lean_ctx/proxy/
chatgpt.rs

1use axum::{
2    body::Body,
3    extract::State,
4    http::{HeaderName, Request, StatusCode},
5    response::Response,
6};
7
8use super::{ProxyState, forward, openai_responses};
9
10/// Codex subscription model turns hit ChatGPT's Responses-compatible rail:
11/// `/backend-api/codex/responses`. Forward through the same compressor/metering
12/// path as OpenAI Responses, but target `https://chatgpt.com`.
13pub async fn codex_responses_handler(
14    State(state): State<ProxyState>,
15    mut req: Request<Body>,
16) -> Result<Response, StatusCode> {
17    // Drop the "responses-lite" marker before forwarding. Codex requests the
18    // reduced lite transport on its HTTP path (`supports_websockets = false`),
19    // but chatgpt.com rejects newer subscription models there
20    // ("This model is not supported when using X-OpenAI-Internal-Codex-Responses-Lite",
21    // seen with gpt-5.5). Stripping it makes chatgpt.com serve the full Responses
22    // rail every model supports; Codex parses the full stream identically
23    // (verified single- + multi-turn `previous_response_id` continuation). #623
24    req.headers_mut().remove(CODEX_RESPONSES_LITE_HEADER);
25    let upstream = state.chatgpt_upstream();
26    forward::forward_request(
27        State(state),
28        req,
29        &upstream,
30        "/backend-api/codex/responses",
31        openai_responses::compress_request_body,
32        "ChatGPT",
33        &[],
34    )
35    .await
36}
37
38/// Codex's HTTP fallback marks the reduced "responses-lite" transport with this
39/// header. chatgpt.com gates newer subscription models behind the full rail, so
40/// the Codex ChatGPT handler strips it (see [`codex_responses_handler`]).
41const CODEX_RESPONSES_LITE_HEADER: &str = "x-openai-internal-codex-responses-lite";
42
43/// ChatGPT's Codex rail rejects WS-only continuation fields such as
44/// `previous_response_id`; ask Codex to retry through the HTTP/SSE path.
45pub async fn codex_responses_ws_handler(
46    State(_state): State<ProxyState>,
47    _headers: axum::http::HeaderMap,
48    _ws: axum::extract::ws::WebSocketUpgrade,
49) -> Response {
50    chatgpt_responses_ws_fallback_response()
51}
52
53fn chatgpt_responses_ws_fallback_response() -> Response {
54    Response::builder()
55        .status(StatusCode::UPGRADE_REQUIRED)
56        .header("content-type", "application/json")
57        .body(Body::from(
58            r#"{"error":{"type":"unsupported_transport","message":"ChatGPT codex responses use HTTP/SSE; retry without WebSocket."}}"#,
59        ))
60        .expect("static response is valid")
61}
62
63/// ChatGPT backend calls outside the model rail are not model JSON and must not be
64/// compressed or cost-metered. They are credential-preserving passthroughs.
65pub async fn backend_api_handler(
66    State(state): State<ProxyState>,
67    req: Request<Body>,
68) -> Result<Response, StatusCode> {
69    let (parts, body) = req.into_parts();
70    let body_bytes = axum::body::to_bytes(body, forward::max_body_bytes())
71        .await
72        .map_err(|_| StatusCode::PAYLOAD_TOO_LARGE)?;
73    let upstream = state.chatgpt_upstream();
74    let path = parts
75        .uri
76        .path_and_query()
77        .map_or("/backend-api", axum::http::uri::PathAndQuery::as_str);
78    let url = format!("{upstream}{path}");
79
80    let mut upstream_req = state.client.request(parts.method.clone(), &url);
81    for (key, value) in &parts.headers {
82        if is_backend_passthrough_request_header(key) {
83            upstream_req = upstream_req.header(key.clone(), value.clone());
84        }
85    }
86
87    let response = upstream_req
88        .body(body_bytes.to_vec())
89        .send()
90        .await
91        .map_err(|e| {
92            tracing::error!("lean-ctx proxy: ChatGPT backend upstream error: {e}");
93            StatusCode::BAD_GATEWAY
94        })?;
95
96    let status = StatusCode::from_u16(response.status().as_u16()).unwrap_or(StatusCode::OK);
97    let headers = response.headers().clone();
98    let is_stream = headers
99        .get("content-type")
100        .and_then(|v| v.to_str().ok())
101        .is_some_and(|ct| ct.contains("text/event-stream"));
102
103    let mut out = Response::builder().status(status);
104    for (key, value) in &headers {
105        if is_backend_passthrough_response_header(key) {
106            out = out.header(key, value);
107        }
108    }
109
110    if is_stream {
111        return out
112            .body(Body::from_stream(response.bytes_stream()))
113            .map_err(|_| StatusCode::INTERNAL_SERVER_ERROR);
114    }
115
116    let bytes = response
117        .bytes()
118        .await
119        .map_err(|_| StatusCode::BAD_GATEWAY)?;
120
121    out.body(Body::from(bytes))
122        .map_err(|_| StatusCode::INTERNAL_SERVER_ERROR)
123}
124
125fn is_backend_passthrough_request_header(name: &HeaderName) -> bool {
126    let lower = name.as_str().to_ascii_lowercase();
127    !matches!(
128        lower.as_str(),
129        "host"
130            | "connection"
131            | "content-length"
132            | "transfer-encoding"
133            | "upgrade"
134            | "keep-alive"
135            | "proxy-authenticate"
136            | "proxy-authorization"
137            | "te"
138            | "trailer"
139            | "accept-encoding"
140    )
141}
142
143fn is_backend_passthrough_response_header(name: &HeaderName) -> bool {
