedOutputTokensDetails {
    reasoning_tokens: u64,
}

async fn process_sse<S>(
    stream: S,
    tx_event: mpsc::Sender<Result<ResponseEvent>>,
    idle_timeout: Duration,
) where
    S: Stream<Item = Result<Bytes>> + Unpin,
{
    let mut stream = stream.eventsource();

    // If the stream stays completely silent for an extended period treat it as disconnected.
    // The response id returned from the "complete" message.
    let mut response_completed: Option<ResponseCompleted> = None;
    let mut response_error: Option<CodexErr> = None;

    loop {
        let sse = match timeout(idle_timeout, stream.next()).await {
            Ok(Some(Ok(sse))) => sse,
            Ok(Some(Err(e))) => {
                debug!("SSE Error: {e:#}");
                let event = CodexErr::Stream(e.to_string(), None);
                let _ = tx_event.send(Err(event)).await;
                return;
            }
            Ok(None) => {
                match response_completed {
                    Some(ResponseCompleted {
                        id: response_id,
                        usage,
                    }) => {
                        let event = ResponseEvent::Completed {
                            response_id,
                            token_usage: usage.map(Into::into),
                        };
                        let _ = tx_event.send(Ok(event)).await;
                    }
                    None => {
                        let _ = tx_event
                            .send(Err(response_error.unwrap_or(CodexErr::Stream(
                                "stream closed before response.completed".into(),
                                None,
                            ))))
                            .await;
                    }
                }
                return;
            }
            Err(_) => {
                let _ = tx_event
                    .send(Err(CodexErr::Stream(
                        "idle timeout waiting for SSE".into(),
                        None,
                    )))
                    .await;
                return;
            }
        };

        trace!("SSE event: {}", sse.data);

        let event: SseEvent = match serde_json::from_str(&sse.data) {
            Ok(event) => event,
            Err(e) => {
                debug!("Failed to parse SSE event: {e}, data: {}", &sse.data);
                continue;
            }
        };

        match event.kind.as_str() {
            // Individual output item finalised. Forward immediately so the
            // rest of the agent can stream assistant text/functions *live*
            // instead of waiting for the final `response.completed` envelope.
            //
            // IMPORTANT: We used to ignore these events and forward the
            // duplicated `output` array embedded in the `response.completed`
            // payload.  That produced two concrete issues:
            //   1. No real‑time streaming – the user only saw output after the
            //      entire turn had finished, which broke the "typing" UX and
            //      made long‑running turns look stalled.
            //   2. Duplicate `function_call_output` items – both the
            //      individual *and* the completed array were forwarded, which
            //      confused the backend and triggered 400
            //      "previous_response_not_found" errors because the duplicated
            //      IDs did not match the incremental turn chain.
            //
            // The fix is to forward the incremental events *as they come* and
            // drop the duplicated list inside `response.completed`.
            "response.output_item.done" => {
                let Some(item_val) = event.item else { continue };
                let Ok(item) = serde_json::from_value::<ResponseItem>(item_val) else {
                    debug!("failed to parse ResponseItem from output_item.done");
                    continue;
                };

                let event = ResponseEvent::OutputItemDone(item);
                if tx_event.send(Ok(event)).await.is_err() {
                    return;
                }
            }
            "response.output_text.delta" => {
                if let Some(delta) = event.delta {
                    let event = ResponseEvent::OutputTextDelta(delta);
                    if tx_event.send(Ok(event)).await.is_err() {
                        return;
                    }
                }
            }
            "response.reasoning_summary_text.delta" => {
                if let Some(delta) = event.delta {
                    let event = ResponseEvent::ReasoningSummaryDelta(delta);
                    if tx_event.send(Ok(event)).await.is_err() {
                        return;
                    }
                }
            }
            "response.reasoning_text.delta" => {
                if let Some(delta) = event.delta {
                    let event = ResponseEvent::ReasoningContentDelta(delta);
                    if tx_event.send(Ok(event)).await.is_err() {
                        return;
                    }
                }
            }
            "response.created" => {
                if event.response.is_some() {
                    let _ = tx_event.send(Ok(ResponseEvent::Created {})).await;
                }
            }
            "response.failed" => {
                if let Some(resp_val) = event.response {
                    response_error = Some(CodexErr::Stream(
                        "response.failed event received".to_string(),
                        None,
                    ));

                    let error = resp_val.get("error");

