Retry compaction after model 404

codex-rs/core/src/compact_model_fallback.rs

@@ -2,8 +2,17 @@ use codex_analytics::CompactionImplementation;
 use codex_analytics::CompactionReason;
 use codex_otel::SessionTelemetry;
 use codex_protocol::error::CodexErr;
+use http::StatusCode;
 use tracing::warn;
 
+pub(crate) fn should_retry_with_current_model(error: &CodexErr) -> bool {
+    match error {
+        CodexErr::InvalidRequest(_) => true,
+        CodexErr::UnexpectedStatus(response) => response.status == StatusCode::NOT_FOUND,
+        _ => false,
+    }
+}
+
 pub(crate) fn record_model_fallback(
     session_telemetry: &SessionTelemetry,
     previous_model: &str,

codex-rs/core/src/compact_remote.rs

@@ -8,6 +8,7 @@ use crate::compact::build_compaction_initial_context;
 use crate::compact::compaction_status_from_result;
 use crate::compact::insert_initial_context_before_last_real_user_or_summary;
 use crate::compact_model_fallback::record_model_fallback;
+use crate::compact_model_fallback::should_retry_with_current_model;
 use crate::context::world_state::WorldState;
 use crate::context_manager::ContextManager;
 use crate::hook_runtime::PostCompactHookOutcome;
@@ -218,7 +219,7 @@ async fn run_remote_compact_task_inner_impl(
             let Some(fallback_step_context) = fallback_step_context else {
                 return Err(error);
             };
-            if !matches!(&error, CodexErr::InvalidRequest(_)) {
+            if !should_retry_with_current_model(&error) {
                 return Err(error);
             }
             let fallback_turn_context = &fallback_step_context.turn;

codex-rs/core/src/compact_remote_v2.rs

@@ -9,6 +9,7 @@ use crate::compact::CompactionAnalyticsDetails;
 use crate::compact::InitialContextInjection;
 use crate::compact::compaction_status_from_result;
 use crate::compact_model_fallback::record_model_fallback;
+use crate::compact_model_fallback::should_retry_with_current_model;
 use crate::compact_remote::process_compacted_history;
 use crate::compact_remote::should_keep_compacted_history_item;
 use crate::hook_runtime::PostCompactHookOutcome;
@@ -230,7 +231,7 @@ async fn run_remote_compact_task_inner_impl(
             let Some(fallback_step_context) = fallback_step_context else {
                 return Err(error);
             };
-            if !matches!(&error, CodexErr::InvalidRequest(_)) {
+            if !should_retry_with_current_model(&error) {
                 return Err(error);
             }
             let fallback_turn_context = &fallback_step_context.turn;

codex-rs/core/tests/suite/compact.rs

@@ -289,6 +289,17 @@ fn invalid_request_response(message: impl Into<String>) -> wiremock::ResponseTem
     }))
 }
 
+fn model_not_found_response(model: &str) -> wiremock::ResponseTemplate {
+    wiremock::ResponseTemplate::new(/*status*/ 404).set_body_json(json!({
+        "error": {
+            "message": format!("Model not found {model}"),
+            "type": "invalid_request_error",
+            "param": "model",
+            "code": null,
+        }
+    }))
+}
+
 fn write_global_file(
     home: &TempDir,
     filename: &str,
@@ -2293,7 +2304,7 @@ async fn pre_sampling_compact_runs_when_comp_hash_changes() {
 }
 
 #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
-async fn pre_sampling_compact_falls_back_from_retired_previous_model_after_rename() {
+async fn pre_sampling_compact_falls_back_when_previous_model_is_not_found() {
     skip_if_no_network!();
 
     let server = MockServer::start().await;
@@ -2320,9 +2331,7 @@ async fn pre_sampling_compact_falls_back_from_retired_previous_model_after_renam
                 ev_assistant_message("m1", "before switch"),
                 ev_completed_with_tokens("r1", /*total_tokens*/ 100),
             ])),
-            invalid_request_response(format!(
-                "The '{retired_model}' model is not supported when using Codex with a ChatGPT account."
-            )),
+            model_not_found_response(retired_model),
             sse_response(sse(vec![
                 json!({
                     "type": "response.output_item.done",
@@ -2384,7 +2393,9 @@ async fn pre_sampling_compact_falls_back_from_retired_previous_model_after_renam
     })
     .await;
 
-    let model_provider = openai_model_provider(&server);
+    let mut model_provider = openai_model_provider(&server);
+    model_provider.request_max_retries = Some(0);
+    model_provider.stream_max_retries = Some(0);
     let mut resumed_builder = test_codex()
         .with_auth(CodexAuth::create_dummy_chatgpt_auth_for_testing())
         .with_model(retired_model)