fix passthought 首token时间计算修复

backend/internal/service/openai_ws_v2/passthrough_relay.go

@@ -82,6 +82,7 @@ type relayState struct {
 	terminalEventType string
 	firstTokenMs      *int
 	turnTimingByID    map[string]*relayTurnTiming
+	activeTurn *relayTurnTiming
 }
 
 type relayExitSignal struct {
@@ -550,6 +551,12 @@ func observeUpstreamMessage(
 		if ms >= 0 {
 			state.firstTokenMs = &ms
 		}
+		if state.activeTurn != nil && state.activeTurn.firstTokenMs == nil {
+			tms := int(now.Sub(state.activeTurn.startAt).Milliseconds())
+			if tms >= 0 {
+				state.activeTurn.firstTokenMs = &tms
+			}
+		}
 	}
 	parsedUsage := parseUsageAndAccumulate(state, message, eventType, onUsageParseFailure)
 	observed := observedUpstreamEvent{
@@ -622,6 +629,7 @@ func openAIWSRelayGetOrInitTurnTiming(state *relayState, responseID string, now
 	if !ok || timing == nil || timing.startAt.IsZero() {
 		timing = &relayTurnTiming{startAt: now}
 		state.turnTimingByID[responseID] = timing
+		state.activeTurn = timing
 		return timing
 	}
 	return timing
@@ -636,6 +644,9 @@ func openAIWSRelayDeleteTurnTiming(state *relayState, responseID string) (relayT
 		return relayTurnTiming{}, false
 	}
 	delete(state.turnTimingByID, responseID)
+	if state.activeTurn == timing {
+		state.activeTurn = nil
+	}
 	return *timing, true
 }
 

backend/internal/service/openai_ws_v2/passthrough_relay_test.go

@@ -750,3 +750,67 @@ func (c *errorOnWriteFrameConn) WriteFrame(_ context.Context, _ coderws.MessageT
 func (c *errorOnWriteFrameConn) Close() error {
 	return nil
 }
+
+func TestRelay_OnTurnComplete_RealOpenAIStream_FirstTokenMs(t *testing.T) {
+	t.Parallel()
+
+	clientConn := newPassthroughTestFrameConn(nil, false)
+	upstreamConn := newPassthroughTestFrameConn([]passthroughTestFrame{
+		{
+			msgType: coderws.MessageText,
+			payload: []byte(`{"type":"response.created","response":{"id":"resp_real"}}`),
+		},
+		{
+			msgType: coderws.MessageText,
+			payload: []byte(`{"type":"response.output_text.delta","delta":"He"}`),
+		},
+		{
+			msgType: coderws.MessageText,
+			payload: []byte(`{"type":"response.output_text.delta","delta":"llo"}`),
+		},
+		{
+			msgType: coderws.MessageText,
+			payload: []byte(`{"type":"response.output_text.delta","delta":" world"}`),
+		},
+		{
+			msgType: coderws.MessageText,
+			payload: []byte(`{"type":"response.completed","response":{"id":"resp_real","usage":{"input_tokens":2,"output_tokens":3}}}`),
+		},
+	}, true)
+
+	firstPayload := []byte(`{"type":"response.create","model":"gpt-5.3-codex","input":[]}`)
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Second)
+	defer cancel()
+
+	base := time.Unix(0, 0)
+	var nowTick atomic.Int64
+	nowFn := func() time.Time {
+		step := nowTick.Add(1)
+		return base.Add(time.Duration(step) * 10 * time.Millisecond)
+	}
+
+	var turn RelayTurnResult
+	result, relayExit := Relay(ctx, clientConn, upstreamConn, firstPayload, RelayOptions{
+		Now: nowFn,
+		OnTurnComplete: func(current RelayTurnResult) {
+			turn = current
+		},
+	})
+	require.Nil(t, relayExit)
+	require.Equal(t, "resp_real", turn.RequestID)
+	require.Equal(t, "response.completed", turn.TerminalEventType)
+
+	require.NotNil(t, turn.FirstTokenMs, "per-turn FirstTokenMs must be captured for real OpenAI streams")
+	require.Greater(t, turn.Duration.Milliseconds(), int64(0))
+
+	require.Less(t,
+		int64(*turn.FirstTokenMs),
+		turn.Duration.Milliseconds(),
+		"per-turn FirstTokenMs (%dms) should be strictly less than Duration (%dms); "+
+			"equality indicates the bug where first_token is mistakenly stamped on the terminal event",
+		*turn.FirstTokenMs, turn.Duration.Milliseconds(),
+	)
+
+	require.NotNil(t, result.FirstTokenMs)
+	require.Greater(t, *result.FirstTokenMs, 0)
+}