feat(codex): stabilize identity across retries

apps/aether-gateway/src/handlers/proxy/websocket/responses/client.rs

@@ -1,5 +1,6 @@
 //! Client-side Responses WebSocket event forwarding and follow-up planning.
 
+use aether_provider_transport::CodexFingerprintConvergenceContext;
 use axum::extract::ws::{Message as AxumWsMessage, WebSocket};
 use futures_util::SinkExt;
 use serde_json::Value;
@@ -248,7 +249,14 @@ pub(super) async fn forward_client_message(
             // derive one strong live control snapshot that every stage below
             // shares. The connection's Upgrade-time decision is only the
             // immutable identity seed.
-            let planning_parts = build_planning_parts(context);
+            let logical_turn_id = Uuid::now_v7().to_string();
+            let mut planning_parts = build_planning_parts(context);
+            let codex_fingerprint_context =
+                crate::ai_serving::codex_context::attach_codex_logical_turn_context(
+                    &mut planning_parts,
+                    &client_event,
+                    &logical_turn_id,
+                );
             let turn_control = match resolve_responses_websocket_turn_control(
                 state,
                 context,
@@ -453,6 +461,8 @@ pub(super) async fn forward_client_message(
                     context,
                     planning_parts,
                     client_event,
+                    logical_turn_id,
+                    codex_fingerprint_context,
                     turn_control,
                     turn_redaction_session,
                 )
@@ -466,6 +476,8 @@ pub(super) async fn forward_client_message(
                 planning_parts,
                 client_event,
                 requested_model,
+                logical_turn_id,
+                codex_fingerprint_context,
                 turn_control,
                 raw_responses_lite_static_config
                     .expect("independent turns always retain their raw static config"),
@@ -513,6 +525,8 @@ async fn forward_pinned_continuation(
     context: &WebSocketRequestContext,
     planning_parts: http::request::Parts,
     client_event: Value,
+    logical_turn_id: String,
+    codex_fingerprint_context: CodexFingerprintConvergenceContext,
     turn_control: ResponsesWebSocketTurnControl,
     turn_redaction_session: Option<RedactionSession>,
 ) -> RelayDisposition {
@@ -556,7 +570,6 @@ async fn forward_pinned_continuation(
         };
 
     let turn_request_id = Uuid::new_v4().to_string();
-    let logical_turn_id = Uuid::new_v4().to_string();
     let planned = match await_owned_responses_websocket_plan(spawn_owned_responses_websocket_plan(
         state.clone(),
         planning_parts,
@@ -766,6 +779,7 @@ async fn forward_pinned_continuation(
     bound.body_normalization = normalization;
     bound.turn_state.begin(
         LogicalTurn::new(client_event, turn_index, logical_turn_id)
+            .with_codex_fingerprint_context(codex_fingerprint_context)
             .with_provider_store(provider_event.get("store") == Some(&Value::Bool(true)))
             .with_turn_control(turn_control),
         turn,
@@ -800,12 +814,13 @@ async fn forward_replanned_response_create(
     planning_parts: http::request::Parts,
     client_event: Value,
     requested_model: String,
+    logical_turn_id: String,
+    codex_fingerprint_context: CodexFingerprintConvergenceContext,
     turn_control: ResponsesWebSocketTurnControl,
     raw_responses_lite_static_config: ResponsesLiteStaticConfig,
     turn_redaction_session: Option<RedactionSession>,
 ) -> RelayDisposition {
     let turn_request_id = Uuid::new_v4().to_string();
-    let logical_turn_id = Uuid::new_v4().to_string();
     let now_unix_secs = current_unix_secs();
     let excluded_key_ids = bound.exhausted_exclusions.key_ids(now_unix_secs);
     let excluded_codex_account_ids = bound.exhausted_exclusions.codex_account_ids(now_unix_secs);
@@ -992,6 +1007,7 @@ async fn forward_replanned_response_create(
         bound.body_normalization = normalization;
         bound.turn_state.begin(
             LogicalTurn::new(client_event.clone(), turn_index, logical_turn_id.clone())
+                .with_codex_fingerprint_context(codex_fingerprint_context.clone())
                 .with_provider_store(provider_event.get("store") == Some(&Value::Bool(true)))
                 .with_turn_control(turn_control),
             turn,
@@ -1076,6 +1092,7 @@ async fn forward_replanned_response_create(
     bound.binding_identity = replacement.binding_identity;
     bound.turn_state.begin(
         LogicalTurn::new(client_event, turn_index, logical_turn_id)
+            .with_codex_fingerprint_context(codex_fingerprint_context)
             .with_provider_store(provider_event.get("store") == Some(&Value::Bool(true)))
             .with_turn_control(turn_control),
         turn,