                    if let Some(error) = error {
                        match serde_json::from_value::<Error>(error.clone()) {
                            Ok(error) => {
                                let delay = try_parse_retry_after(&error);
                                let message = error.message.unwrap_or_default();
                                response_error = Some(CodexErr::Stream(message, delay));
                            }
                            Err(e) => {
                                debug!("failed to parse ErrorResponse: {e}");
                            }
                        }
                    }
                }
            }
            // Final response completed – includes array of output items & id
            "response.completed" => {
                if let Some(resp_val) = event.response {
                    match serde_json::from_value::<ResponseCompleted>(resp_val) {
                        Ok(r) => {
                            response_completed = Some(r);
                        }
                        Err(e) => {
                            debug!("failed to parse ResponseCompleted: {e}");
                            continue;
                        }
                    };
                };
            }
            "response.content_part.done"
            | "response.function_call_arguments.delta"
            | "response.in_progress"
            | "response.output_item.added"
            | "response.output_text.done" => {
                // Currently, we ignore this event, but we handle it
                // separately to skip the logging message in the `other` case.
            }
            "response.reasoning_summary_part.added" => {
                // Boundary between reasoning summary sections (e.g., titles).
                let event = ResponseEvent::ReasoningSummaryPartAdded;
                if tx_event.send(Ok(event)).await.is_err() {
                    return;
                }
            }
            "response.reasoning_summary_text.done" => {}
            other => debug!(other, "sse event"),
        }
    }
}

/// used in tests to stream from a text SSE file
async fn stream_from_fixture(
    path: impl AsRef<Path>,
    provider: ModelProviderInfo,
) -> Result<ResponseStream> {
    let (tx_event, rx_event) = mpsc::channel::<Result<ResponseEvent>>(1600);
    let f = std::fs::File::open(path.as_ref())?;
    let lines = std::io::BufReader::new(f).lines();

    // insert \n\n after each line for proper SSE parsing
    let mut content = String::new();
    for line in lines {
        content.push_str(&line?);
        content.push_str("\n\n");
    }

    let rdr = std::io::Cursor::new(content);
    let stream = ReaderStream::new(rdr).map_err(CodexErr::Io);
    tokio::spawn(process_sse(
        stream,
        tx_event,
        provider.stream_idle_timeout(),
    ));
    Ok(ResponseStream { rx_event })
}

fn rate_limit_regex() -> &'static Regex {
    static RE: OnceLock<Regex> = OnceLock::new();

    #[expect(clippy::unwrap_used)]
    RE.get_or_init(|| Regex::new(r"Please try again in (\d+(?:\.\d+)?)(s|ms)").unwrap())
}

fn try_parse_retry_after(err: &Error) -> Option<Duration> {
    if err.code != Some("rate_limit_exceeded".to_string()) {
        return None;
    }

    // parse the Please try again in 1.898s format using regex
    let re = rate_limit_regex();
    if let Some(message) = &err.message
        && let Some(captures) = re.captures(message)
    {
        let seconds = captures.get(1);
        let unit = captures.get(2);

        if let (Some(value), Some(unit)) = (seconds, unit) {
            let value = value.as_str().parse::<f64>().ok()?;
            let unit = unit.as_str();

            if unit == "s" {
                return Some(Duration::from_secs_f64(value));
            } else if unit == "ms" {
                return Some(Duration::from_millis(value as u64));
            }
        }
    }
    None
}

#[cfg(test)]
mod tests {
    use super::*;
    use serde_json::json;
    use tokio::sync::mpsc;
    use tokio_test::io::Builder as IoBuilder;
    use tokio_util::io::ReaderStream;

    // ────────────────────────────
    // Helpers
    // ────────────────────────────

    /// Runs the SSE parser on pre-chunked byte slices and returns every event
    /// (including any final `Err` from a stream-closure check).
    async fn collect_events(
        chunks: &[&[u8]],
        provider: ModelProviderInfo,
    ) -> Vec<Result<ResponseEvent>> {
        let mut builder = IoBuilder::new();
        for chunk in chunks {
            builder.read(chunk);
        }

        let reader = builder.build();
        let stream = ReaderStream::new(reader).map_err(CodexErr::Io);
        let (tx, mut rx) = mpsc::channel::<Result<ResponseEvent>>(16);
        tokio::spawn(process_sse(stream, tx, provider.stream_idle_timeout()));

        let mut events = Vec::new();
        while let Some(ev) = rx.recv().await {
            events.push(ev);
        }
        events
    }

    /// Builds an in-memory SSE stream from JSON fixtures and returns only the
    /// successfully parsed events (panics on internal channel errors).
    async fn run_sse(
        events: Vec<serde_json::Value>,
        provider: ModelProviderInfo,
    ) -> Vec<ResponseEvent> {
        let mut body = String::new();
        for e in events {
            let kind = e
                .get("type")
                .and_then(|v| v.as_str())
                .expect("fixture event missing type");
            if e.as_object().map(|o| o.len() == 1).unwrap_or(false) {
                body.push_str(&format!("event: {kind}\n\n"));
            } else {
                body.push_str(&format!("event: {kind}\ndata: {e}\n\n"));
            }
        }

        let (tx, mut rx) = mpsc::channel::<Result<ResponseEvent>>(8);
        let stream = ReaderStream::new(std::io::Cursor::new(body)).map_err(CodexErr::Io);
        tokio::spawn(process_sse(stream, tx, provider.stream_idle_timeout()));

        let mut out = Vec::new();
        while let Some(ev) = rx.recv().await {
            out.push(ev.expect("channel closed"));
        }
        out
    }

