llm/providers/codex/
provider.rs

1use super::mappers::{map_messages, map_tools};
2use super::oauth::CodexTokenManager;
3use super::streaming::process_response_stream;
4use crate::oauth::credential_store::OAuthCredentialStore;
5use crate::provider::{LlmResponseStream, ProviderFactory, StreamingModelProvider, get_context_window};
6use crate::{Context, LlmError, Result};
7use async_openai::types::responses::{
8    CreateResponse, IncludeEnum, InputParam, Reasoning, ReasoningEffort, ReasoningSummary, ResponseStreamEvent,
9    ResponseTextParam, TextResponseFormatConfiguration, Verbosity,
10};
11use eventsource_stream::Eventsource;
12use futures::StreamExt;
13use reqwest::header::{AUTHORIZATION, HeaderMap, HeaderValue};
14use std::sync::Arc;
15use tracing::debug;
16
17const CODEX_API_BASE: &str = "https://chatgpt.com/backend-api/codex";
18
19#[derive(Clone)]
20pub struct CodexProvider {
21    client: reqwest::Client,
22    model: String,
23    token_manager: Arc<CodexTokenManager<OAuthCredentialStore>>,
24}
25
26impl CodexProvider {
27    pub fn new(token_manager: CodexTokenManager<OAuthCredentialStore>) -> Self {
28        Self { client: reqwest::Client::new(), model: "gpt-5.4".to_string(), token_manager: Arc::new(token_manager) }
29    }
30
31    fn build_request(&self, context: &Context) -> Result<CreateResponse> {
32        let (system_prompt, input) = map_messages(context.messages())?;
33        let tools = if context.tools().is_empty() { None } else { Some(map_tools(context.tools())?) };
34
35        let codex_effort = context.reasoning_effort().map_or(ReasoningEffort::Medium, to_codex_effort);
36
37        Ok(CreateResponse {
38            model: Some(self.model.clone()),
39            input: InputParam::Items(input),
40            instructions: system_prompt,
41            tools,
42            store: Some(false),
43            stream: Some(true),
44            reasoning: Some(Reasoning { effort: Some(codex_effort), summary: Some(ReasoningSummary::Auto) }),
45            include: Some(vec![IncludeEnum::ReasoningEncryptedContent]),
46            text: Some(ResponseTextParam {
47                format: TextResponseFormatConfiguration::Text,
48                verbosity: Some(Verbosity::Medium),
49            }),
50            prompt_cache_key: context.prompt_cache_key().map(String::from),
51            ..Default::default()
52        })
53    }
54
55    async fn build_headers(&self) -> Result<HeaderMap> {
56        let (access_token, account_id) = self.token_manager.get_valid_token().await?;
57
58        let mut headers = HeaderMap::new();
59        headers.insert(
60            AUTHORIZATION,
61            HeaderValue::from_str(&format!("Bearer {access_token}"))
62                .map_err(|e| LlmError::InvalidApiKey(e.to_string()))?,
63        );
64        headers.insert(
65            "chatgpt-account-id",
66            HeaderValue::from_str(&account_id).map_err(|e| LlmError::InvalidApiKey(e.to_string()))?,
67        );
68        headers.insert("OpenAI-Beta", HeaderValue::from_static("responses=experimental"));
69        headers.insert("originator", HeaderValue::from_static("codex_cli_rs"));
70
71        Ok(headers)
72    }
73
74    /// Send the request and return a stream of SSE lines parsed into typed events.
75    ///
76    /// Uses manual SSE parsing because the Codex API does not return a
77    /// `Content-Type: text/event-stream` header, which `reqwest_eventsource`
78    /// (used by `async-openai`'s `create_stream`) requires.