apps/aether-gateway/src/handlers/proxy/websocket/responses/quota.rs

@@ -146,6 +146,7 @@ pub(super) async fn retry_active_turn_after_quota_exhaustion(
     };
     let turn_index = active.turn_index;
     let logical_turn_id = active.logical_turn_id.clone();
+    let codex_fingerprint_context = active.codex_fingerprint_context.clone();
     let turn_attempt = active.turn_attempt;
 
     let retry_exclusion_until_unix_secs = bound
@@ -154,7 +155,13 @@ pub(super) async fn retry_active_turn_after_quota_exhaustion(
     let exhausted_key = record_exhausted_bound_key(bound, retry_exclusion_until_unix_secs);
     let exhausted_key_id = exhausted_key.as_ref().map(|(key_id, _)| key_id.clone());
 
-    let planning_parts = build_planning_parts(context);
+    let mut planning_parts = build_planning_parts(context);
+    if let Some(codex_fingerprint_context) = codex_fingerprint_context.as_ref() {
+        crate::ai_serving::codex_context::restore_codex_logical_turn_context(
+            &mut planning_parts,
+            codex_fingerprint_context,
+        );
+    }
     let turn_request_id = Uuid::new_v4().to_string();
     let now_unix_secs = current_unix_secs();
     let excluded_key_ids = bound.exhausted_exclusions.key_ids(now_unix_secs);

apps/aether-gateway/src/handlers/proxy/websocket/responses/session.rs

@@ -444,7 +444,14 @@ async fn bootstrap_responses_websocket(
     let raw_responses_lite_static_config =
         ResponsesLiteStaticConfig::from_response_create(&first_event);
 
-    let planning_parts = build_planning_parts(context);
+    let first_logical_turn_id = Uuid::now_v7().to_string();
+    let mut planning_parts = build_planning_parts(context);
+    let first_codex_fingerprint_context =
+        crate::ai_serving::codex_context::attach_codex_logical_turn_context(
+            &mut planning_parts,
+            &first_event,
+            &first_logical_turn_id,
+        );
     let turn_control = match resolve_responses_websocket_turn_control(
         &state,
         context,
@@ -860,7 +867,6 @@ async fn bootstrap_responses_websocket(
                 return None;
             }
         };
-    let first_logical_turn_id = Uuid::new_v4().to_string();
     let first_turn_decision = prepare_responses_websocket_turn_decision(
         &decision,
         context.trace_id.clone(),
@@ -954,6 +960,7 @@ async fn bootstrap_responses_websocket(
     }
     bound.turn_state.begin(
         LogicalTurn::new(first_event, 1, first_logical_turn_id)
+            .with_codex_fingerprint_context(first_codex_fingerprint_context)
             .with_provider_store(first_provider_event.get("store") == Some(&Value::Bool(true)))
             .with_turn_control(turn_control),
         first_turn,

apps/aether-gateway/src/handlers/proxy/websocket/responses/turn_state.rs

@@ -5,6 +5,7 @@
 //! 非法组合只能靠调用点的 if 和「记得同时改另外两个字段」来避免。这里把它收敛成
 //! 一个枚举：合法组合由类型保证，转换只能走受控 API。
 
+use aether_provider_transport::CodexFingerprintConvergenceContext;
 use serde_json::Value;
 
 use super::control::ResponsesWebSocketTurnControl;
@@ -26,6 +27,9 @@ pub(super) struct LogicalTurn {
     pub(super) provider_store: bool,
     pub(super) turn_index: u64,
     pub(super) logical_turn_id: String,
+    /// Immutable Codex client identity for every provider attempt belonging to
+    /// this logical turn. A transparent re-plan must never mint a new turn.
+    pub(super) codex_fingerprint_context: Option<CodexFingerprintConvergenceContext>,
     pub(super) turn_attempt: u32,
     pub(super) retry_attempted: bool,
     pub(super) retry_unsafe_reason: Option<&'static str>,
@@ -42,6 +46,7 @@ impl LogicalTurn {
             provider_store: false,
             turn_index,
             logical_turn_id,
+            codex_fingerprint_context: None,
             turn_attempt: 1,
             retry_attempted: false,
             retry_unsafe_reason: None,
@@ -54,6 +59,14 @@ impl LogicalTurn {
         self
     }
 
+    pub(super) fn with_codex_fingerprint_context(
+        mut self,
+        context: CodexFingerprintConvergenceContext,
+    ) -> Self {
+        self.codex_fingerprint_context = Some(context);
+        self
+    }
+
     pub(super) fn with_provider_store(mut self, provider_store: bool) -> Self {
         self.provider_store = provider_store;
         self