144    let lower = name.as_str().to_ascii_lowercase();
145    !matches!(
146        lower.as_str(),
147        "connection"
148            | "content-length"
149            | "transfer-encoding"
150            | "upgrade"
151            | "keep-alive"
152            | "proxy-authenticate"
153            | "proxy-authorization"
154            | "te"
155            | "trailer"
156    )
157}
158
159#[cfg(test)]
160mod tests {
161    use std::sync::Arc;
162    use std::time::Duration;
163
164    use tokio::io::{AsyncReadExt, AsyncWriteExt};
165
166    use super::*;
167    use crate::core::config::Upstreams;
168
169    fn proxy_state(chatgpt_upstream: String) -> ProxyState {
170        let (_tx, rx) = tokio::sync::watch::channel(Arc::new(Upstreams {
171            anthropic: "https://api.anthropic.com".into(),
172            openai: "https://api.openai.com".into(),
173            chatgpt: chatgpt_upstream,
174            gemini: "https://generativelanguage.googleapis.com".into(),
175        }));
176        ProxyState {
177            client: reqwest::Client::new(),
178            port: 0,
179            stats: Arc::new(crate::proxy::ProxyStats::default()),
180            introspect: Arc::new(crate::proxy::introspect::IntrospectState::default()),
181            upstreams: rx,
182        }
183    }
184
185    #[test]
186    fn codex_responses_ws_requests_trigger_http_fallback() {
187        let response = chatgpt_responses_ws_fallback_response();
188        assert_eq!(response.status(), StatusCode::UPGRADE_REQUIRED);
189        assert_eq!(
190            response
191                .headers()
192                .get(axum::http::header::CONTENT_TYPE)
193                .unwrap(),
194            "application/json"
195        );
196    }
197
198    async fn spawn_streaming_upstream() -> (String, tokio::sync::oneshot::Receiver<String>) {
199        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
200        let addr = listener.local_addr().unwrap();
201        let (tx, rx) = tokio::sync::oneshot::channel();
202        tokio::spawn(async move {
203            let (mut socket, _) = listener.accept().await.unwrap();
204            let mut buf = Vec::new();
205            loop {
206                let mut chunk = [0_u8; 1024];
207                let n = socket.read(&mut chunk).await.unwrap();
208                if n == 0 {
209                    break;
210                }
211                buf.extend_from_slice(&chunk[..n]);
212                if buf.windows(4).any(|w| w == b"\r\n\r\n") {
213                    break;
214                }
215            }
216            let _ = tx.send(String::from_utf8_lossy(&buf).into_owned());
217            socket
218                .write_all(
219                    b"HTTP/1.1 200 OK\r\n\
220                      content-type: text/event-stream\r\n\
221                      mcp-session-id: server-session\r\n\
222                      cache-control: no-cache\r\n\
223                      x-custom-backend-state: passthrough\r\n\
224                      \r\n\
225                      event: message\n\
226                      data: {\"jsonrpc\":\"2.0\"}\n\n",
227                )
228                .await
229                .unwrap();
230            tokio::time::sleep(Duration::from_secs(2)).await;
231        });
232        (format!("http://{addr}"), rx)
233    }
234
235    #[tokio::test]
236    async fn backend_api_streams_mcp_sse_and_preserves_session_headers() {
237        let (upstream, seen_request) = spawn_streaming_upstream().await;
238        let state = proxy_state(upstream);
239        let req = Request::builder()
240            .method("POST")
241            .uri("/backend-api/ps/mcp?transport=streamable")
242            .header("Authorization", "Bearer codex-token")
243            .header("Mcp-Session-Id", "client-session")
244            .header("Last-Event-ID", "event-7")
245            .header("X-OpenAI-Product-Sku", "codex")
246            .header("X-OpenAI-Internal-Codex-Residency", "us")
247            .header("Originator", "codex_cli_rs")
248            .header("Accept", "application/json, text/event-stream")
249            .body(Body::empty())
250            .unwrap();
251
252        let response = tokio::time::timeout(
253            Duration::from_millis(500),
254            backend_api_handler(State(state), req),
255        )
256        .await
257        .expect("SSE passthrough must return after upstream headers")
258        .expect("backend request should succeed");
259
260        assert_eq!(response.status(), StatusCode::OK);
261        assert_eq!(
262            response.headers().get("mcp-session-id").unwrap(),
263            "server-session"
264        );
265        assert_eq!(
266            response.headers().get("x-custom-backend-state").unwrap(),
267            "passthrough"
268        );
269
270        let request = seen_request.await.unwrap().to_ascii_lowercase();
271        assert!(request.contains("post /backend-api/ps/mcp?transport=streamable http/1.1"));
272        assert!(request.contains("authorization: bearer codex-token"));
273        assert!(request.contains("mcp-session-id: client-session"));
274        assert!(request.contains("last-event-id: event-7"));
275        assert!(request.contains("x-openai-product-sku: codex"));
276        assert!(request.contains("x-openai-internal-codex-residency: us"));
277        assert!(request.contains("originator: codex_cli_rs"));
278    }
279}
lean_ctx/proxy/chatgpt.rs

lean_ctx/proxy/
chatgpt.rs