79    async fn send_request(
80        &self,
81        request: CreateResponse,
82        headers: HeaderMap,
83    ) -> Result<impl futures::Stream<Item = Result<ResponseStreamEvent>>> {
84        let url = format!("{CODEX_API_BASE}/responses");
85
86        debug!("Sending request to Codex API: {url}");
87        debug!(
88            "Codex request body: {}",
89            serde_json::to_string(&request).unwrap_or_else(|_| "<failed to serialize>".to_string())
90        );
91
92        let response = self
93            .client
94            .post(&url)
95            .headers(headers)
96            .json(&request)
97            .send()
98            .await
99            .map_err(|e| LlmError::ApiRequest(e.to_string()))?;
100
101        if !response.status().is_success() {
102            let status = response.status();
103            let error_text = response.text().await.unwrap_or_else(|_| "Unknown error".to_string());
104
105            if status.as_u16() == 401 || status.as_u16() == 403 {
106                self.token_manager.clear_cache().await;
107            }
108
109            return Err(LlmError::ApiError(format!("Codex API request failed with status {status}: {error_text}")));
110        }
111
112        let event_stream = response.bytes_stream().eventsource().filter_map(|result| {
113            std::future::ready(match result {
114                Ok(event) if event.data == "[DONE]" => None,
115                Ok(event) => match serde_json::from_str::<ResponseStreamEvent>(&event.data) {
116                    Ok(parsed) => Some(Ok(parsed)),
117                    Err(e) => {
118                        debug!("Failed to parse Codex SSE line: {} - Error: {e}", event.data);
119                        None
120                    }
121                },
122                Err(e) => Some(Err(LlmError::IoError(e.to_string()))),
123            })
124        });
125
126        Ok(event_stream)
127    }
128}
129
130impl ProviderFactory for CodexProvider {
131    async fn from_env() -> Result<Self> {
132        let store = OAuthCredentialStore::new(super::PROVIDER_ID);
133        let token_manager = CodexTokenManager::new(store, super::PROVIDER_ID);
134        Ok(Self::new(token_manager))
135    }
136
137    fn with_model(mut self, model: &str) -> Self {
138        self.model = model.to_string();
139        self
140    }
141}
142
143impl StreamingModelProvider for CodexProvider {
144    fn model(&self) -> Option<crate::LlmModel> {
145        format!("{}:{}", super::PROVIDER_ID, self.model).parse().ok()
146    }
147
148    fn context_window(&self) -> Option<u32> {
149        get_context_window(super::PROVIDER_ID, &self.model)
150    }
151
152    fn stream_response(&self, context: &Context) -> LlmResponseStream {
153        let provider = self.clone();
154        let context = match self.model() {
155            Some(model) => context.filter_encrypted_reasoning(&model),
156            None => context.clone(),
157        };
158
159        Box::pin(async_stream::stream! {
160            let headers = match provider.build_headers().await {
161                Ok(h) => h,
162                Err(e) => {
163                    yield Err(e);
164                    return;
165                }
166            };
167
168            let request = match provider.build_request(&context) {
169                Ok(r) => r,
170                Err(e) => {
171                    yield Err(e);
172                    return;
173                }
174            };
175
176            let event_stream = match provider.send_request(request, headers).await {
177                Ok(s) => s,
178                Err(e) => {
179                    yield Err(e);
180                    return;
181                }
182            };
183
184            let mut response_stream = Box::pin(process_response_stream(event_stream));
185            while let Some(result) = response_stream.next().await {
186                yield result;
187            }
188        })
189    }
190
191    fn display_name(&self) -> String {
192        format!("Codex ({})", self.model)
193    }
194}
195
196fn to_codex_effort(effort: crate::ReasoningEffort) -> ReasoningEffort {
197    match effort {
198        crate::ReasoningEffort::Low => ReasoningEffort::Low,
199        crate::ReasoningEffort::Medium => ReasoningEffort::Medium,
200        crate::ReasoningEffort::High => ReasoningEffort::High,
201        crate::ReasoningEffort::Xhigh => ReasoningEffort::Xhigh,
202    }
203}
204
205#[cfg(test)]
206mod tests {
207    use super::*;
208    use crate::ChatMessage;
209    use crate::ContentBlock;
210    use crate::ToolDefinition;
211    use crate::types::IsoString;
212
213    fn create_test_provider() -> CodexProvider {
214        let store = OAuthCredentialStore::new("codex-test");
215        let tm = CodexTokenManager::new(store, "codex-test");
216        CodexProvider::new(tm).with_model("gpt-5.4")
217    }
218
219    #[test]
220    fn build_request_simple() {
221        let provider = create_test_provider();
222        let context = Context::new(