    // ────────────────────────────
    // Tests from `implement-test-for-responses-api-sse-parser`
    // ────────────────────────────

    #[tokio::test]
    async fn parses_items_and_completed() {
        let item1 = json!({
            "type": "response.output_item.done",
            "item": {
                "type": "message",
                "role": "assistant",
                "content": [{"type": "output_text", "text": "Hello"}]
            }
        })
        .to_string();

        let item2 = json!({
            "type": "response.output_item.done",
            "item": {
                "type": "message",
                "role": "assistant",
                "content": [{"type": "output_text", "text": "World"}]
            }
        })
        .to_string();

        let completed = json!({
            "type": "response.completed",
            "response": { "id": "resp1" }
        })
        .to_string();

        let sse1 = format!("event: response.output_item.done\ndata: {item1}\n\n");
        let sse2 = format!("event: response.output_item.done\ndata: {item2}\n\n");
        let sse3 = format!("event: response.completed\ndata: {completed}\n\n");

        let provider = ModelProviderInfo {
            name: "test".to_string(),
            base_url: Some("https://test.com".to_string()),
            env_key: Some("TEST_API_KEY".to_string()),
            env_key_instructions: None,
            wire_api: WireApi::Responses,
            query_params: None,
            http_headers: None,
            env_http_headers: None,
            request_max_retries: Some(0),
            stream_max_retries: Some(0),
            stream_idle_timeout_ms: Some(1000),
            requires_openai_auth: false,
        };

        let events = collect_events(
            &[sse1.as_bytes(), sse2.as_bytes(), sse3.as_bytes()],
            provider,
        )
        .await;

        assert_eq!(events.len(), 3);

        matches!(
            &events[0],
            Ok(ResponseEvent::OutputItemDone(ResponseItem::Message { role, .. }))
                if role == "assistant"
        );

        matches!(
            &events[1],
            Ok(ResponseEvent::OutputItemDone(ResponseItem::Message { role, .. }))
                if role == "assistant"
        );

        match &events[2] {
            Ok(ResponseEvent::Completed {
                response_id,
                token_usage,
            }) => {
                assert_eq!(response_id, "resp1");
                assert!(token_usage.is_none());
            }
            other => panic!("unexpected third event: {other:?}"),
        }
    }

    #[tokio::test]
    async fn error_when_missing_completed() {
        let item1 = json!({
            "type": "response.output_item.done",
            "item": {
                "type": "message",
                "role": "assistant",
                "content": [{"type": "output_text", "text": "Hello"}]
            }
        })
        .to_string();

        let sse1 = format!("event: response.output_item.done\ndata: {item1}\n\n");
        let provider = ModelProviderInfo {
            name: "test".to_string(),
            base_url: Some("https://test.com".to_string()),
            env_key: Some("TEST_API_KEY".to_string()),
            env_key_instructions: None,
            wire_api: WireApi::Responses,
            query_params: None,
            http_headers: None,
            env_http_headers: None,
            request_max_retries: Some(0),
            stream_max_retries: Some(0),
            stream_idle_timeout_ms: Some(1000),
            requires_openai_auth: false,
        };

        let events = collect_events(&[sse1.as_bytes()], provider).await;

        assert_eq!(events.len(), 2);

        matches!(events[0], Ok(ResponseEvent::OutputItemDone(_)));

        match &events[1] {
            Err(CodexErr::Stream(msg, _)) => {
                assert_eq!(msg, "stream closed before response.completed")
            }
            other => panic!("unexpected second event: {other:?}"),
        }
    }

    #[tokio::test]
    async fn error_when_error_event() {
        let raw_error = r#"{"type":"response.failed","sequence_number":3,"response":{"id":"resp_689bcf18d7f08194bf3440ba62fe05d803fee0cdac429894","object":"response","created_at":1755041560,"status":"failed","background":false,"error":{"code":"rate_limit_exceeded","message":"Rate limit reached for gpt-5 in organization org-AAA on tokens per min (TPM): Limit 30000, Used 22999, Requested 12528. Please try again in 11.054s. Visit https://platform.openai.com/account/rate-limits to learn more."}, "usage":null,"user":null,"metadata":{}}}"#;

