fix(openai-ws): exclude terminal events from TTFT

backend/internal/service/openai_ws_v2/passthrough_relay.go

@@ -953,23 +953,7 @@ func shouldParseUsage(eventType string) bool {
 }
 
 func isTokenEvent(eventType string) bool {
-	if eventType == "" {
-		return false
-	}
-	switch eventType {
-	case "response.created", "response.in_progress", "response.output_item.added", "response.output_item.done":
-		return false
-	}
-	if strings.Contains(eventType, ".delta") {
-		return true
-	}
-	if strings.HasPrefix(eventType, "response.output_text") {
-		return true
-	}
-	if strings.HasPrefix(eventType, "response.output") {
-		return true
-	}
-	return eventType == "response.completed" || eventType == "response.done"
+	return strings.HasSuffix(strings.TrimSpace(eventType), ".delta")
 }
 
 func minDuration(a, b time.Duration) time.Duration {

backend/internal/service/openai_ws_v2/passthrough_relay_internal_test.go

@@ -246,7 +246,7 @@ func TestHelperFunctionsCoverage(t *testing.T) {
 
 	require.True(t, isTokenEvent("response.output_text.delta"))
 	require.True(t, isTokenEvent("response.output_audio.delta"))
-	require.True(t, isTokenEvent("response.completed"))
+	require.False(t, isTokenEvent("response.completed"))
 	require.False(t, isTokenEvent(""))
 	require.False(t, isTokenEvent("response.created"))
 
@@ -407,8 +407,29 @@ func TestIsTokenEventCoverageBranches(t *testing.T) {
 	require.False(t, isTokenEvent("response.in_progress"))
 	require.False(t, isTokenEvent("response.output_item.added"))
 	require.True(t, isTokenEvent("response.output_audio.delta"))
-	require.True(t, isTokenEvent("response.output"))
-	require.True(t, isTokenEvent("response.done"))
+	require.True(t, isTokenEvent("response.function_call_arguments.delta"))
+	require.True(t, isTokenEvent("response.reasoning_summary_text.delta"))
+	require.False(t, isTokenEvent("response.output"))
+	require.False(t, isTokenEvent("response.output_text.done"))
+	require.False(t, isTokenEvent("response.output_audio.done"))
+	require.False(t, isTokenEvent("response.output_text.annotation.added"))
+	require.False(t, isTokenEvent("response.done"))
+}
+
+func TestTerminalAndTokenEventSetsAreDisjoint(t *testing.T) {
+	t.Parallel()
+
+	for _, eventType := range []string{
+		"response.completed",
+		"response.done",
+		"response.failed",
+		"response.incomplete",
+		"response.cancelled",
+		"response.canceled",
+	} {
+		require.True(t, isTerminalEvent(eventType), eventType)
+		require.False(t, isTokenEvent(eventType), eventType)
+	}
 }
 
 func TestShouldParseUsageTerminalEvents(t *testing.T) {

backend/internal/service/openai_ws_v2/passthrough_relay_test.go

@@ -202,7 +202,7 @@ func TestRelay_BasicRelayAndUsage(t *testing.T) {
 	require.Equal(t, 7, result.Usage.InputTokens)
 	require.Equal(t, 3, result.Usage.OutputTokens)
 	require.Equal(t, 2, result.Usage.CacheReadInputTokens)
-	require.NotNil(t, result.FirstTokenMs)
+	require.Nil(t, result.FirstTokenMs)
 	require.Equal(t, int64(1), result.ClientToUpstreamFrames)
 	require.Equal(t, int64(1), result.UpstreamToClientFrames)
 	require.Equal(t, int64(0), result.DroppedDownstreamFrames)
@@ -865,6 +865,64 @@ func (c *errorOnWriteFrameConn) Close() error {
 	return nil
 }
 
+func TestRelay_NoDeltaTerminalSequence_FirstTokenMsNil(t *testing.T) {
+	t.Parallel()
+
+	for _, terminalEvent := range []string{"response.completed", "response.done"} {
+		terminalEvent := terminalEvent
+		t.Run(terminalEvent, func(t *testing.T) {
+			t.Parallel()
+
+			clientConn := newPassthroughTestFrameConn(nil, false)
+			upstreamConn := newPassthroughTestFrameConn([]passthroughTestFrame{
+				{
+					msgType: coderws.MessageText,
+					payload: []byte(`{"type":"response.created","response":{"id":"resp_no_delta"}}`),
+				},
+				{
+					msgType: coderws.MessageText,
+					payload: []byte(`{"type":"response.in_progress","response":{"id":"resp_no_delta"}}`),
+				},
+				{
+					msgType: coderws.MessageText,
+					payload: []byte(`{"type":"response.output_text.done","response_id":"resp_no_delta","text":""}`),
+				},
+				{
+					msgType: coderws.MessageText,
+					payload: []byte(`{"type":"response.content_part.done","response_id":"resp_no_delta"}`),
+				},
+				{
+					msgType: coderws.MessageText,
+					payload: []byte(`{"type":"response.output_item.done","response_id":"resp_no_delta"}`),
+				},
+				{
+					msgType: coderws.MessageText,
+					payload: []byte(`{"type":"` + terminalEvent + `","response":{"id":"resp_no_delta","usage":{"input_tokens":2,"output_tokens":0}}}`),
+				},
+			}, true)
+
+			ctx, cancel := context.WithTimeout(context.Background(), 2*time.Second)
+			defer cancel()
+
+			var turn RelayTurnResult
+			result, relayExit := Relay(
+				ctx,
+				clientConn,
+				upstreamConn,
+				[]byte(`{"type":"response.create","model":"gpt-5.3-codex","input":[]}`),
+				RelayOptions{OnTurnComplete: func(current RelayTurnResult) { turn = current }},
+			)
+
+			require.Nil(t, relayExit)
+			require.Equal(t, terminalEvent, turn.TerminalEventType)
+			require.Nil(t, turn.FirstTokenMs)
+			require.Equal(t, terminalEvent, result.TerminalEventType)
+			require.Nil(t, result.FirstTokenMs)
+			require.Equal(t, int64(6), result.UpstreamToClientFrames)
+		})
+	}
+}
+
 func TestRelay_OnTurnComplete_RealOpenAIStream_FirstTokenMs(t *testing.T) {
 	t.Parallel()