Merge pull request #5307 from lenzhang/fix/ws-prewarm-continuation

fix(ws): allow Codex prewarm continuation

backend/internal/service/openai_ws_forwarder_ingress_test.go

@@ -551,13 +551,22 @@ func TestNormalizeOpenAIWSPayloadWithoutInputAndPreviousResponseID(t *testing.T)
 	t.Parallel()
 
 	normalized, err := normalizeOpenAIWSPayloadWithoutInputAndPreviousResponseID(
-		[]byte(`{"model":"gpt-5.1","input":[1],"previous_response_id":"resp_x","metadata":{"b":2,"a":1}}`),
+		[]byte(`{"model":"gpt-5.1","input":[1],"previous_response_id":"resp_x","client_metadata":{"request_start_ms":"1"},"stream_options":{"include_usage":true},"generate":false,"metadata":{"b":2,"a":1}}`),
 	)
 	require.NoError(t, err)
 	require.False(t, gjson.GetBytes(normalized, "input").Exists())
 	require.False(t, gjson.GetBytes(normalized, "previous_response_id").Exists())
+	require.False(t, gjson.GetBytes(normalized, "client_metadata").Exists())
+	require.False(t, gjson.GetBytes(normalized, "stream_options").Exists())
+	require.False(t, gjson.GetBytes(normalized, "generate").Exists())
 	require.Equal(t, float64(1), gjson.GetBytes(normalized, "metadata.a").Float())
 
+	normalized, err = normalizeOpenAIWSPayloadWithoutInputAndPreviousResponseID(
+		[]byte(`{"model":"gpt-5.1","generate":true}`),
+	)
+	require.NoError(t, err)
+	require.True(t, gjson.GetBytes(normalized, "generate").Bool())
+
 	_, err = normalizeOpenAIWSPayloadWithoutInputAndPreviousResponseID(nil)
 	require.Error(t, err)
 
@@ -662,6 +671,36 @@ func TestShouldKeepIngressPreviousResponseID(t *testing.T) {
 		require.Equal(t, "strict_incremental_ok", reason)
 	})
 
+	t.Run("codex_prewarm_to_business_keep", func(t *testing.T) {
+		prewarmPayload := []byte(`{
+			"type":"response.create",
+			"model":"gpt-5.1",
+			"store":false,
+			"generate":false,
+			"client_metadata":{"x-codex-ws-stream-request-start-ms":"100"},
+			"stream_options":{"include_usage":true},
+			"input":[{"type":"input_text","text":"hello"}]
+		}`)
+		businessPayload := []byte(`{
+			"type":"response.create",
+			"model":"gpt-5.1",
+			"store":false,
+			"client_metadata":{"x-codex-ws-stream-request-start-ms":"200"},
+			"previous_response_id":"resp_prewarm",
+			"input":[{"type":"input_text","text":"hello"}]
+		}`)
+
+		keep, reason, err := shouldKeepIngressPreviousResponseID(
+			prewarmPayload,
+			businessPayload,
+			"resp_prewarm",
+			false,
+		)
+		require.NoError(t, err)
+		require.True(t, keep)
+		require.Equal(t, "strict_incremental_ok", reason)
+	})
+
 	t.Run("missing_previous_response_id", func(t *testing.T) {
 		payload := []byte(`{"type":"response.create","model":"gpt-5.1","input":[]}`)
 		keep, reason, err := shouldKeepIngressPreviousResponseID(previousPayload, payload, "resp_turn_1", false)

backend/internal/service/openai_ws_forwarder_payload.go

@@ -449,6 +449,17 @@ func normalizeOpenAIWSPayloadWithoutInputAndPreviousResponseID(payload []byte) (
 	}
 	delete(decoded, "input")
 	delete(decoded, "previous_response_id")
+	// Codex changes transport-only metadata for every response.create. These fields
+	// do not alter the context referenced by previous_response_id and are excluded
+	// from Codex's own websocket reuse comparison.
+	delete(decoded, "client_metadata")
+	delete(decoded, "stream_options")
+	// Official Codex prewarms a connection with generate=false, then omits the
+	// field on the business request that continues from the prewarm response.
+	// Only normalize false so a meaningful generate=true change remains visible.
+	if generate, ok := decoded["generate"].(bool); ok && !generate {
+		delete(decoded, "generate")
+	}
 	return json.Marshal(decoded)
 }