223            vec![ChatMessage::User { content: vec![ContentBlock::text("Hello")], timestamp: IsoString::now() }],
224            vec![],
225        );
226
227        let request = provider.build_request(&context).unwrap();
228        assert_eq!(request.model.as_deref(), Some("gpt-5.4"));
229        assert_eq!(request.store, Some(false));
230        assert_eq!(request.stream, Some(true));
231        assert!(request.tools.is_none());
232        assert!(request.instructions.is_none());
233        if let InputParam::Items(items) = &request.input {
234            assert_eq!(items.len(), 1);
235        } else {
236            panic!("Expected InputParam::Items");
237        }
238    }
239
240    #[test]
241    fn build_request_with_system_and_tools() {
242        let provider = create_test_provider();
243        let context = Context::new(
244            vec![
245                ChatMessage::System { content: "You are helpful".to_string(), timestamp: IsoString::now() },
246                ChatMessage::User { content: vec![ContentBlock::text("Hello")], timestamp: IsoString::now() },
247            ],
248            vec![ToolDefinition {
249                name: "bash".to_string(),
250                description: "Run a command".to_string(),
251                parameters: r#"{"type": "object", "properties": {"cmd": {"type": "string"}}}"#.to_string(),
252                server: None,
253            }],
254        );
255
256        let request = provider.build_request(&context).unwrap();
257        assert!(request.instructions.is_some());
258        if let InputParam::Items(items) = &request.input {
259            assert_eq!(items.len(), 1);
260        } else {
261            panic!("Expected InputParam::Items");
262        }
263        assert!(request.tools.is_some());
264        assert_eq!(request.tools.as_ref().unwrap().len(), 1);
265    }
266
267    #[test]
268    fn display_name_includes_model() {
269        let provider = create_test_provider();
270        assert_eq!(provider.display_name(), "Codex (gpt-5.4)");
271    }
272
273    #[test]
274    fn build_request_defaults_to_medium_effort() {
275        let provider = create_test_provider();
276        let context = Context::new(
277            vec![ChatMessage::User { content: vec![ContentBlock::text("Hi")], timestamp: IsoString::now() }],
278            vec![],
279        );
280
281        let request = provider.build_request(&context).unwrap();
282        let json = serde_json::to_value(&request).unwrap();
283        assert_eq!(json["reasoning"]["effort"], "medium");
284    }
285
286    #[test]
287    fn build_request_uses_context_reasoning_effort() {
288        let provider = create_test_provider();
289        let mut context = Context::new(
290            vec![ChatMessage::User { content: vec![ContentBlock::text("Think hard")], timestamp: IsoString::now() }],
291            vec![],
292        );
293        context.set_reasoning_effort(Some(crate::ReasoningEffort::High));
294
295        let request = provider.build_request(&context).unwrap();
296        let json = serde_json::to_value(&request).unwrap();
297        assert_eq!(json["reasoning"]["effort"], "high");
298    }
299
300    #[test]
301    fn build_request_serializes_correctly() {
302        let provider = create_test_provider();
303        let context = Context::new(
304            vec![ChatMessage::User { content: vec![ContentBlock::text("Hi")], timestamp: IsoString::now() }],
305            vec![],
306        );
307
308        let request = provider.build_request(&context).unwrap();
309        let json = serde_json::to_value(&request).unwrap();
310
311        assert_eq!(json["model"], "gpt-5.4");
312        assert_eq!(json["store"], false);
313        assert_eq!(json["stream"], true);
314        assert_eq!(json["reasoning"]["effort"], "medium");
315        assert_eq!(json["text"]["verbosity"], "medium");
316        assert_eq!(json["include"][0], "reasoning.encrypted_content");
317    }
318
319    #[test]
320    fn build_request_includes_prompt_cache_key_when_set() {
321        let provider = create_test_provider();
322        let mut context = Context::new(
323            vec![ChatMessage::User { content: vec![ContentBlock::text("Hi")], timestamp: IsoString::now() }],
324            vec![],
325        );
326        context.set_prompt_cache_key(Some("session-abc".to_string()));
327
328        let request = provider.build_request(&context).unwrap();
329        assert_eq!(request.prompt_cache_key.as_deref(), Some("session-abc"));
330    }
331
332    #[test]
333    fn build_request_omits_prompt_cache_key_when_unset() {
334        let provider = create_test_provider();
335        let context = Context::new(
336            vec![ChatMessage::User { content: vec![ContentBlock::text("Hi")], timestamp: IsoString::now() }],
337            vec![],
338        );
339
340        let request = provider.build_request(&context).unwrap();
341        assert!(request.prompt_cache_key.is_none());
342    }
343}
llm/providers/codex/provider.rs

llm/providers/codex/
provider.rs
