Record reasoning replay removal metrics (#61017)

## What changed

Parse reasoning removals from response `input_transformations` and carry the counts through completion events to sampling telemetry.

- Add `codex.responses.reasoning_replay` to distinguish unavailable reports, no reported removals, prefix mismatches, and other removals.
- Add `codex.responses.reasoning_items_removed` histograms by bounded removal reason, grouping unknown or missing reasons under `other`.
- Include zero counts for usable reports; exclude missing, null, or malformed reports from histograms without failing the response.
- Use the sampling step's session telemetry when recording response events.

## Testing

Add SSE and WebSocket integration coverage for report availability, removal counts, unknown reasons, zero observations, and `product_sku` tags in exported metrics.

GitOrigin-RevId: 83c70ca6067fa4b34bf5c463958b4632a7a0a188

codex-rs/codex-api/src/common.rs

@@ -16,6 +16,7 @@ use serde::Deserialize;
 use serde::Serialize;
 use serde_json::Value;
 use serde_json::value::RawValue;
+use std::collections::BTreeMap;
 use std::collections::HashMap;
 use std::pin::Pin;
 use std::sync::Arc;
@@ -78,6 +79,40 @@ pub struct MemorySummarizeOutput {
 #[derive(Clone, Debug)]
 pub struct ResponseId(pub String);
 
+/// A bounded reason for removing a replayed reasoning item from model input.
+#[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd, Ord)]
+pub enum ReasoningRemovalReason {
+    OrganizationMismatch,
+    EndUserMismatch,
+    ModelMismatch,
+    PrefixMismatch,
+    TurnBoundary,
+    Other,
+}
+
+impl ReasoningRemovalReason {
+    /// Every metric category, including unrecognized API reasons.
+    pub const ALL: [Self; 6] = [
+        Self::OrganizationMismatch,
+        Self::EndUserMismatch,
+        Self::ModelMismatch,
+        Self::PrefixMismatch,
+        Self::TurnBoundary,
+        Self::Other,
+    ];
+
+    pub fn as_str(self) -> &'static str {
+        match self {
+            Self::OrganizationMismatch => "organization_mismatch",
+            Self::EndUserMismatch => "end_user_mismatch",
+            Self::ModelMismatch => "model_mismatch",
+            Self::PrefixMismatch => "prefix_mismatch",
+            Self::TurnBoundary => "turn_boundary",
+            Self::Other => "other",
+        }
+    }
+}
+
 #[derive(Debug)]
 pub enum ResponseEvent {
     Created {
@@ -104,6 +139,9 @@ pub enum ResponseEvent {
         response_id: String,
         token_usage: Option<TokenUsage>,
         usage_metadata: Option<ResponseUsageMetadata>,
+        /// Reported reasoning removals by bounded reason. Missing or non-array
+        /// reports are unavailable; an empty map means no removals were reported.
+        reasoning_removals: Option<BTreeMap<ReasoningRemovalReason, i64>>,
         /// Did the model affirmatively end its turn? Some providers do not set this,
         /// so we rely on fallback logic when this is `None`.
         end_turn: Option<bool>,

codex-rs/codex-api/src/lib.rs

@@ -42,6 +42,7 @@ pub use crate::common::RawMemory;
 pub use crate::common::RawMemoryMetadata;
 pub use crate::common::Reasoning;
 pub use crate::common::ReasoningContext;
+pub use crate::common::ReasoningRemovalReason;
 pub use crate::common::ReasoningSummaryDelivery;
 pub use crate::common::ResponseCreateWsRequest;
 pub use crate::common::ResponseEvent;

codex-rs/codex-api/src/sse/responses.rs

@@ -1,4 +1,5 @@
 use super::responses_error::parse_failed_response;
+use crate::common::ReasoningRemovalReason;
 use crate::common::ResponseEvent;
 use crate::common::ResponseStream;
 use crate::common::SafetyBuffering;
@@ -21,6 +22,7 @@ use eventsource_stream::Eventsource;
 use futures::StreamExt;
 use serde::Deserialize;
 use serde_json::Value;
+use std::collections::BTreeMap;
 use std::sync::Arc;
 use std::sync::OnceLock;
 use std::time::Duration;
@@ -372,6 +374,29 @@ impl ResponsesEventError {
     }
 }
 
+// Diagnostic fields are optional and must never make an otherwise valid response fail.
+fn reasoning_removals(response: &Value) -> Option<BTreeMap<ReasoningRemovalReason, i64>> {
+    let transformations = response.get("input_transformations")?.as_array()?;
+    let mut counts = BTreeMap::new();
+    for transformation in transformations {
+        if transformation.get("type").and_then(Value::as_str) != Some("reasoning")
+            || transformation.get("action").and_then(Value::as_str) != Some("removed")
+        {
+            continue;
+        }
+        let reason = match transformation.get("reason").and_then(Value::as_str) {
+            Some("organization_mismatch") => ReasoningRemovalReason::OrganizationMismatch,
+            Some("end_user_mismatch") => ReasoningRemovalReason::EndUserMismatch,
+            Some("model_mismatch") => ReasoningRemovalReason::ModelMismatch,
+            Some("prefix_mismatch") => ReasoningRemovalReason::PrefixMismatch,
+            Some("turn_boundary") => ReasoningRemovalReason::TurnBoundary,
+            _ => ReasoningRemovalReason::Other,
+        };
+        *counts.entry(reason).or_default() += 1;
+    }
+    Some(counts)
+}
+
 pub fn process_responses_event(
     event: ResponsesStreamEvent,
 ) -> std::result::Result<Option<ResponseEvent>, ResponsesEventError> {
@@ -492,6 +517,7 @@ pub fn process_responses_event(
                 }
             }
             if let Some(resp_val) = event.response {
+                let reasoning_removals = reasoning_removals(&resp_val);
                 let metadata = resp_val
                     .get("usage")
                     .filter(|usage| !usage.is_null())
@@ -505,6 +531,7 @@ pub fn process_responses_event(
                             response_id: resp.id,
                             token_usage: resp.usage.map(Into::into),
                             usage_metadata: resp.usage_metadata,
+                            reasoning_removals,
                             end_turn: if interrupted {
                                 Some(false)
                             } else {
@@ -881,6 +908,7 @@ mod tests {
                 response_id,
                 token_usage,
                 usage_metadata,
+                reasoning_removals: _,
                 end_turn,
             }) => {
                 assert_eq!(response_id, "resp1");
@@ -1079,6 +1107,7 @@ mod tests {
                 response_id,
                 token_usage,
                 usage_metadata,
+                reasoning_removals: _,
                 end_turn,
             }) => {
                 assert_eq!(response_id, "resp1");
@@ -1675,6 +1704,7 @@ mod tests {
                 response_id,
                 token_usage: None,
                 usage_metadata: None,
+                reasoning_removals: None,
                 end_turn: None,
             } if response_id == "resp-1"
         );
@@ -1716,6 +1746,7 @@ mod tests {
                 response_id,
                 token_usage: None,
                 usage_metadata: None,
+                reasoning_removals: None,
                 end_turn: None,
             } if response_id == "resp-1"
         );
@@ -1752,6 +1783,7 @@ mod tests {
                 response_id,
                 token_usage: None,
                 usage_metadata: None,
+                reasoning_removals: None,
                 end_turn: None,
             } if response_id == "resp-1"
         );
@@ -1788,6 +1820,7 @@ mod tests {
                 response_id,
                 token_usage: None,
                 usage_metadata: None,
+                reasoning_removals: None,
                 end_turn: None,
             } if response_id == "resp-1"
         );