        let sse1 = format!("event: response.failed\ndata: {raw_error}\n\n");
        let provider = ModelProviderInfo {
            name: "test".to_string(),
            base_url: Some("https://test.com".to_string()),
            env_key: Some("TEST_API_KEY".to_string()),
            env_key_instructions: None,
            wire_api: WireApi::Responses,
            query_params: None,
            http_headers: None,
            env_http_headers: None,
            request_max_retries: Some(0),
            stream_max_retries: Some(0),
            stream_idle_timeout_ms: Some(1000),
            requires_openai_auth: false,
        };

        let events = collect_events(&[sse1.as_bytes()], provider).await;

        assert_eq!(events.len(), 1);

        match &events[0] {
            Err(CodexErr::Stream(msg, delay)) => {
                assert_eq!(
                    msg,
                    "Rate limit reached for gpt-5 in organization org-AAA on tokens per min (TPM): Limit 30000, Used 22999, Requested 12528. Please try again in 11.054s. Visit https://platform.openai.com/account/rate-limits to learn more."
                );
                assert_eq!(*delay, Some(Duration::from_secs_f64(11.054)));
            }
            other => panic!("unexpected second event: {other:?}"),
        }
    }

    // ────────────────────────────
    // Table-driven test from `main`
    // ────────────────────────────

    /// Verifies that the adapter produces the right `ResponseEvent` for a
    /// variety of incoming `type` values.
    #[tokio::test]
    async fn table_driven_event_kinds() {
        struct TestCase {
            name: &'static str,
            event: serde_json::Value,
            expect_first: fn(&ResponseEvent) -> bool,
            expected_len: usize,
        }

        fn is_created(ev: &ResponseEvent) -> bool {
            matches!(ev, ResponseEvent::Created)
        }
        fn is_output(ev: &ResponseEvent) -> bool {
            matches!(ev, ResponseEvent::OutputItemDone(_))
        }
        fn is_completed(ev: &ResponseEvent) -> bool {
            matches!(ev, ResponseEvent::Completed { .. })
        }

        let completed = json!({
            "type": "response.completed",
            "response": {
                "id": "c",
                "usage": {
                    "input_tokens": 0,
                    "input_tokens_details": null,
                    "output_tokens": 0,
                    "output_tokens_details": null,
                    "total_tokens": 0
                },
                "output": []
            }
        });

        let cases = vec![
            TestCase {
                name: "created",
                event: json!({"type": "response.created", "response": {}}),
                expect_first: is_created,
                expected_len: 2,
            },
            TestCase {
                name: "output_item.done",
                event: json!({
                    "type": "response.output_item.done",
                    "item": {
                        "type": "message",
                        "role": "assistant",
                        "content": [
                            {"type": "output_text", "text": "hi"}
                        ]
                    }
                }),
                expect_first: is_output,
                expected_len: 2,
            },
            TestCase {
                name: "unknown",
                event: json!({"type": "response.new_tool_event"}),
                expect_first: is_completed,
                expected_len: 1,
            },
        ];

        for case in cases {
            let mut evs = vec![case.event];
            evs.push(completed.clone());

            let provider = ModelProviderInfo {
                name: "test".to_string(),
                base_url: Some("https://test.com".to_string()),
                env_key: Some("TEST_API_KEY".to_string()),
                env_key_instructions: None,
                wire_api: WireApi::Responses,
                query_params: None,
                http_headers: None,
                env_http_headers: None,
                request_max_retries: Some(0),
                stream_max_retries: Some(0),
                stream_idle_timeout_ms: Some(1000),
                requires_openai_auth: false,
            };

            let out = run_sse(evs, provider).await;
            assert_eq!(out.len(), case.expected_len, "case {}", case.name);
            assert!(
                (case.expect_first)(&out[0]),
                "first event mismatch in case {}",
                case.name
            );
        }
    }

    #[test]
    fn test_try_parse_retry_after() {
        let err = Error {
            r#type: None,
            message: Some("Rate limit reached for gpt-5 in organization org- on tokens per min (TPM): Limit 1, Used 1, Requested 19304. Please try again in 28ms. Visit https://platform.openai.com/account/rate-limits to learn more.".to_string()),
            code: Some("rate_limit_exceeded".to_string()),
        };

        let delay = try_parse_retry_after(&err);
        assert_eq!(delay, Some(Duration::from_millis(28)));
    }

    #[test]
    fn test_try_parse_retry_after_no_delay() {
        let err = Error {
            r#type: None,
            message: Some("Rate limit reached for gpt-5 in organization <ORG> on tokens per min (TPM): Limit 30000, Used 6899, Requested 24050. Please try again in 1.898s. Visit https://platform.openai.com/account/rate-limits to learn more.".to_string()),
            code: Some("rate_limit_exceeded".to_string()),
        };
        let delay = try_parse_retry_after(&err);
        assert_eq!(delay, Some(Duration::from_secs_f64(1.898)));
    }
}